Helping People Work Through a Shared Decision

An applied information design method for examining one decision through the responsibilities of the people who have a part in it.

Use comparable AI research and source inspection to identify relevant information and prepare specific questions for the people involved.

AN APPLICATION OF INFORMATION DESIGN

A focused investigation within the guide family

Information Design provides the purpose: understand what information must enable for people. This note teaches one applied research exercise within that practice. AI Visibility provides methods for observing the AI answer environment; Business Impact investigates organizational consequence when that question arises. The three guides retain their existing responsibilities. [R01–R03]

Distinguish tool observations, source findings, and human requirements

The AI research record establishes what the recorded system searched for, retrieved, cited, or said under the specified conditions.

Source inspection establishes what a document contains and what its evidence supports. The source’s account and independent verification of that account are distinct findings.

A proposed requirement describes what a person may need information to enable. When it comes only from prompt assumptions or AI behavior, it remains a hypothesis. Records, expertise, direct inquiry, or other evidence about the people doing the work can strengthen, qualify, or overturn it. [R01, R02]

Use the family’s evidence states

Provenance means where a statement comes from: for example, a prompt, a run record, an inspected passage, or a conversation. Record it beside the statement. Use an evidence state when the strength of a claim matters to the next decision. [R01-R03]

1. FRAME THE SHARED DECISION

Describe one decision and one common situation

Start with one consequential choice or commitment. Name who can make or authorize it, who contributes judgment or practical support, and who will live with its effects. These parts may overlap. Record the circumstances, the conditions that matter, and the realistic alternatives. Adapt the scope to the decision in front of you, whether it concerns work, education, a household, or another setting.

Throughout this note, the worked example is Asteron Industrial Systems, a synthetic industrial company deciding what to do next with an AI-visibility program. The program has produced encouraging visibility results while direct referral traffic remains small and downstream business effects remain uncertain. The CMO must recommend whether to continue, expand, change, or reallocate the investment – and prepare that recommendation for serious scrutiny from colleagues with different responsibilities. [R04-R06]

The Asteron decision used throughout this note is:

What should Asteron do with the next period of AI-visibility investment, and what evidence would help the people involved evaluate that commitment well?

Separate the shared decision from the research decision. The people in the situation make the shared choice. The person using this method decides what to research, inspect, verify, or ask next. In Asteron, the company decides what to fund; the investigator decides what evidence can responsibly be carried into that discussion.

Write down the common situation once

Save one version of the common factual block and use it across the role prompts being compared. When a fact applies to everyone, keep it in that shared block. When a difference is specific to a represented responsibility, identify it explicitly. This lets the reader see which prompt features changed. The Asteron research used a frozen situation and explicit prompt-construction rules for this purpose. [R04, R05]

Keep all realistic outcomes available

Ask what would justify each realistic option, including proceeding, revising, choosing an alternative, gathering more evidence, deferring, or stopping. Record the criteria people actually use when those are known. Identify proposed criteria as assumptions to check. [R01]

When external tools are used, use approved and appropriately scoped material. Keep private identities and sensitive facts out of external systems unless disclosure is authorized. Asteron is synthetic and exists to make the method inspectable.

WHAT TO DO Write the shared decision in one sentence. Save a dated version of the common facts and their sources. Name material rules or policies, or record them as unresolved. If the research assumes a governing context, record that assumption. State what this first research pass must help you decide to inspect, ask, or change.

2. NAME THE RESPONSIBILITIES

Describe what each person is responsible for judging or contributing

In this note, a responsibility is the judgment, choice, task, or contribution being represented in a research prompt. A role is its label in the working record. Use the person’s actual part in the situation where it is known; record a proposed part as a hypothesis. A job title is optional, and one person may have several responsibilities. [R01]

Record the distinction between decision authority, advice, practical contribution, and being affected by the result. People can share consequences while retaining different rights and choices. Ask them about the constraints, preferences, and commitments that research cannot establish for them. A shared decision may end in agreement, a revised option, a stated disagreement, or a decision to wait.

Asteron example: the CMO prepares colleagues to review a recommendation

In the Asteron teaching case, the CMO must recommend what happens next with the AI-visibility program. She needs to understand the result and prepare information that colleagues can examine from their own responsibilities. This is one business application of the broader method.

In the Asteron example, the CMO needs to give each colleague enough of the record to ask the questions their responsibility requires. A colleague may support the recommendation, add a condition, correct it, or change it. The CEO then carries the resulting recommendation, including material disagreements and unresolved questions, into Board review.

These are modeled responsibilities in the Asteron study, rather than interview findings about particular people. The research examines what the selected system does when those responsibilities are supplied. [R04-R06]

Start with what you actually know about each responsibility

An existing record, conversation, practitioner account, or other evidence may ground a responsibility. You may also begin with a plausible responsibility to investigate. Preserve that difference. When the initial account is hypothetical, the research shows what follows from the modeled responsibility; it does not establish that the real person holds it.

The full Asteron experiment used eight organizational roles plus a Neutral condition. A live investigation can begin with the few responsibilities most likely to change the decision, recommendation, or questions that need to be resolved. [R05, R06]

WHAT TO DO For each represented role, state what the person needs to judge, choose, or contribute. Record the basis for that description, their authority or practical influence, and the questions about their part in the decision that remain open.

3. ASK COMPARABLE QUESTIONS

Keep the situation the same while you change the responsibility being represented

Give each represented responsibility its own fresh AI research request. Preserve the common factual record and use a consistent evidence request. Change the responsibility clause deliberately and save its exact wording. This makes the resulting differences easier to inspect. [R04]

A small Asteron starting design can use the CMO, CFO, Legal, and Content conditions plus a role-agnostic Neutral baseline, all at the same knowledge level and decision moment. The baseline asks about the shared commitment while omitting explicit role identity. The full study expands this logic to 81 conditions. [R05, R06]

The baseline shows what the system does with the common situation alone. A role comparison adds both the title and the responsibility as written. Its result concerns that combined prompt condition. A question about a bare-title effect requires a separate test. [R04, R06]

Preserve what you put into the prompt

Add more context only when it answers a specific research question

A second comparison can ask what changes when more of the situation is specified. Preserve the original prompt and add the extra facts. Keep a record of the addition, including uncertainty and alternative explanations. Compare levels within the same role and decision moment. [R04]

Use additional facts to investigate a defined question. Costs, timing, relevant conditions, constraints, and alternative explanations may change the research. Save the added material exactly, including any named sources or methods. In Asteron, the protocol added context while leaving the system to select methods and publishers. [R04, R05]

A separate comparison can examine different research systems. Use the same prompt family and record the model, search service, settings, and dates for each. Asteron used one documented system; comparisons across systems remain further work. [R06]

Decide repetitions based on what you want to claim

One run can provide an initial source list to inspect. Repeating the same condition helps reveal ordinary variation. Choose the repetition plan to match the claim, stakes, and available resources. Preserve failed runs and runs with no visible search. One run per exact condition remains exploratory. [R02, R06]

WHAT TO DO Save the prompt family and common record before collection. Label the comparison you are making: role, supplied knowledge, decision moment, or system.

A RESEARCH PROMPT YOU CAN ADAPT

Start with the shared decision, then describe one responsibility

WORKING TEMPLATE. Use this structure for a new investigation. The historical Asteron workbook remains the exact instrument for reproducing the reported experiment. [R04, R05]

We are investigating the information available around this shared decision: [decision].

Here is the common situation record: [paste the approved facts, their conditions, and the important unknowns].

Approach this from the perspective of [person or role]. This person is responsible for [specific judgment or task] within the shared decision. Their authority, practical influence, and relevant constraints are [details, with assumptions identified].

Investigate the questions that arise for that responsibility. Use current external evidence where relevant and cite the sources you rely on. Explain which statements the sources support, which conditions affect applicability, and what remains unresolved.

Distinguish information supplied in this prompt, information added by external sources, and your interpretation of the situation. State what the person may need to establish and the basis for each proposed requirement. Describe what further evidence would change the assessment.

For a role-agnostic baseline, keep the common facts and evidence request the same while leaving out a particular responsibility. Ask what needs to be established for the shared decision. Save this prompt with the role-specific versions.

Asteron responsibility clauses

These summarize the responsibilities represented in the Asteron design; use the frozen workbook for the executed wording. The role clause supplies a concern to the model. The research then shows what the system adds, qualifies, or challenges within that framing. [R04-R06]

Separate the question from the instrument

The prompt asks for research and a sourced answer. The instrument determines which search queries, surfaced URLs, explicit opens, and citation records can actually be captured. Save exposed tool events through the instrument; treat a model-written account of its searches as a statement to verify against the available trace.

WHAT TO DO Run each condition independently. Preserve the exact prompt, returned answer, exposed events, and run identifier before asking an assistant to summarize or compare them.

HOW THE ASSISTANT AND RESEARCH TOOL WORK TOGETHER

Four different jobs in one investigation

You frame the decision and approve the comparison. Your assistant helps prepare requests and work with the returned records. The research instrument executes prompts and exposes activity. The investigator then checks the records, reads sources, and states what the evidence supports.

MCP, or Model Context Protocol, is the connection through which an assistant application accesses an external tool. The tool supplies its particular functions and records. [R22]

Choose between a new run and a new analysis

A new research run tests a new condition or repeats an existing one. A new analysis can instead revisit stored runs: compare another pair, inspect an uncited source, or examine a qualification in an answer. Preserve the original prompt, settings, and run identifier in either case. [R21]

Separate capture, recovery, and preservation

Capture is the record the instrument exposes. Recovery is how the assistant or client obtains that record. Preservation is how the returned data are saved for later inspection. Check each capability. Keep this use of recovery distinct from retrieved sources, which refers to URLs encountered during the model’s research.

For the dated Thinking Inspector implementation used in this research, see Appendix K. It covers the recorded tool functions, response formats, existing-run recovery, settings, and the remaining access and archival questions.

WHAT TO DO Identify who designs the investigation, which system performs the research, which records are available, and how the records will be preserved.

4. CAPTURE WHAT THE TOOL EXPOSES

Establish what is observable before comparing runs

Begin with a capability check. An instrument runs a prompt against a research system and exposes some record of what happened. Describe those capabilities separately from the route used to recover and save the record.

Capture. Record access to the exact prompt and settings, search queries, original source URLs, explicit open/find actions where exposed, cited URLs, full answer, stable run identifier, and any query-to-URL linkage.

Complete recovery. Establish whether the relevant response is full or condensed, whether results are paginated, and how returned counts are reconciled with the expected total.

Preservation. Demonstrate a direct route for saving complete returned data, through a native export or an authorized client-side capture. Preserve original response data separately from analytical tables. When a route remains unresolved, keep the authoritative run identifiers and state the archival gap.

The dated TI briefing documents the available responses and reports no file-writing command among the seven tools visible to that operator. Access to a full answer and a tested way to save it are separate capabilities. Appendix K explains both, with their verification status. [R21]

Use the records actually available for the comparison they can support. AI Visibility provides the broader tool-evaluation method. [R02]

In this note, retrieved means the surfaced or consulted URLs exposed for a run. The original Asteron dataset separately labels SURFACED as true and RETRIEVED as NOT_OBSERVABLE. Preserve those raw values. Explicitly opened remains a separate action flag. Appendix E records the full terminology and historical examples. [R07]

A citation can appear without a separately recorded page open. An answer-only tool may reveal citations while leaving the wider retrieval set unknown. Capture the layers actually exposed, then describe the analysis at that scope. [R02, R07]

A matched Asteron slice shows what a compact capture inventory can look like. These four L1 DECIDE runs share the same decision moment and knowledge level; only the represented responsibility changes. [R06, R07]

The difference in visible search activity is itself part of the record. Explicit opens are optional recorded actions, rather than a stage every citation must pass through. The cited-set differences below are therefore interpreted from the surfaced, opened, cited, and answer records separately. [R06, R07]

Keep prompts, runs, and source records linked

Use a stable run identifier connecting the prompt, role, common-record version, settings, answer, and URLs. Preserve original URL strings; record any grouping of subdomains, aliases, or page variants separately. For a larger collection, one row per run and original URL makes recurrence and source states easier to compare. [R07]

Keep the model’s page-opening events separate from the investigator’s later reading. Connect a specific query to a URL only when the instrument exposes that link. In Asteron, queries and source sets are available by run, while a reliable query-to-URL link for each source remains unavailable. [R06, R07]

Handle a run with no visible external research

Retain the result as part of the record. A further prompt variant can test another question about retrieval, and a separately logged source check can investigate the evidence directly. Each is a new research action. The original condition continues to describe the original run.

WHAT TO DO Save a simple capture inventory: available layers, missing layers, instrument definition, collection date, and any anomalies. Then begin the role comparison.

5. COMPARE THE SOURCES AND READ THEM

Start by listing the questions the system actually investigated

Query fan-out (QFO) is the set of smaller search queries an AI system generates while answering the larger prompt you supplied. When the instrument records them, these queries are concrete mechanical information-demand signals: requests for information made by the system under the recorded situation and responsibility. Read the wording, including any named source, alongside the original prompt. It shows what was searched for; whether the sources answer it and whether people need that information are separate questions. [R02, R06, R28]

For each represented responsibility, summarize the main questions visible in the trace and keep the raw strings available. Separate concerns already assigned in the prompt from new conditions, exceptions, methods, or source questions introduced during research.

For example, a role clause that explicitly asks about claim substantiation already supplies that concern. The useful additional learning is which evidence the system seeks, which distinctions the sources support, and what it leaves unresolved. Asteron’s role conditions deliberately combined accountability and situation; the resulting differences belong to that construction. [R04, R06]

Compare sources at more than one level

A domain identifies a host or source organization; a page identifies a selected resource; a passage shows what a source actually says. Compare at the level needed for your question. Shared domains may contain different pages or qualifications. Read the relevant material before interpreting similarity or difference as useful to a person. [R13, R18]

A count is optional when direct inspection is enough. For larger sets, the following section describes the measures and their limits. State whether the items are URLs, domains, or publisher groups, and how URLs were normalized. [R06]

PRIORITIZE FREQUENTLY RETRIEVED, UNCITED SOURCES

When a relevant page or domain appears repeatedly in retrieval but rarely or never appears in the same runs’ citations, give it priority for inspection. Pay particular attention when a QFO explicitly names the source. Use this as a lead for developing questions that a subject-matter expert or someone with relevant experience could help answer.

For example, a generated search might contain reddit or site:reddit.com. The first names Reddit in the search text; the second asks for results from that domain. Check the full query: a search about Reddit policy and a search for firsthand experience on Reddit ask for different information. These examples explain what to look for; they are not quotations from a specific Asteron run.

State the scope and counting unit: runs containing the source, distinct pages, or page appearances across runs. In Asteron, the reported 2,111 Reddit retrieval records and zero cited records count run-by-URL appearances across the batch. They identify a source pool to inspect, rather than 2,111 distinct threads or proof that each query requested Reddit. [R06, R07]

Read the pages to develop useful questions

Read the original question, substantive responses, and relevant disagreements. Look for a number with an unclear basis, a workaround, a decision criterion, a reported result, or a condition that changes the answer. A page may already contain an informative account. Your next question can ask someone to explain its circumstances, supply supporting records, or contribute contrasting experience. [R29]

Suppose an inspected thread describes a team continuing a visibility program because of sales-team feedback despite small referral numbers. Commenters disagree about whether that feedback justified the spending. This hypothetical thread could prompt the question below; it is an illustration, not a finding from an inspected Asteron page.

“Have you used sales-team feedback when deciding whether to continue a marketing program? Describe a specific decision. What did the feedback establish, how did you check it, and what other evidence affected your choice?”

Keep the QFO, run identifiers, URL, relevant passage, and resulting interview question together. Ask the respondent for their own evidence and experience, including accounts that challenge the original claim. Preserve the original source’s contribution and distinguish what the interview independently adds.

The recorded pattern establishes search, retrieval, and citation status. The source’s influence on the original answer and the reason it was uncited remain open. Link a query directly to a page only where the instrument records that relationship; otherwise describe their connection as your assessment within the run. [R06, R07]

WHAT TO DO: Flag recurring retrieved-but-uncited sources, especially those explicitly named in QFOs. Read the relevant pages and prepare questions that could add explanation, supporting records, or contrasting experience. Step 8 shows how to choose respondents, use their accounts, and assess any resulting publication.

Read each source for what it actually establishes

Record the specific claim, its supporting passage or data, the source’s basis, and the applicable date, version, setting, and limits. Note sponsorship and editorial control where they affect the source relationship. A manufacturer can establish its stated capability or terms; an independent test can establish an observed result under its test conditions; an official document can establish a requirement within its remit.

WHAT TO DO For each material source, write: “This page supports ____ under ____ conditions. It leaves ____ unresolved for the represented responsibility.”

HOW WE COMPARE CITATION SETS

Sources used together and sources shared across answers

This method compares the sources that role-conditioned answers cite. Two established bibliometric ideas describe that comparison, and they answer different questions.

Sources used together. Co-citation counts how often two sources are cited together; Small (1973) introduced the measure. Here one answer supplies one co-occurrence. Across answers, those occurrences can show recurring bundles of displayed evidence. Inspect the claims those sources support before interpreting the bundle. [R23]

Sources shared across answers. Bibliographic coupling relates documents through shared references (Kessler, 1963). Here, the documents are AI-generated answers. Asteron’s Jaccard comparisons measure how much their cited URL sets overlap. This describes selected evidence under the recorded conditions. [R06, R23]

For example, the Asteron CMO and Content L1 DECIDE answers share two cited URLs. That is a comparison across answers. A co-citation question instead asks which sources appear together within one answer, and whether that pairing recurs across answers. Answer those questions separately. [R06, R07]

How this application differs from the original bibliometric use

The application here uses deliberately constructed prompts: the recorded system, shared situation, and evidence request are held fixed while responsibility wording varies. The result describes those generated answers. Use the source map to guide source inspection and provisional role requirements; direct claims about people require evidence about their work.

Say exactly what you are counting

State whether you are counting exact URL strings, pages, hostnames, or publisher groups. A query parameter or alternate address can make the same page appear as two exact URLs. Grouping at domain level answers a broader question about which organizations appear. Keep the original addresses and document every grouping rule. [R07, R13, R18]

Asteron’s worked comparison uses exact cited URL strings. Apply that rule consistently when reproducing its reported values. For a new analysis, decide in advance whether to combine page variants, tracking parameters, trailing slashes, or mirror hosts; then report the rule with the result. [R06, R07]

Explain the denominator and the limits of an overlap measure

Jaccard divides the number of shared items by all distinct items in the two sets. In the Asteron CMO-Content comparison, the answers cite 12 and 9 URLs and share 2. Their union contains 19 URLs, so Jaccard is 2/19, or 0.105. The overlap coefficient instead divides by the smaller set: 2/9, or approximately 0.222. It answers how much of the smaller set is shared. Report the set sizes with either measure. [R06, R07]

A zero-overlap pair, a both-empty pair, and a pair where one condition performed no external search are three different observations. Record each as what it is.

State only what the comparison itself establishes

A cited-set difference establishes that the recorded system displayed different supporting sources under the compared prompts. Claims about which sources people would use, source quality, or causal influence require additional evidence. Examine answers separately: in Asteron, different cited sets coexisted with the analyst-reported broad funding recommendation. [R06-R08]

WHAT TO DO For each comparison, record the counting unit, normalization rule, measure, set sizes, and whether you are describing sources used together within an answer or sources shared across answers. Then record the specific question the difference makes worth checking.

DISTINGUISH THE PROMPT, THE TOOL RECORD, AND THE FINDING

Show how each interpretation is supported

A material finding should show how you moved from the prompt to the interpretation. Keep four layers separate: what the prompt supplied, what the tool recorded, what source inspection supports, and what you infer about the person or next action. [R01, R02]

Describe coverage accurately

Use statements such as “The sources inspected in this pass provide X and leave Y unresolved.” Record what was inspected, how it was found, and what remained inaccessible. A broader claim about information being absent requires a broader search and a clear account of its scope.

A source can appear in the retrieved record without appearing among the answer’s citations. Record that distinction directly. Relevance, credibility, repetition, snippets, and answer construction are possible explanations to investigate; the presence and citation flags leave the selection reason unresolved. [R06, R07]

Look for corrections and contrary evidence

Read sources that qualify a convenient claim as carefully as those supporting it. Check whether seemingly independent pages rely on the same original study. Preserve substantive differences between settings or versions. A disagreement may identify a condition that changes the evidence requirement rather than a contest one source must win. [R01]

Keep the work proportionate

Go deeper when an answer could change the next research step, conversation, verification, or publishing decision. A modest correction may already have enough support. Budget time for source reading as well as tool execution.

WHAT TO DO Preserve the sequence from supplied premise to tool behavior to source support to a provisional requirement or next action. State the unresolved step plainly.

6. WRITE EACH ROLE’S FINDINGS

Write a findings page that another person can check

The first deliverable is one evidence account for each represented responsibility. It should let another person see what the investigation found, what sources support it, and how tentative any proposed human requirement remains.

Use one role findings page within the Decision Evidence Record. Reference the common situation rather than repeating it. Keep the role clause and its evidence basis alongside the result.

State provisional requirements as provisional

A useful form is: “For [person or role] considering [judgment within the shared decision], the information may need to make [fact or distinction] verifiable under [conditions], because [source evidence or stated responsibility]. The remaining question is [uncertainty].”

Use stronger language when appropriate evidence already establishes the requirement. Keep the provenance visible. A public account can establish that a difficulty occurred. A technical source can establish a product condition. A review-policy document can establish a formal requirement. Their combination may support an inference; actual prevalence and applicability still require the evidence suitable to those claims. [R01]

State what the research added

Identify what changed in your understanding: the concern you supplied, any new distinction or source the tool introduced, what the source check confirms, and the question still unresolved. When the response repeats an uncertainty already in the prompt, record that as supplied context.

Classify a repeated prompt uncertainty as supplied context. Then identify the extra work the response contributes: a useful decomposition, external condition, source, or challenge. In Asteron, a role prompt can likewise contain the concern that later dominates the search. Treat that concern as supplied context; count as an addition the new distinction, source, method, threshold, or unresolved condition the research introduces. [R06]

A completed record can leave the human requirement hypothetical and still provide a useful result: a verified source distinction, a corrected description of the available evidence, or a precise question for one of the people involved. That is a valid endpoint for an information-only investigation.

WHAT TO DO Finish each role page with the strongest source-grounded statement, one next check, and the questions left open. Keep its information requirement and proposed intervention effect separately classified.

7. COMPARE FINDINGS AND PLAN WHAT TO SHARE

Put the responsibility findings together

Put the findings for the represented responsibilities alongside one another. Identify shared facts, different uses of those facts, and information someone will need from another person. Keep all of these findings connected to the same decision.

What is shared?

Several roles may depend on the same product fact, contract term, study, or test. A shared source can provide a common factual basis. Inspect whether the roles use the same passage and whether its conditions remain intact in each answer.

What does the baseline show?

The role-agnostic baseline retains the shared situation and evidence request. In Asteron, the Neutral conditions show what the same shared situation and decision moment produced without an explicit organizational role. Use matched Neutral comparisons to describe how the role-framed source set differed; separating the causal contribution of title, accountability wording, and other prompt features requires further tests. [R06]

What differs?

One person may need to establish that something is possible. Another may need to judge whether it fits their constraints. A third may need to confirm cost, timing, eligibility, risk, or whether an appropriate review occurred. These are different uses of information within one shared decision. [R01]

What information must be carried to the next person?

Name the recipient and what they need to receive: the supporting source, date, applicable conditions, limitations, open questions, and any relevant verification or approval. This transfer is a handoff. It can be a conversation or a document. Record the person’s actual need where known and identify any proposed need as a hypothesis. [R01]

Use these findings to select the next question to investigate. Existing information may already be adequate. The next useful step may be a conversation, a specialist check, a private record, or a decision to maintain the material already in use.

Write a summary someone else can examine

A useful readout states the shared decision, the responsibilities examined, the main source findings, the provisional requirements, what remains open about the people, and the unknowns most likely to change the decision. It also records where the available evidence is adequate and what would justify additional work.

WHAT TO DO Summarize what the comparisons and source checks established. Identify what someone else needs to examine, what remains uncertain, and the next question to resolve before proposing a change.

8. CHOOSE WHAT TO CHECK OR CHANGE NEXT

Choose the next action based on the question that remains

The investigation should now give you a more specific account of the shared decision: the responsibilities involved, the information found, what the inspected sources support, and what remains uncertain. Choose the next action according to the question it would help resolve and how the answer could affect the decision. Sometimes the available information is already adequate for the next conversation. [R01-R03]

Use the parts of this section that fit your next step. You may finish with a source check or a private discussion. Interviews, information production, and publication are possible continuations when they address a specific need; the original research remains useful on its own.

Use QFOs and inspected pages to prepare the questions

Start with the priority leads recorded in Step 5. These include recurring retrieved-but-uncited pages, especially where a QFO explicitly requested material associated with their domain. Use the passages you inspected to prepare questions for people who can explain a practice, substantiate a claim, describe relevant experience, or challenge an account. [R28, R29]

A QFO provides the exact mechanical request: perhaps a comparison, measurement method, condition, or documented experience. Reading the returned pages may reveal additional questions inside their text. Record these origins separately. A question raised by a forum discussion is a source-reading finding; it becomes an observed QFO only when the query log actually contains it.

For each lead, record what the search requested, what the inspected material already answers, and what remains unresolved for the shared decision. Keep the run ID, exact query, URL, and passage with your proposed question. Where query-to-URL linkage is unavailable, label the connection as your analysis of the run. [R06, R07]

Read answers and substantive disagreements as well as unanswered questions. A useful interview can establish the circumstances of a reported success, examine a disputed number, or compare a practice with another person’s experience. Frame the question so the respondent can qualify or contradict the original account. A respondent’s credentials do not verify the earlier account; their independently supported contribution needs its own record.

Include questions raised by participants, specialist knowledge, or your examination of the decision even when they are absent from the QFOs. Ask whether an answer could change how people assess their options. Record whether the question came from an observed search, a source passage, a participant, or an investigator’s proposed concern. [R29]

Ask people who can supply the missing information

Some questions concern the people making this particular decision. Their preferences, available time, commitments, and willingness to accept a trade-off need to come from them. Use the research to make that conversation specific and preserve each person’s ability to correct your account of their part in the choice. [R01]

Other questions concern how people have handled a comparable decision. Ask a subject-matter expert or experienced respondent about a particular event: what they needed to establish, which evidence they consulted, what uncertainty remained, what they chose, and what followed. Their experience may inform the present decision while its applicability remains a separate judgment. [R28, R29]

Follow up on dates, circumstances, direct involvement, and supporting records. Distinguish recollection, interpretation, and statements checked against documents. A person can describe a decision they took; a stronger claim about what caused its outcome needs further evidence.

For several interviews, use a shared set of questions about comparable circumstances, followed by questions specific to each respondent’s responsibility. Ask everyone what they needed to establish, what evidence they used, what they decided, what would make them reconsider, and who else contributed. Save the question set and record changes between rounds. This makes similarities and differences easier to examine without assuming every respondent faced identical conditions. [R29]

When contact is unavailable, preserve the source findings and specify the human question that remains open. The research can still be complete for the information-side question it was designed to answer. [R01, R02]

Asteron example: ask about actual funding decisions

Proposed continuation. The existing Asteron research records different questions and sources under the CMO, CFO, Legal, and Content prompts. The interview-program draft proposes asking people in comparable real organizations how they handled similar decisions. These interviews and the publication tests below have not been completed as part of the reported Asteron study. [R06, R28, R29]

A question for someone with marketing responsibility could be:

“Describe a recent decision about continuing or changing spending on AI visibility. What evidence did you bring to your colleagues, what did they question, and how did the recommendation change?”

For someone with finance responsibility:

“Describe a time you had to decide whether to authorize further spending while direct AI referral traffic remained small. What evidence did you ask for, what did you receive, and how did it affect the decision?”

A follow-up could ask:

“What happened at the next review, and which records would help another person understand whether the decision worked out as expected?”

These are proposed interview questions, rather than verbatim QFOs. The passage-level leads recorded in Step 5 can make them more specific. For example, an inspected disagreement about sales-team feedback could prompt the finance question about how that feedback was checked. Record the actual passage before treating that connection as part of the investigation.

Respondents could also review the synthetic Asteron situation. Keep a proposed response to that scenario separate from an account of something they actually did. Add any interview as a new, dated source; retain the original prompts, run records, and synthetic company facts unchanged.

Produce information suited to the people who need it

Begin with what someone needs to understand, verify, compare, or discuss. For a particular shared decision, the output might be a private summary, a comparison of options, an explanation of a disputed measure, or a record of conditions attached to a proposal. Existing material may need a correction or clearer organization. Keep sources and unresolved questions available to the people using it. [R01]

When interviews produce useful accounts and publication is appropriate, possible outputs include individual interviews, a report comparing responses to common questions, or an explanation of how people with different responsibilities assessed similar choices. The interview-program draft proposes these formats; their usefulness and citation performance remain to be tested. [R29]

State how respondents were selected, what they were asked, when interviews occurred, and how answers were edited or summarized. Preserve supporting records and the circumstances behind each comparison. Report counts with their actual denominator and describe findings as accounts from the participating respondents. Broader applicability needs additional support.

Agree on consent, attribution, confidentiality, and factual review before release. Credit original sources where their material contributes to the publication, alongside what the interviews independently add. Record sponsorship and editorial relationships. Protect personal information and each person’s ability to question the account. [R01, R19]

Publication can also make the information available to AI search

A published interview or report can make new evidence available to people working through similar decisions. A further outcome to test is whether the publication appears in results returned for later QFOs and whether its findings enter AI-generated answers. A source might be retrieved, cited, or explicitly attributed to its respondent or organization. These are possible outcomes of the proposed publishing approach. [R28, R29]

For Asteron’s proposed continuation, a finance interview might document how someone evaluated spending under uncertain measurement. A later research run could retrieve that account and cite it alongside broader studies or measurement guidance. State which question the publication addresses, what it adds to the inspected sources, and what a reader can establish from it.

The same possibility applies to an interview question developed from a retrieved-but-uncited forum page. A documented answer may add evidence about the underlying issue. If the new publication is later cited, that establishes a new citation outcome; it leaves the earlier page’s influence and the reason for its citation absence unresolved. [R06, R07, R29]

Check what happens after publication

Repeat the relevant original decision prompts, leaving the new publication unmentioned so the system can encounter it through its own research. Preserve settings where possible, record the date and any changes, and choose repetitions appropriate to the claim. Inspect the new fan-out: the exact original queries may not recur. Separately label any search that names or directly supplies the new publication as a targeted check. [R02, R29]

Preserve retrieved-but-uncited outcomes for the new publication too. Where query-to-URL linkage is unavailable, describe retrieval within the run rather than assigning it to a particular query. Similar wording alone leaves source influence unresolved. [R06, R07]

A repeat run records what appeared under those conditions. Establishing that publication caused a broader change requires an appropriate comparison and attention to other changes. Assessing whether the information helped people requires evidence from their use. Continue these investigations through AI Visibility and Information Design, respectively. [R01, R02]

Record the next action and how it will be assessed

Add the action to the existing Decision Evidence Record. State the question, the QFO or other evidence that prompted it, who will carry it out, what they will check or produce, who needs to use it, and how its value will be assessed. Link source passages and interview materials as attachments rather than creating another separate worksheet.

Keep established evidence separate from expected effects. A completed interview records an account; source checks may support particular claims. Publication, retrieval, citation, accurate representation, and usefulness to people each need their own observation.

Use the family’s action labels when helpful: RESEARCH FIRST for further reading, interviews, or specialist consultation; TEST for a defined change ready to evaluate; MONITOR for further observation; MAINTAIN / DEFEND for keeping useful information current; and PAUSE when other work is better justified. These describe the investigation’s next action. The people involved retain responsibility for their shared choice. For an AI Visibility handoff, PAUSE corresponds to DO NOT PURSUE YET. [R02]

WHAT TO DO: Identify the unresolved question that matters most to the decision. Record the query, source passage, or other reason for pursuing it; what existing evidence establishes; and who or what could answer it. Choose what to check or produce. When publication is appropriate, specify its intended use for people and the retrieval, citation, or answer-representation outcomes you will test.

WHEN PUBLISHING IS THE NEXT ACTION

Choose the publication route and maintain the record

Use this section after Step 8 identifies a publishable contribution: an interview, report, correction, reference, or independently examined account. Step 8 describes how to obtain and assess the information. Here, decide where it should appear, who can verify it, who decides what is published, and who will maintain it. A private conversation or record remains a valid outcome for other shared decisions. [R01, R02, R29]

In the Asteron teaching case, a possible contribution is a reference explaining how the program’s visibility measures were collected and what those measures establish. Another is a claim-by-claim account of product descriptors and their applicable conditions, checked by the responsible experts. These are proposed actions based on the modeled decision; their effects on people or AI answers remain to be tested. [R04-R07]

Choose a format suited to the evidence and its intended use: an interview, a comparison of documented accounts, a methods page, a policy, an implementation reference, or a scoped case. Explain the evidence source separately from the publication location. A company statement reproduced elsewhere remains company-originated; an independent publisher makes its own editorial judgment. [R02, R12, R17, R29]

Offer independent publishers material they can examine

Begin with the work their audience is doing. Identify the page, the question it helps answer, and the specific evidence you can add or correct. Useful contributions may include data, methods, equipment access, a technical demonstration, a documented case, or an expert who can explain conditions and limitations. [R15]

A practical opening is: “This person needs to establish ____ before deciding ____. Your resource currently supports ____. We can supply ____ to add or correct ____, with these conditions preserved: ____. Your team can assess whether it improves the resource.”

The factual owner verifies what is supplied. The editor or independent examiner decides what their publication can support. Preserve sponsorship, commissioning, and editorial-control relationships. The contribution’s value is the evidence it adds to the reader’s work. [R15, R19]

Preserve the different kinds of support

Where an external authority establishes a requirement, make its scope and its relationship to company evidence clear. Qualified specialists determine applicability. A governing source and a product test perform different jobs.

Community pages can supply questions and reported experiences worth investigating through the Step 5 callout. An interview or source check may add independently supported detail. Preserve both contributions and evaluate any effect on people or AI answers separately. [R01, R14, R29]

Keep the contribution route concrete

Name the page or record, the permitted contribution route, the person who decides what appears, the factual verifier, and the maintenance owner. Agree how future corrections and changes will be recorded. Recurrence in a source log helps prioritize inspection; access to edit or contribute remains a separate practical question. [R14, R15]

WHAT TO DO Use the placement and publisher views of the record to specify one useful contribution. Preserve its evidence basis and expected effect separately.

END WITH A CLEAR RECORD AND NEXT STEP

Record what is established, what remains open, and who acts next

The primary output is the Decision Evidence Record: one shared situation, the represented responsibilities, exact research conditions, source findings for each responsibility, a combined account of what remains to be established, and the questions still open about the people. In professional publishing work, an Evidence Placement Brief can follow when the record supports a specific contribution.

An Asteron readout could say: “The CMO and CFO L1 DECIDE runs each cited 12 URLs and shared one. We have identified their source lists and the different questions in their research. Reading those sources and checking the company’s own records are the next steps before using their claims in a funding discussion.” [R06, R07]

That statement identifies a recorded result and the work still required. Source findings, when added, need their own dates and supporting passages. Any proposed information requirement remains separate from evidence that the people actually hold it. [R01, R02]

Use the shared Evidence to Carry Forward block. Preserve an established result verbatim with its evidence state, provenance, scope, and limits. Enter the next requirement or expected effect as a separate hypothesis when appropriate. [R01–R03]

What the method is intended to provide

A person can begin with a plausible account of one shared decision, examine how a research tool approaches each represented responsibility, read the underlying sources, and leave with a clearer record of the information available and the questions still unresolved.

The people involved remain responsible for the judgment itself. The method provides a structured research record that can make the next conversation, verification, or information change more specific.

End when the record is sufficient for the next question or action. Preserve useful findings even when some human requirements remain unresolved. Record what would justify further work, who could do it, and which decisions remain with the people involved.

RESEARCH / WHAT THE STUDIES SHOW

Earlier studies recorded different search behavior under different prompt conditions

Observed / documented. Earlier studies varied intent, specificity, situation, and role framing and reported changes in searches and source selection. The pharmaceutical-services contrast and the repeated Role × Situation × Intent tests examine different manipulations, so their numerical overlaps retain their own units and scope. [R09, R11]

In Asteron, calculator-bearing conditions rose from 0 of 27 at L1 to 16 of 27 at L2 and 24 of 27 at L3. More detailed prompts supplied costs, revenue, margin, and measurement conditions; two high-knowledge decision traces included power-calculation expressions. [R06]

Interpretation. More complete conditions can make additional analytical operations available. The correctness and appropriateness of those operations remain separate questions.

Practice hypothesis. Use the added conditions to investigate what each role may need to establish. A resource that preserves those conditions is one proposed response; its effect remains to be tested.

Asteron recorded different cited source sets under different responsibility conditions

Asteron Industrial Systems is the frozen synthetic company used throughout this note and in the 81-condition experiment. Its internal roles evaluated an AI-visibility program with improving visibility, small direct referrals, and uncertain downstream results. [R04, R05]

The following questions paraphrase modeled responsibilities. They are neither verbatim executed prompts nor evidence from interviews with people doing those jobs.

Observed / documented. Mean exact cited-URL Jaccard overlap with matched Neutral conditions was 0.027 for Legal; the other role means ranged from 0.116 to 0.202. Legal searched for claims, substantiation, and enforcement; PR introduced disclosure, editorial independence, and professional or platform policies. [R06]

Jaccard compares shared items with all distinct items in two sets: 0 indicates no shared item and 1 identical sets. These values describe URL overlap, rather than reasoning agreement or human similarity. Legal also had the largest summed citation count. Evaluate matched set sizes alongside overlap; this edition leaves their contribution to the low mean unresolved. See “How we compare citation sets.” [R06]

Interpretation. Accountability, the wording that expresses it, supplied context, and available authority can all contribute. Asteron measures the constructed role/accountability prompts; it leaves the effects of bare titles, individual clauses, and exact-condition variation unresolved. Actual people establish their information requirements.

Asteron also provides a four-responsibility teaching slice

Observed / documented. At L1 DECIDE, CMO, CFO, Legal, and Content received the same shared situation and decision moment with different represented responsibilities. Their exact cited sets were materially different: CMO × CFO shared 1 of 23 distinct cited URLs (Jaccard 0.043); CMO × Legal shared none (0.000); CMO × Content shared 2 of 19 (0.105). Search activity also differed: 12 queries for CMO, 24 for CFO, 32 for Legal, and 14 for Content. These worked-record figures compare each represented responsibility directly with the CMO; Appendix G4 instead reports role-versus-Neutral means across nine matched conditions. [R06, R07]

Interpretation. Under these constructed conditions, the same proposed investment produced different research paths when the represented responsibility changed. For the Asteron teaching case, those differences identify questions and sources the CMO may want available when colleagues examine the recommendation. They remain observations about the prompts and recorded system, not evidence of what real executives require.

Scope. One run per condition. The full 81-condition matrix supplies the broader descriptive context. Repeatability, cross-system behavior, and human usefulness remain separate evidence questions.

In some prior tests, identifiable authoritative sources concentrated the cited source set

Observed / documented. In the cross-setting tests, the risk-review role drew heavily on formal sources in the cases classified as having an authoritative source for the concern. The source analysis used a ratio of canon pages to canon-plus-vendor pages. It reported values from 0.75 to 1.0 in those cases and 0.0 in the two software cases classified as lacking a comparable source. Other source classes sit outside that ratio, so it is a compositional measure rather than the share of all citations. [R10]

The software results were especially informative. The model cited vendor terms, export guidance, and related documentation. The available general review material did not become an equivalent cited substitute for those particular questions. [R10]

Interpretation. A useful source relationship is often specific to the thing being established. Vendor documentation can be primary evidence for published terms. An official source can establish a requirement. A test can establish an observed result under stated conditions. Those functions can complement each other.

Scope and next evidence. The classification of an authority being present or absent was part of the research design and analysis. It is not a legal finding that a risk is wholly unregulated. The studies also varied substantive topics and had limited repetition within each setting. The proposed explanation is a bounded pattern to test more broadly.

Retrieval is much broader than visible citation

Observed / documented. In Asteron, Reddit supplied 2,111 run-by-URL retrieval records and Wikipedia supplied 614. Neither domain had a cited record in that batch. The same findings report 214 retrieved records and 76 cited records for Pew Research Center, and 711 retrieved records and 82 cited records for arXiv. These are counts of source appearances across runs, not counts of distinct people or measures of influence. [R06, R07]

Earlier studies also recorded substantial community retrieval with little citation. However, the recovered record includes Reddit and Wikipedia citations in some prior conditions. Academic material likewise changed markedly by topic: a weakly cited source class in some product-selection studies became a recurring cited source in the Asteron AI-search discussion. [R09, R14]

Interpretation. Citation selection depends on more than encounter frequency. The relevance of the particular page, the question being answered, and the evidentiary job are worth inspecting.

Scope and next evidence. These observations identify neither a permanent source blacklist nor the internal reason for selecting one source over another. Credibility, applicability, redundancy, available snippets, and answer construction remain possible explanations. Establishing influence would require additional evidence beyond retrieval or co-occurrence.

In Asteron, broad recommendations converged while cited source sets differed

Observed / documented. The Asteron analyst’s answer screen reports that the runs directly asked to choose among funding actions converged on forms of capped, redesigned, better-instrumented continuation. The same report records no recommendation for immediate expansion or outright termination across its reviewed answers. Meanwhile, the cited source sets differed substantially by condition. [R06]

This is the analyst-runner’s reported answer screen. The archived corpus and mixed-review description limit detailed answer-category claims; Appendix J states those limits. [R06, R08]

Interpretation. Agreement at the level of a broad recommendation can coexist with different evidence paths. A bibliography comparison and an answer comparison answer different questions.

Scope and next evidence. The synthetic scenario supplied improving visibility, small direct referrals, and uncertain downstream effects. Those conditions may have favored bounded continuation. The study therefore gives an example of convergence; it leaves open how frequently this happens in other situations.

For the role-by-role investigation, a shared broad recommendation still leaves distinct work to support. In the modeled situation, Finance must judge the next expenditure, Legal must evaluate claims and controls, Content must choose roadmap changes, and PR must assess credible external contributions. Those are different decisions or tasks within the same initiative. The practical question is which evidence each person would need to make their own judgment inspectable. Human validation remains with the people doing that work. [R04, R05]

RESEARCH / HOW THE PROGRAM DEVELOPED

A sequence of questions built the research program

The early studies noticed different source selections around different role-framed questions. Later comparisons asked whether sector, responsibility, question specificity, or the available authority better explained those differences. Tests across consumer, international, and software settings added boundaries; version comparisons showed that the recorded system belongs in every finding’s description. [R09, R10, R12, R13]

In parallel, source-log analyses distinguished retrieval, citation, and practical contribution routes. Asteron then brought role, supplied knowledge, and decision moment into one matrix with matched Neutral conditions and a separate answer screen. The Product X specimen applied the resulting workflow to a small matrix from a frozen teaching case. Appendix A shows what the research families added.

Later studies changed several earlier interpretations

The research changed its working explanations. Job-title language gave way to closer examination of responsibility and specificity. Broad sector labels gave way to the authority available for a particular question. An expected fallback to independent review media was contradicted in the software tests, where vendor documentation was selected instead. [R09, R10]

Citation-rate summaries also became more precise as the studies distinguished counting units and source mixtures. Within Asteron, an early Neutral knowledge-ladder interpretation was retired after the full matrix was considered. The example prompted a further check: record unspecified context and synthetic identity before attributing differences to the roles. Appendix H preserves the useful corrections. [R06, R07, R09, R14, R20]

The Asteron design in plain language

The experiment used one frozen synthetic business record. Its $420,000 pilot cost, $120,000 typical first-year contract value, 48% gross margin, visibility gains, and uncertain downstream results were invented inputs. The observed result was how AI researched and answered those inputs. [R05]

Nine role conditions, including Neutral, were crossed with three knowledge levels and three decision moments: 9 × 3 × 3 = 81 conditions. L1 supplied a compact plausible account; L2 added intervention and measurement detail; L3 added material economics, conditions, and uncertainty. Respond asked about disruption, Interpret about the evidence, and Decide about justified action. [R04, R05]

Each prompt level retained the prior material and added context. A common evidence request supported comparison. After a six-condition Neutral/CFO pilot, the batch completed all 81 conditions on September 8, 2026, using the model recorded as gpt-5.6 with OpenAI API web search through Thinking Inspector, with documented coverage and preservation limits. There was one run per exact condition. Jobs ran concurrently, with ordering preserved within jobs. [R04, R06, R08]

What each analysis step added to the record

These steps improve interpretability; one run per exact condition still leaves stochastic variation unresolved. The tables reproduce documented analyses rather than a newly recomputed master corpus. Appendices D–J preserve definitions, methods, selected results, and material source limits. [R04, R06–R09, R20]

Use claim classes alongside the family evidence states

A claim class describes what kind of analytical work supports a statement: a count, comparison, recurrence, interpretation, or proposed application. An evidence state describes what is established. Provenance records where the support came from. These fields serve different purposes.

Appendix I records claim class, evidence state, and provenance for the principal claims. Preserve each supporting analysis’s scope and archive limits.

WORKING TOOLS / DECISION EVIDENCE RECORD

Set up one shared decision

Use one record for the situation. Repeat the role findings page as needed. Attach run logs and source details rather than squeezing them into the summary. Evidence state and provenance belong beside each material statement.

A useful record states the decision, responsibilities, source findings, provisional requirements, unresolved questions, and next check. Keep the basis for each hypothesis available.

WORKING TOOLS / ROLE FINDINGS

Keep the supplied responsibility and the learned evidence distinct

Complete one page per role, linked to the common situation record. Use concise statements and source or run references. Repeat the page when different questions require separate treatment.

Keep reading records and counting rules clear

Date the model’s recorded activity separately from your source inspection. Keep the counting unit, URL-normalization rule, set sizes, and comparison measure with the results. Appendix F explains the measures; Appendix K explains the TI records.

WORKED ROLE FINDING / ASTERON LEGAL

Use the Legal source pattern to prepare a more specific next question

This example uses the documented Asteron analysis. The company and role responsibilities are synthetic inputs; the reported tool behavior comes from the executed batch. It summarizes the nine Legal conditions, rather than presenting them as nine repeats of an identical prompt. [R04–R07]

WORKED RECORD / ASTERON DECISION FROM FOUR RESPONSIBILITIES

How the same Asteron decision looks under four responsibilities

This worked record uses existing Asteron L1 DECIDE runs. It follows the CMO as the person preparing the recommendation, then uses matched CFO, Legal, and Content conditions to show what different responsibilities pulled into the research. The dialogue and meeting language are teaching constructions; the query, surfaced, opened, cited, and run identifiers come from the recorded experiment. [R04-R07]

Setup

Shared decision: What should Asteron do with the next period of AI-visibility investment?

CMO’s working assignment: turn the current result into a recommendation that colleagues can interrogate from the responsibilities they own, improve together, and carry forward with its conditions intact.

What the CMO’s run makes visible

The CMO trace searched across AI-summary click behavior, AI referral traffic and conversion, B2B buying research, and information behavior among technical buyers. Its cited set included buyer research, AI-search measurement studies, Google documentation, and industry research. [R07]

A teaching formulation for the CMO is:

“This visibility result is useful. Help me turn it into a recommendation we can examine together. I need to show what we learned, what another commitment would require, and what questions my colleagues need to help us answer.”

That sentence is an instructional scene, not a verbatim Asteron output.

Finance changes the interrogation

The matched CFO run issued twice as many queries as the CMO run. Its searches emphasized zero-click evidence, referral traffic and conversion, B2B buying behavior, measurement and attribution, and evidence for further spending. Its cited set shared only one exact URL with the CMO’s set. [R07]

The resulting preparation question for the CMO is:

What would another period of investment need to deliver, and what evidence would let Finance judge whether the commitment remains worthwhile?

The research can help identify candidate measures, assumptions, and external evidence. Asteron’s own financial records and Finance’s judgment establish the company-specific answer.

Legal changes the interrogation again

The matched Legal run issued 32 queries and cited 18 pages. Its searches moved toward advertising substantiation, endorsements, review rules, native advertising, false advertising, platform policies, and AI-risk guidance. It shared no exact cited URLs with the CMO run in this matched pair. [R07]

The preparation question becomes:

Which claims are we asking the organization to repeat, what supports them, and which conditions or disclosures need specialist review before the recommendation travels onward?

Content changes what execution means

The matched Content run searched for AI-feature documentation, citation research, buyer research, technical-buyer information, and content implications. It shared two exact cited URLs with the CMO run. [R07]

The preparation question becomes:

What work would this recommendation create, what current commitments would move, and which evidence gaps are important enough to justify that change?

What the comparison shows and what it makes worth checking

The useful result is not that the roles “disagree.” The observed result is that the recorded system assembled substantially different supporting evidence under different responsibility clauses. The practice uses those differences to prepare a stronger shared examination.

What the CMO needs to carry into the discussion

For this teaching case, the CMO’s discussion record can include:

1. What the program result establishes, with its measurement conditions and limits.

2. What the CMO recommends doing next and the alternatives considered.

3. What the next commitment requires in money, time, people, and dependencies.

4. Which questions each colleague needs to examine from their responsibility.

5. Which conditions, disagreements, and unknowns the CEO and Board need to receive.

The CMO uses the record to invite thorough questions and improve the proposal. Colleagues may support the same action for different reasons, attach conditions, challenge it, or prefer another option. Preserve those judgments and their basis so the CEO can accurately explain the recommendation and unresolved issues to the Board.

Possible next step: collect accounts from people with relevant experience

The recorded differences give the investigator questions to inspect, rather than completed answers about real people. Step 8 proposes interviews with marketing and finance practitioners about comparable funding decisions. The CMO could use the resulting documented accounts to ask more specific questions of her colleagues. Suitability for Asteron’s circumstances, and any benefit to the discussion, would still need checking. [R28, R29]

The reported Asteron Reddit retrieval-and-citation pattern also supplies a practical inspection lead. Select relevant pages from the actual run records, read their contents, and connect any resulting interview question to its passage. The hypothetical thread in Step 5 demonstrates the procedure; it does not stand in for that reading. Any interview, publication, or follow-up AI test would be new work, with its own date and evidence.

WORKING TOOLS / EVIDENCE PLACEMENT BRIEF

An optional professional action view

Use this when a role finding supports a specific evidence contribution. Link to the Decision Evidence Record and carry existing findings at their established state. The expected effect remains a separate hypothesis.

Choose TEST, RESEARCH FIRST, MONITOR, MAINTAIN / DEFEND, or PAUSE as appropriate. PAUSE maps to AI Visibility’s DO NOT PURSUE YET code. A verified correction, an appropriate decision to maintain the current source, or a referral to another method can all be useful outcomes.

WHAT MUST REMAIN PROTECTED? Consider necessary scrutiny, accurate conditions, appropriate escalation, privacy, existing useful records, and editorial independence. [R19]

WORKING TOOLS / PUBLISHER CONTRIBUTION BRIEF

RECIPIENT VIEW. Draw this from the Evidence Placement Brief when the next action is a publisher conversation. Tailor the evidence to the resource and preserve editorial independence.

Opening sentence: “This person needs to decide or accomplish ____. Your resource helps them establish ____. We can supply ____ to add or correct ____, with ____ conditions preserved. Your team can assess whether it improves the resource.”

Providing evidence and obtaining editorial coverage are different outcomes. Carry each at the state established.

WORKING TOOLS / EVIDENCE TO CARRY FORWARD

SHARED FAMILY BLOCK. Copy the established result verbatim. Preserve its state and provenance; enter the next hypothesis as a new uncertainty. Field names follow the family’s full Evidence to Carry Forward block. [R01–R03]

Attach the relevant Decision Evidence Record and any selected Evidence Placement Brief. A proposed contribution can enter the receiving guide as a HYPOTHESIS-state intervention candidate while the starting observation retains its own established state.

APPENDIX A — RESEARCH PROGRAM TIMELINE

How the research program developed

This is a selected map of the research families that inform the note. It distinguishes early specimens, repeated comparisons, interpretive work, and production proposals. The recovery package inventories additional studies, including items located but not fully read during recovery. Those items remain discovery leads rather than primary support for this manuscript. Client situations use descriptive labels. [R09]

What counts as accumulation

A later study can add repetition, vary an alternative explanation, extend a finding to another setting, or identify a boundary. These are different contributions. The evidence inventory preserves that distinction rather than treating every row as an independent replication of the same claim.

The controlled work sits within one model vendor family. Some sources use “single-engine” as shorthand even when versions differ. This note uses the more explicit description: model family, recorded version, search surface, date, and situation. [R09]

APPENDIX B — THE ASTERON SITUATION RECORD

A synthetic company and a synthetic pilot

The following is a condensed account of the frozen experimental situation. All company figures and intervention outcomes in this appendix are invented inputs. The observed research result is the AI’s response to those inputs. The original situation record is identified as SYNTHETIC_ASTERON_v1.0. [R05]

Company. Asteron Industrial Systems sells industrial edge-computing gateways, industrial PCs, device-management software, and support. The scenario gives it approximately $220 million in annual revenue, a blended gross margin of 48%, and a typical first-year new-customer contract or deployment value of $120,000. The usual new-customer sales cycle is four to seven months. Buying involves commercial, technical, security, procurement, and implementation work.

Historical organic search. The prior full year has approximately 1.8 million organic visits, or 150,000 per month. Last-touch attribution credits organic with about 8% of new-customer pipeline. Multi-touch analysis shows organic appearing in about 22% of opportunities, explicitly as an association rather than a causal share.

The disruption. Organic visits fall 38% year over year. Tracked nonbrand impressions fall 6%, while average ranking across the top 200 tracked nonbrand terms changes by less than one position overall. Organic last-touch pipeline falls 24%; total new-customer pipeline falls 9% amid softer demand. Management suspects zero-click and AI-answer changes are important, with their contribution relative to other causes left uncertain.

The pilot. Asteron conducts six months of owned-evidence consolidation, an operated offsite practitioner knowledge base, selected work with publishers, practitioner-oriented resources, community participation, and tracking of a finite prompt set. These activities are part of the invented scenario rather than interventions executed by this research batch.

Cost and alternatives. The fully loaded pilot estimate is $420,000: $210,000 in external program spending, $150,000 in internal marketing and subject-matter work, and $60,000 in engineering, legal, analytics, and operational support. Continuing at similar intensity would require roughly $400,000–$450,000 for the following six months; expansion would require roughly $650,000–$750,000. Other uses of the money remain available.

Measurement setup inside the scenario. The pilot tracks 180 prompts across five cohorts and four AI answer surfaces. This fictional 180-prompt pilot is distinct from the actual 81-condition experiment examining responses to the scenario.

Uneven gains. Category questions improve most in citation and mention. Comparison questions still show competing sources selected despite relevant company evidence. Use-case answers improve in recommendation while retaining gaps around difficult conditions. Current-version consolidation improves accuracy, with some stale third-party descriptions remaining.

Claim conditions. The scenario includes externally repeated descriptors such as “fast deployment,” “lowest cost,” and “enterprise security.” Their evidentiary status varies. The company has no approved universal lowest-cost claim, and other descriptors require configuration-specific support. These details create substantive responsibilities for Legal and PR in the prompts.

Downstream observations. Routine clarification requests fall about 11%, alongside changed tagging and case mix. A small security-review comparison moves from 19 to 17 days using unmatched samples of 34 and 31 reviews. Win rate moves from about 24% to 25%, and reported sales cycle from 138 to 133 days. The scenario explicitly treats these as uncertain observations rather than established effects of the pilot. Support, implementation, retention, and expansion have no reliable intervention-linked improvement established in the record.

Concurrent changes. Paid digital spending rises approximately 12%; there is a product update, two trade stories, a competitor availability disruption, soft demand, AI-platform changes, and improvements in some internal records. These conditions limit causal interpretation of the fictional business outcomes.

Role decisions. Marketing considers allocation; the CEO considers priorities; Finance considers continued expenditure; Legal considers claim and governance controls; PR considers credible external work; Content considers the roadmap; SEO/AEO considers the next diagnosis and test; the Board considers oversight and materiality. Neutral expresses an organizational question without a named accountable role.

APPENDIX C — THE 81-CONDITION MATRIX

Three dimensions, one shared situation

The design crosses nine role conditions with three knowledge levels and three decision moments. Eight named organizational roles generate 72 conditions. Neutral adds nine matched baseline conditions. Each exact condition has one recorded run, R1. [R04, R06]

Every row contains the same nine combinations:

L1: a plausible compact account of the concern. L2: the L1 account plus intervention and measurement detail. L3: the earlier account plus material economics, uncertainty, constraints, and alternative explanations. “Knowledge” names the information supplied to the model; it is not an assessment of a real person’s expertise.

Each condition ID combines role, moment, and level: CFO_INTERPRET_L3, for example. A run ID joins the prompt, trace, sources, and answer. Each level/moment has one Neutral baseline. The original prompt workbook remains the exact replication instrument. [R04, R08]

APPENDIX D — PROMPT CONSTRUCTION

Preserve ancestry and make the changes inspectable

The protocol describes five conceptual components: situation facts; role and accountability; decision moment; the knowledge addition; and a constant evidence request. The Neutral conditions retain the organizational situation while leaving the explicit role identity out. [R04]

L2 was designed as an additive extension of L1, and L3 as an additive extension of L2. The purpose was to let the analysis ask which new behavior appeared when additional facts entered the question. The highest level supplied decision conditions rather than a prescribed bibliography or a list of specialist methods to invoke.

The constant close asked for current external evidence where relevant, citations to relied-on sources, a distinction between external evidence and inference about the synthetic company, and visible uncertainty where stronger conclusions lacked support. Thus, the experiment explicitly requested evidence; results describe that prompting condition.

The role conditions included responsibility-specific work. Consequently, a comparison with Neutral measures the constructed role/accountability prompt, including its relevant wording and concerns. A bare-title effect would require a separate manipulation.

Longer knowledge prompts also change length and the number of supplied premises. Additive construction reduces arbitrary rewriting, while leaving those features coupled to the knowledge manipulation. The reported comparisons should be read with that fact in view.

For reproducibility, the exact executed workbook version, model, evidence request, and run settings matter. For human-facing examples, condensed paraphrases should be labeled as paraphrases and kept distinct from the executed instrument. The Product X stress test used the Revision 3 template, with its executed wording preserved in the frozen prompt file. Use that file for replication; this edition’s adaptable template has been edited for broader use. [R20]

APPENDIX E — OBSERVABILITY AND TERMINOLOGY

What each record means

The raw Asteron schema has a RETRIEVED field containing NOT_OBSERVABLE. Its SURFACED field is true for each source-event row. This note’s terminology intentionally maps “retrieved” to that observed surfaced set, while preserving the raw schema’s distinctions. The Product X stress-test source table uses the same raw-state convention. [R07, R20]

The Asteron dataset records 140 opened run-by-URL rows, including 121 also marked cited. There are 621 cited rows without an explicit opened flag and 19 opened rows without a citation flag. Fourteen runs have no recorded explicit open. In the Product X specimen, 42 of 42 cited rows across four runs include 9 with an opened flag; the baseline’s one opened row was not cited. An explicit open is therefore an optional observed action rather than a required stage through which every citation passes. [R07, R20]

These counts establish recorded tool states. They leave open how source snippets, prior knowledge, and the model’s internal processing contributed to an answer. They also leave open whether a particular result was fully read in the ordinary human sense.

The findings report that URL-to-query linkage is unavailable at the individual source-row level. A run can have known queries and known source sets without a reliable record connecting every query to every URL. This note therefore describes run-level relationships and avoids assigning a specific query as the cause of a particular source appearance. [R06]

APPENDIX F — ANALYTICAL METHODS

Source-event normalization

The source-event artifact has one row for each run and original URL combination. It preserves role, knowledge level, decision moment, run identifier, original URL, mechanical hostname and domain fields, and source-state flags. The Asteron schema reports 81 run IDs, 9,690 rows, and 3,443 distinct original URLs. The Product X stress-test table reports 4 run IDs, 288 rows, and 241 distinct original URLs. Domain derivation is a mechanical heuristic; editorial source families are a separate analytical choice. [R07, R20]

Original URL strings and normalized groupings answer different questions. Exact-address overlap retains distinctions between pages and variants. Hostname or publisher-family views group sources more broadly. A reproducible comparison states which grouping it uses and how aliases, subdomains, and URL variants are handled, and does so before counting.

Descriptive measures and denominators

A run’s source count measures the size of its recorded source set. A domain’s recurrence can mean runs containing the domain or the number of run-by-URL rows from it. Those counts should be labeled separately. Calculator actions are separated from web-search queries when the trace supports that distinction.

Asteron’s reported distinct-source fraction is 213 / 3,443, approximately 6.2%. Its run-by-URL citation fraction is 742 / 9,690, approximately 7.7%. The former counts each address once across the batch. The latter counts an address again when it appears in another run. Neither fraction measures accuracy, influence, or human usefulness. [R06, R07]

Historical studies used pooled, per-run, domain-specific, and source-kind-specific denominators. Their percentages remain attached to their original units. This note treats their shared directional pattern as useful while leaving a harmonized pooled estimate for a separate analysis. [R09]

Matched comparisons

The Asteron analysis compares each role with Neutral at the same knowledge level and decision moment. It also compares levels within a role and decision moment, and decision moments within a role and level. Aggregating the resulting comparisons by role describes patterns within this matrix. It supplies neither independent repeat runs for an exact cell nor a population estimate across all possible prompts.

Sources used together and sources shared across answers

The comparisons in this note use two bibliometric ideas that answer different questions. Co-citation (Small, 1973) counts sources cited together within one answer and reads their co-occurrence as relatedness for that purpose. Bibliographic coupling (Kessler, 1963) counts the sources two answers share and reads that as similarity between the answers. The Jaccard figures in Appendix G and in the Product X specimen are coupling measures between answers. Within-answer co-citation remains a candidate analysis of the historical source-event rows. [R23]

This application compares deliberately generated answers. In a matched role contrast, the recorded system, shared situation, and evidence request are fixed while responsibility wording varies. The comparison describes those outputs; wider claims require further evidence.

Source-set overlap

Jaccard similarity divides shared items by the distinct items in the union. Report both set sizes. When useful, also report the overlap coefficient: shared items divided by the smaller set. In the Product X stress-test specimen, reviewer and Procurement have reported values of 0.176 and 0.375, from three shared URLs. In the separate Asteron study, Legal has 121 summed cited run-by-URL appearances, compared with 69-92 for the other roles. Those totals alone leave the contribution of set size to the matched mean unresolved. Preserve the counting unit and normalization rule. [R06, R20]

Low overlap can coexist with agreement on a claim. High overlap can coexist with different interpretations. Answer review therefore remains separate. Both-empty citation sets should be recorded as an absence of cited evidence rather than silently treated as evidence agreement; the present selected tables concern nonempty reported comparisons.

Query analysis and source classification

The analyses examined query vocabulary, domain-scoped searches, and recurring source families. Token overlap describes wording. A semantic interpretation of what the query seeks is a further analytical step. The older co-query work explicitly distinguishes vocabulary, query purpose, and citation structure, and reports limitations in its heuristic labels and pooled role mapping. [R16]

Source-kind categories such as vendor, official, community, and review media were study-specific classifications. The page’s function can differ from the publisher’s general identity. Source-kind comparisons therefore require their original coding rules and local scope. Instrument-supplied site-type labels are retained as-is and not used for source-kind claims.

Answer screen and synthesis

The Asteron report describes a sentence-level screen plus manual review of early outputs. Its aggregate conclusion is that funding-choice answers favored bounded continuation. The available archive and mixed-review details limit how strongly fine-grained category counts can be used. This note carries the summary as the analyst’s reported finding and gives the unresolved items in Appendix J. [R06, R08]

The final synthesis relates source behavior to proposed publishing choices. It supplies hypotheses for intervention selection. Establishing an intervention’s effect requires a baseline, an appropriate comparison, and a record of other relevant changes, using the AI Visibility or Information Design method as appropriate. [R01, R02]

APPENDIX G — SELECTED PRIMARY RESULTS

G1. Asteron coverage

gpt-5.6; OpenAI Responses API web search; September 8, 2026; one run per exact condition. Reported counts from Full-Matrix Findings and Source-Event Schema v2. [R06, R07]

G2. Role-level source counts

Summed run-by-URL appearances across nine conditions per role; an address is counted again in another run. Source: Full-Matrix Findings, section 1. [R06]

G3. Supplied knowledge and decision moment

Calculator figures count runs with an action. Source counts are rounded per-run means. Source: Full-Matrix Findings, sections 1 and 5c. [R06]

G4. Matched cited-URL overlap with Neutral

Each value is the mean Jaccard similarity over nine matched role-versus-Neutral pairs, holding knowledge level and decision moment fixed. Source: Full-Matrix Findings, section 3b. [R06]

Descriptive R1 summaries: role trajectories differ, and exact-condition repeatability remains to be measured.

G5. Selected retrieved and cited source appearances

These are the domain groupings and run-by-URL counts reported by the analyst, reproduced rather than independently recounted for this edition. Source: Full-Matrix Findings, sections 4 and 6a. [R06]

G6. Asteron L1 DECIDE teaching slice: cited-set coupling

Unit: exact cited URL strings from the recorded L1 DECIDE runs. These values are descriptive of one run per condition. [R06, R07]

APPENDIX H — HOW OUR INTERPRETATIONS CHANGED

Five interpretations that changed as more evidence was added

From role label to specified work. Early observations associated different role prompts with different source sets. Specificity and accountability probes then showed that the content of the question matters alongside the label. The public practice therefore begins with the required proposition and responsibility, while treating title-only personalization as a separate hypothesis. [R09]

From sector label to the particular evidence requirement. Early industrial-versus-clinical comparisons suggested different source mixtures. Later committee and question-level tests showed important differences within those settings. The public practice assigns source roles question by question. [R09, R11]

From expected review-media substitution to an observed vendor-documentation route. The software tests contradicted the expectation that the risk-review role would automatically use general independent media when a canonical source was absent. It selected vendors’ own documents. The effect of purpose-built independent examination remains open. [R10]

From a universal citation percentage to defined local rates. Historical summaries used a narrow approximate band, but the underlying studies mixed denominators and source kinds. Asteron reports distinct-URL and run-by-URL fractions separately. The public practice preserves those units and treats source selection as local to the observed question and environment. [R07, R09, R14]

From role-only interpretation to inspection of the shared record. The Product X stress-test specimen combined an unspecified governing regime and a synthetic product; all 73 recorded search queries were site-scoped to formal-source hosts, and no vendor, trade, community, or review-media page surfaced. Those shared inputs are plausible contributors to that pattern; their separate effects remain untested. In a live project, name the governing context when it is known and remember that a real product identity makes product-specific documentation available to the research system. [R20]

Two additional cautions retained in this edition

The source record documents model-version changes and a pilot interpretation that failed to hold in the full Asteron matrix. Both argue for keeping versions, dates, and the exact compared conditions visible. [R06, R09]

Earlier analytical prose sometimes treats retrieval as influence or citation absence as rejection. This note keeps those as possible interpretations to investigate. Its reported observations remain retrieval, explicit tool actions where available, citation, and answer content.

APPENDIX I — CLAIM REGISTER

I1. Different prompt conditions produced different source sets

EVIDENCE STATE OBSERVED / DOCUMENTED — reported comparisons of constructed conditions; causal explanations remain separate.

Claim class: reported controlled comparison. Support: Asteron matched contrasts, the Product X specimen, and the earlier role/situation/specificity studies. Scope: selected constructed prompts, settings, and recorded models. Boundary: the role/accountability condition bundles several meaningful wording differences; exact-condition stochastic variation remains. Next test: replicate the fixed cells and vary responsibility clauses separately. [R06, R09, R20]

I2. An identifiable authoritative source can organize a risk-review answer’s evidence

EVIDENCE STATE INFERRED — a supported interpretation of the documented cross-setting pattern.

Claim class: recurring cross-setting pattern and supported interpretation. Support: repeated committee observations and cross-setting tests. Scope: the particular concerns and classifications studied. Boundary: one situation per comparison setting in the broadening work; the authority ratio excludes other source categories; the source-availability classification is not a legal determination. Next test: vary the relevant concern and authority availability within a carefully matched design. [R10]

I3. The recorded retrieved environment is broader than the cited set

EVIDENCE STATE OBSERVED / DOCUMENTED — source records and reported counts under defined units.

Claim class: documented observation recurring across studies. Support: source logs and reported counts, including Asteron’s 9,690 / 742 run-by-URL comparison and the specimen’s 288 / 42. Boundary: explicit retrieval definitions and denominators vary in the historical program; influence remains unmeasured. Next test: harmonize source units before pooling numerical rates. [R07, R09, R20]

I4. Community retrieval frequently coexisted with little citation in the studied samples

EVIDENCE STATE OBSERVED / DOCUMENTED — a recurring pattern in the named samples, with exceptions.

Claim class: recurring observation with exceptions. Support: early specimens, stress tests, industrial source analysis, and Asteron. Boundary: earlier Reddit and Wikipedia citations are recorded; results are conditional on topic, prompt, model, and period; the Product X specimen recorded no community sources under its particular input conditions. Next test: replicate across different topics and search implementations. [R06, R09, R13, R14, R20]

I5. Supplied situational facts were associated with additional calculator use

EVIDENCE STATE OBSERVED / DOCUMENTED — calculator-bearing run counts; correctness and usefulness require further examination.

Claim class: within-matrix observation. Support: Asteron L1 / L2 / L3 run counts. Boundary: fact content, prompt length, and the number of available premises vary together; calculator use does not certify the resulting analysis. Next test: replicate and separately assess correctness and decision relevance. [R06]

I6. Different cited sets coexisted with a shared broad funding recommendation

EVIDENCE STATE OBSERVED / DOCUMENTED — the analyst-reported screen and source comparison, with the archive and coverage limits retained.

Claim class: analyst-reported answer screen combined with source comparison. Support: Asteron funding-choice summary; the same pattern in the four Product X stress-test answers. Boundary: one synthetic scenario per study, mixed answer-review descriptions, six pilot texts pending in the last preservation report, and unresolved detailed category totals. Next test: complete a uniform archived-answer coding pass, then test another scenario. [R06, R08, R20]

I7. Frequently retrieved sources differ in how they can be responsibly changed

EVIDENCE STATE INFERRED — possible contribution routes inferred from source inspection and the recorded practice analysis.

Claim class: source inspection and practice analysis. Support: industrial addressability work and the Offsite Targeting Instrument. Boundary: a route to contribute is specific to a page, policy, relationship, and time; opportunity mapping is distinct from measured intervention success. Next test: record completed contributions and their observed effects. [R14, R15]

I8. Evidence placement should follow the proposition and credible source relationship

EVIDENCE STATE HYPOTHESIS — the proposed placement practice and its expected effects require application and testing.

Claim class: practice proposal. Support: synthesis of the findings with Citation Labs’ Information Design and publishing disciplines. Boundary: the research informs the proposal; it does not establish that a particular asset or contribution will improve citations or human work. Next test: use the proposal in bounded applications and measure the outcome through the appropriate guide. [R01, R02, R09]

I9. A tool-led responsibility comparison can produce a more inspectable record of the information around a shared decision

EVIDENCE STATE HYPOTHESIS — the applied workflow’s usefulness requires evaluation in live use.

Claim class: practice adaptation. Basis: Information Design, AI Visibility source inspection, the documented role-comparison studies, and the Product X method stress test with its later source-check addendum. Proposed scope: one consequential shared decision, with represented responsibilities and declared tool coverage. The broader use in household and other settings is a proposed application, distinct from the original business and product studies. The intended output is a source map, verified source findings after inspection, provisional information requirements, unresolved questions, and a justified next action. Improved decisions and the needs of real people require separate evidence. Next test: apply the method to live organizational and non-work decisions and compare the records with an unstructured research pass. [R01, R02, R04, R06, R20]

I10. A matched Asteron decision slice produced strongly different cited sets across responsibilities

Observed. CMO, CFO, Legal, and Content L1 DECIDE conditions shared the same Asteron situation, decision moment, system, and knowledge level while varying the represented responsibility. Pairwise exact-URL Jaccard ranged from 0.000 to 0.105 among those four runs. [R06, R07]

What this supports. Role-conditioned source comparison can reveal materially different evidence paths around one shared investment question.

What remains open. Repeatability, cross-system behavior, causal attribution to particular prompt components, and whether the resulting preparation actually helps people deliberate.

I11. Recurring uncited retrieval can guide questions for interviews and new information

EVIDENCE STATE HYPOTHESIS for the value of prioritizing these leads and for the effects of any resulting publication. Query strings and source-state counts remain OBSERVED / DOCUMENTED where the run record supports them.

Claim class: proposed research and information-production practice. Basis: Asteron’s retrieved-but-uncited pattern, query records, and the interview concept and panel drafts. The proposed route is to inspect pages, identify a material question, obtain independently supported accounts, and evaluate the information produced. The proposal does not establish why an earlier source was uncited, whether it influenced an answer, or that an interview or aggregate report will be cited. Next test: preserve the lead-to-question record, assess what interviews add, and examine reader usefulness and later AI retrieval, citation, and representation separately. [R06, R07, R28, R29]

APPENDIX J — DATA, PROVENANCE, AND SOURCE NOTES

J1. Evidence availability

The Asteron trace registers and findings identify 81 completed conditions. The source-event QA document reports full source-row coverage, consistent per-run surfaced and cited counts, and a preserved original JSON. The most recent answer-preservation report in the material reviewed for this edition records 75 archived answer texts, with six Neutral/CFO interpretation answers pending export. The analyst reports having reviewed those six earlier. [R06–R08]

The four-condition stress-test record (PRODX_R3_4C_B1) is preserved as: the full Decision Evidence Record; the frozen prompts; and the 288-row source-event table with SURFACED / OPENED / CITED flags and RETRIEVED marked NOT_OBSERVABLE. The original preservation report leaves four verbatim answers in TI under recorded run IDs. The later operator check confirms full text in a response but no native file-export command on the inspected MCP surface. An out-of-band export or authorized client-side capture remains to be demonstrated. [R21] The instrument’s site-type field is a heuristic label retained as-is and not used for source-kind claims. [R20]

This note reproduces selected documented results and interprets the research analyses. It does not provide a newly recomputed historical master dataset. Exact raw prompt and answer archives are distinct from the selected tables in this document.

Client identities are replaced with industry descriptions. Source titles below are descriptive where the internal original names a client. Raw source lists, prompts, and filenames can reveal identities, so their public release requires an additional disclosure review. The current source notes provide a traceable research inventory without linking readers into those identity-bearing records.

J2. Items requiring reconciliation for a public data release

Search totals. The Asteron findings’ headline reports 1,323 combined query entries and 46 calculator calls. Its nine seat query totals sum to 1,343, and the register-based analysis in the working record reports 47 calculator calls. This note leaves the disputed combined total out of its result tables. Calculator-bearing run counts are a separate measure.

Answer-screen coverage. The findings describe 75 stored answers plus six manually reviewed pilots, while a later limitations paragraph describes 12 early outputs screened manually. The listed expansion categories sum to 75. The document therefore uses the bounded funding-choice summary and attributes it to the analyst, rather than publishing an exact 81-answer category distribution.

Opened flags. The source-event QA notes disagreements between compact counters and per-URL flags for five conditions. It preserves the per-URL flags and records restoration of affected flags from an inline collection result. The schema reports that original URLs and cited flags were unaffected.

Product X specimen normalization. The archived stress-test record counted exact URL strings; one regulatory landing page appears with and without a tracking parameter. A normalized rule could change that specimen’s overlap figures. Appendix G6 reports the separate Asteron L1 DECIDE comparison. Preserve each study’s original counting rule when reproducing its results. [R20]

Historical denominators and scope. Early study summaries include differing aggregate counts, variable definitions of “read,” and heterogeneous pooling. The cross-setting authority brief also uses differing scope counts in its table and summary wording. This note names the individual example settings and preserves study-specific denominators rather than creating a new pooled claim.

These limits apply to the identified numbers and interpretations. The remaining results are reported at their documented scope. This publication presents selected reported findings; it makes no claim to supply a newly recomputed or publicly released raw corpus. [R06–R10, R20]

J3. Source notes

[R01] Citation Labs. Information Design for Humans at Work. 2026 working manuscript. Purchase-decision application, people and responsibilities, source inspection, provisional requirements, evidence boundaries, and the shared family handoff. Supplies the parent method and the Industrial Gateway Security Review teaching case.

[R02] Citation Labs. AI Visibility. 2026 working manuscript. Exact prompts and measurement conditions; query decomposition; source inspection; retrieval, citation and representation; observation limits and intervention testing. Supplies observation methods adapted to the role-by-role investigation and the subsequent AI-measurement handoff.

[R03] Citation Labs. Tracing Business Impact Beyond the Click. 2026 working manuscript. Organizational consequence, evidence strength, specialist ownership, and carrying bounded findings forward. Used for the organizational-impact boundary.

[R04] Citation Labs. Role × Knowledge × Decision Co-Citation Experiment: Research Brief and Agent Execution Protocol. September 8, 2026. Research questions; matrix; Neutral control; knowledge definitions; prompt ancestry; constant evidence request; independent-request execution requirements. Design document. The title’s “co-citation” is retained as the historical name; see Appendix F for the distinction this note now draws.

[R05] Citation Labs. Canonical Synthetic Situation Record: Asteron Industrial Systems. Approved v1.0, September 8, 2026. Frozen scenario SYNTHETIC_ASTERON_v1.0. Primary source for the invented business facts and role-specific decision horizons.

[R06] Citation Labs / Claude analyst-runner. Asteron Full-Matrix Findings: 81 of 81 Conditions. September 8, 2026. Batch RKD_V1_B1. Sections 1–5: reported counts, overlap, source gating, query vocabulary and calculator use. Section 6: retrieved-but-uncited counts and answer screen. Sections 7–9: interpretation, limits and subsequent questions. Quantitative discrepancies are listed in J2.

[R07] Citation Labs / Claude analyst-runner. Asteron Source-Event Dataset: Schema and QA. Version 2, September 8, 2026. Row definition, 15 fields, counts, joins, raw-state meanings, explicit-open anomalies and restored flags. Primary support for source-event coverage and denominator definitions.

[R08] Citation Labs / Claude analyst-runner. Asteron Run Index v2; Stage 0 Observability Verdict; final-answer preservation reports. September 8, 2026. Execution and archive provenance. The index and observability document establish the recorded environment; preservation reports describe answer-file coverage and checksums.

[R09] Citation Labs / Claude. AI Evidence-Environment Research Program: Prior Study Recovery and Reconciliation. September 8, 2026. The user-supplied recovery package inventories 26 research entries, identifies source documents and run references, and separates observation, interpretation, framework synthesis, and practice implication. It is a secondary reconstruction; items marked located-only retain that status.

[R10] Citation Labs. Cross-setting authoritative-source tests: Canon-Floor Reviser Brief. Version 1.0, July 15, 2026. Sections 2–4: comparison design, reported authority ratios, software boundary cases and vendor-documentation selection. Section 7: model, situation, repetition and classification limits. Interpretive brief grounded in named jobs.

[R11] Citation Labs. Pharmaceutical-services recommendation-intent comparison: Canon-Density Test 1 Results. July 3, 2026. Five recommendation prompts on gpt-5.5, compared with an earlier definitional/decisioning baseline. Per-run vendor/canon/other coding and a reported approximately 93% vendor share. Client-identifying title replaced by a descriptive title.

[R12] Citation Labs. Site-Operator Probe: Source Owner, Verification Load and Substitute Sources. Design dated July 13; results appended July 14, 2026. Controlled tests A–C and their bounds; nameable-owner comparison; source mix in the substitute-source test. The design’s proposed outcome test is distinct from the reported completed tests.

[R13] Citation Labs. QFO Stress-Test Synthesis. Version 1.0, July 2026. Seven exploratory batteries and a dated rerun, generally one replicate per prompt. Reports search variation, source overlap, international source selection, and exceptions. Its broader “law” language is treated here as a proposed interpretation of the recorded tests.

[R14] Citation Labs. Industrial-computing source logs: Consideration Frequency, Citation Selection by Kind, and Addressability. Findings 15, July 23, 2026. Four situations, reported 40 runs per situation, matched recorded versions. Source recurrence, local source-kind rates, and possible contribution routes. The document explicitly identifies its trust-display account as a hypothesis.

[R15] Citation Labs. Offsite Targeting Instrument. Version 0.1, July 23, 2026. Proposed queue-building method combining source recurrence, citation behavior and addressability. Used as prior practice design, with retrieval kept separate from the influence interpretation present in the original.

[R16] Citation Labs. Role-to-Role Co-Query Analysis: First Findings. Version 0.1, July 2026. Query-term overlap and an exploratory intent-profile recomputation. Contains explicit caveats about pooled role mappings, heuristic labels and unmatched citation-layer comparisons.

[R17] Citation Labs. Evidence-Site Template Kit. Version 1.3, July 2026. Production guidance on primary-source records, verification, page purpose, provenance and measurement. Used as practice background; production examples and design proposals are separate from experimentally established effects.

[R18] Citation Labs. The Three-Layer Funnel: Co-Query, Co-Read, Co-Cited. Version 0.1, July 13, 2026. Source-resolution discussion and methodological vocabulary. This note retains its domain/page/proposition distinction while using the newer operational retrieval definition and treating inferred rejection or influence as hypotheses.

[R19] Citation Labs / supplied editorial draft. Evidence Placement Brief and Crosswalk Draft. Working-record proposal, sections A–F and the family-case illustration. Used for onward-recipient, protection, contribution, and expanded-record fields. Proposed source-to-surface and claim-state crosswalks were evaluated separately; the note retains flexible surface relationships and statement-specific evidence states.

[R20] Citation Labs / Claude analyst-runner. Four-Condition Method Stress Test: Decision Evidence Record PRODX_R3_4C_B1. September 9, 2026. First execution of the published workflow: setup, four role pages, cited-set and surfaced-set coupling, stress-test log S1–S9, and Evidence to Carry Forward. Source-event table (288 rows) and frozen prompts attached. Source inspection was pending in that original pass; the later source-check addendum is preserved with Revision 5 and the supporting Product X materials. Appendix J5 in this edition concerns Asteron. Verbatim answers were reported held in TI under recorded run IDs.

[R21] Citation Labs / Claude operator. Thinking Inspector MCP – Briefing for the Assistant, with dated response examples and operator confirmation, 9 September 2026. Sections 5a-5b supply the single-run and URL-grouped response shapes and visible-interface export limitation. These are operator-reported checks; this revision inspected the briefing rather than executing TI. Earlier conflicting summary wording is reconciled in Appendix K.

[R22] Model Context Protocol. Architecture overview, versioned documentation accessed 9 September 2026. Used for the general connection explanation. Official architecture reference.

[R23] Small, H. (1973). Co-citation in the scientific literature: A new measure of the relationship between two documents. Journal of the American Society for Information Science, 24(4), 265-269. DOI 10.1002/asi.4630240406. Kessler, M. M. (1963). Bibliographic coupling between scientific papers. American Documentation, 14(1), 10-25. DOI 10.1002/asi.5090140103. Publisher records checked 9 September 2026. These supply lineage rather than validation of this AI application.

[R24] CISA and partner agencies. Secure by Demand: Priority Considerations for Operational Technology Owners and Operators when Selecting Digital Products, January 2025. Secure communications section. Read on the official Australian co-publisher’s Cyber.gov.au copy, 9 September 2026. The original CISA-hosted PDF was unavailable in this check.

[R25] NIST / NCCoE. SP 1800-16, Volume B, Securing Web Transactions: TLS Server Certificate Management. Sections 1.2, 5.1.1, and 5.2.2. Official HTML volume. Inspected 9 September 2026.

[R26] NIST. SP 800-82 Revision 3, Guide to Operational Technology (OT) Security, September 2023. Section 5.2.3.1, printed page 73 (PDF page 90). Official PDF. Inspected 9 September 2026.

[R27] FDA. Computer Software Assurance for Production and Quality Management System Software, issued 3 February 2026. Cover, section III, and section V.A (printed pages 6-7). Official PDF. Inspected 9 September 2026.

[R28] Citation Labs. Interviewing the People the Machine Can’t Find. Concept note, September 2026. User-supplied proposal to derive interview questions from recorded AI searches, publish attributable accounts, and repeat the research after publication. Used here as a practice proposal. Claims about unmet demand, source influence, preferred source formats, and guaranteed citation are hypotheses rather than established results.

[R29] Citation Labs. The AI Visibility Decision Panel – Program Skeleton. Draft v0.2 (B2B + B2C), 10 September 2026; supplied as Role-Holder_Evidence_Panel_Skeleton.md. Sections 3-5, 7-8, and Appendix B inform question derivation, common and responsibility-specific interview questions, publication options, retesting, and inspection of retrieved-but-uncited pages. The program is proposed. Uncollected market/role variants, staffing and timing estimates, claimed gaps, and predicted publication effects retain that status; the historical counts in this note remain attached to their original study sources.

J4. Research still needed

A repeated Asteron batch would help estimate exact-condition variation. A carefully matched change in the concern assigned to Legal would help test how accountability and available authority interact. A uniform archived-answer screen would strengthen the convergence finding. Cross-system studies would examine the vendor-family boundary. Completed publishing interventions, followed by appropriate measurement and human-work validation, would test the practice proposed here.

The Product X stress-test specimen leaves four design questions worth testing separately: exact-condition variation; the effect of specifying the governing regime; the effect of synthetic versus real product identity on available source routes; and cross-system behavior. Those are supporting-method questions. Asteron’s next-evidence queue in J5 concerns the investment-decision example itself. Application of the workflow to a live decision should also test whether the record’s yields, especially the questions it leaves open, are more useful than an unstructured tool pass. [R20]

The Step 5 priority rule and Step 8 interview-and-publication continuation require their own evaluation. Preserve the exact search/source lead, what an interview independently adds, and how people use any resulting material. A later citation of that publication establishes a new result; it does not by itself test the original source’s influence or explain its citation absence. [R28, R29]

J5. Asteron source-inspection queue for the worked example

The Asteron worked record above is grounded in the executed trace. A publication-grade source-reading pass can deepen it without changing the original run history.

Priority checks:

1. Pew Research Center, 22 July 2025. Read the study method and exact measured behavior before carrying its click-rate findings into the CMO or CFO account.

2. Google Search documentation and performance-report materials. Separate what Google documents about measurement and eligibility from claims about commercial impact.

3. Gartner B2B buyer research. Preserve the surveyed population, date, and wording before using the result to support a finance or executive discussion.

4. FTC substantiation and disclosure sources in the Legal run. Record the specific proposition each source can establish and route applicability to the appropriate legal reviewer.

5. Technical-buyer and content research in the CMO and Content runs. Separate publisher-reported audience behavior from Asteron’s own customer evidence.

Also inspect relevant retrieved-but-uncited pages from the selected Asteron runs, starting with recurring sources and checking for explicit source naming in their QFOs. Save the exact query, URL, passage, counting scope, and interview question derived from it. This is a proposed inspection task, not a completed reading of the Reddit corpus. [R06, R07, R29]

Each later source check should record the inspection date, exact passage or data relied on, conditions, and what remains unresolved. These later readings enrich the source layer while leaving the September 8 run record unchanged. [R06, R07]

APPENDIX K – USING THINKING INSPECTOR WITH CLAUDE

K1. Understand the setup

A worked implementation of the role-by-role method

This appendix describes the Thinking Inspector (TI) interface reported by the Claude operator on 9 September 2026. It uses the revised briefing and its sanitized response examples. Treat it as a dated implementation account: discover the tools available to your connection before using the procedure. The main method follows capabilities and can be used with another instrument. [R21]

Claude helps organize and operate the investigation. TI submits approved prompts to the selected OpenAI model with API web search and returns recorded activity. The model and search service perform the research under study. The investigator compares the records and reads the sources. The MCP connection is how the assistant application accesses TI’s functions. [R21, R22]

Two useful starting points

Recover an existing investigation. Use stored run IDs to revisit answers, source selections, and recorded queries. A different analytical question can often be investigated using the same recorded runs.

Prepare a new comparison. Freeze a situation and prompt set, obtain approval for the execution settings and spend, then submit new runs. Repeating an existing prompt is another execution, with a new record to preserve. [R21]

What the seven reported tools do

These names and behaviors are documented in the operator briefing. The response examples below illustrate particular successful calls; optional fields and error cases require the current tool definition. [R21]

WHAT TO DO Choose an existing-run analysis or an approved new execution. Keep the distinction visible in the record.

K2. Check access and settings

Establish the connection you actually have

Obtain TI access and the approved connection instructions from the service administrator. Confirm the assistant environment, authentication, permissions, charges, and handling of stored prompts and answers. The operator’s schema check leaves those account and product facts unresolved. This appendix therefore begins its concrete procedure from a configured connection. [R21]

Ask the assistant to identify the tools exposed by that connection and their current input definitions. A read-only recovery of one known run is a useful connection check. Record the date and which capabilities work. Keep credentials in the approved connection mechanism.

Make execution choices explicit

These are settings to check, rather than a universal prescription for a model or repetition count. Additional repeats support a different claim from a first exploratory result. The appendix leaves connection-wide charges and retention terms to administrator confirmation. [R21]

WHAT TO DO Save a dated settings card before execution. Ask the operator to explain any material default that is absent from the returned metadata.

K3. Recover and examine existing runs

Start with one complete result

Locate a known job or stored run. Save the job ID and every selected run ID. Retrieve each individual result and match its exact query prompt to your condition. Build a crosswalk among your role name, condition, replicate, original run ID, and any temporary comparison label such as A or B. That mapping belongs to the investigation. [R21]

A single-run response already includes a sources array with each URL’s opened and cited flags. Use it to examine one role before assembling a larger comparison. The operator supplied the following response shape from a successful call. Placeholder strings and omitted fields make this an excerpt, rather than a complete data export. [R21]

{
  “role”: “A”,
  “run_id”: “<stored-run-id>”,
  “query”: “<complete original prompt>”,
  “model”: “gpt-5.6”,
  “effort”: “medium”,
  “status”: “completed”,
  “output”: “<complete markdown answer>”,
  “sources”: [
    {“url”: “<original URL>”, “title”: “<title>”,
    “domain”: “<domain>”, “cited”: true, “opened”: false}
  ]
}

The dated example also includes created_at, usage, error, counts, and optional reasoning. Preserve the actual returned values; use the current schema to interpret missing fields. [R21]

Select the comparison view for the question

Use compact results to locate questions. Use complete relevant records to support whole-set counts, rare-source findings, or claims about a particular answer. Retrieve missing data from the stored run rather than reconstructing it from an assistant’s summary. [R21]

Existing Asteron teaching-slice identifiers

Use existing runs when learning the recovery workflow; this avoids creating new model executions merely to practice the instrument.

The Asteron run index is the authoritative join key across prompt, query, source, and answer records. The historical batch used gpt-5.6, one run per condition, and TI profile 2026-06-03. [R06, R07, R21]

K4. Read the returned data correctly

One run and one comparison use different shapes

Single run. get_run_result.sources is a flat source list for that run. Attach the response’s run_id to each entry to form run-by-URL records.

Comparison. compare_runs.read_urls groups records by original URL. The per_role entries record participation and flags within that comparison. Resolve their role labels to original run IDs before creating a run-by-URL table. When repeated prompts make that mapping ambiguous, use the individual run records. Retain the original grouped response alongside the derived table. [R21]

The following is an abridged shape from the operator’s dated example. It illustrates a grouped URL and the pagination metadata, with placeholders for identifiers and addresses. [R21]

{
  “meta”: {
    “run_ids”: [“<run-A>”, “<run-B>”],
    “runCount”: 2,
    “profileVersion”: “2026-06-03”,
    “generatedAt”: “<comparison-generation-time>”
  },
  “read_urls”: [{
    “url”: “<original URL>”,
    “runs_read”: 1, “total_reads”: 1,
    “cited_any”: true, “opened_any”: false,
    “per_role”: [{“role”: “A”, “cited”: true,
                  “opened”: false}]
  }],
  “read_urls_page”: {
    “offset”: 0, “limit”: 3, “returned”: 3, “total”: 132
  }
}

The excerpt omits additional URL rows and classification fields. The actual example’s returned count describes its full response, rather than the shortened display above.

Preserve these distinctions

The later dated examples clarify two earlier summary descriptions in the briefing: aggregate read_urls is URL-grouped, and the reasoning field is a summary. This appendix follows those examples. [R21]

Recover every required page

The briefing reports read_urls_limit and read_urls_offset, with a maximum of 1,000 rows per call. Check the current schema and returned read_urls_page. Record each offset, actual returned count, and reported total. Reconcile coverage across the pages; a zero-length response before the total is covered is an explicit gap to investigate. Keep the same selected run IDs and comparison scope while paging. [R21]

Rank, the snippets the model saw, reliable row-level query-to-URL linkage, and whether each surfaced page was fetched remain unknown in the documented record. Preserve those limits. URL presence establishes recorded encounter; causal influence requires additional evidence. [R20, R21]

WHAT TO DO Declare the record shape and mapping before counting. Save the raw response and a separate, reproducible transformation.

K5. Preserve the research record

Separate availability from a saved archive

The September 9 operator check reports seven TI tools returning data, with no native file-writing command on that inspected connection. It confirms complete answer text in get_run_result.output. A supported out-of-band export or client-side capture remains an operational question. [R21]

A native TI download and an authorized client that saves returned responses are possible routes to investigate. Demonstrate whichever route is available. Record the actual procedure, check completeness, and keep its status distinct from a proposed implementation.

What a preserved package should contain

Move original bytes or directly captured response data through the supported file route. Preserve the complete string value when decoding a JSON answer; log any subsequent display conversion. Model-authored summaries belong in analysis files alongside the originals.

What is reported preserved for Asteron

The Asteron preservation reports identify the approved prompt workbook, run index, trace registers, and source-event dataset covering all 81 conditions. The latest answer-preservation report reviewed for this note records 75 archived answers and six early Neutral/CFO interpretation answers pending export. Appendix J describes the scope and discrepancies. These are historical archive reports; this editorial revision adds no new TI execution or answer-export verification. [R06-R08]

Keep the authoritative identifiers when direct preservation remains unavailable. A clearly scoped source comparison remains useful; a claim requiring unrecovered text waits for that text.

Recover from interruption

Resume by locating the existing job and checking its status. Preserve completed and failed conditions. Treat any new submission as a separate attempt and link it to the original condition. Report which records were recovered, which remain missing, and whether another execution was approved. [R21]

WHAT TO DO Demonstrate preservation on one existing result before planning a large collection. Record the available archive and its exact remaining dependencies.

K6. Prepare a new comparison

Approve the design and settings together

Freeze the common record and role prompts from Steps 1-3. Select the exact system, effort, repetitions, and descriptive label. Establish the available budget and the preservation route. For a role comparison, keep the shared facts and evidence request constant and preserve each responsibility clause.

Submit only the approved prompts and settings. Save the returned job ID, check progress, and classify each condition as completed, failed, or still pending. Resolve original run IDs against exact prompts. Recover the views needed for the analysis, then use Steps 5-8 to inspect sources and write the findings. [R21]

Copyable instruction for an assistant with TI access

Use this guide and the attached situation and prompt files. First inspect the TI tools available to this connection and report the supported read, run, and preservation routes.

For an existing investigation, recover the specified job and runs. Match each exact prompt to its condition and replicate, then save the mapping from comparison labels to original run IDs. Retrieve complete answers, recorded queries, and the source records needed for the proposed analysis. Use compact views for orientation and complete records for counts.

For a new investigation, present the exact prompt set, model, effort, profile, repetitions, expected number of runs, budget constraints, and preservation plan for approval. Execute after that approval and record failures and retries as distinct attempts.

Preserve original response data through a supported file route where available. Report pagination coverage, missing records, and any unverified preservation step. Keep requested settings separate from returned metadata.

Compare the source sets at a declared unit. Read relevant source passages. Separate what the prompt supplied, what the tool recorded, what source inspection establishes, and what remains a hypothesis about the people. Return the Decision Evidence Record and the strongest supported next action.

This is an agent instruction, rather than an executable API payload. The connected agent uses the current tool schemas for calls. The operator briefing’s dated response examples provide a reference, while the live definitions govern supported parameters. [R21]

Use the record with the people involved

Use the record to prepare the next conversation, verification, or information change. Explain what was researched, what the sources support, and which questions remain open. Keep each person’s actual choices and responsibilities clear, including the decisions that require their own judgment.

WHAT TO DO Execute the approved comparison, preserve its record, and return to source inspection and role findings. Keep tool mechanics in service of the investigation.

Garrett French

Garrett French is the founder of Citation Labs, where he helps brands stay visible in AI answers and search through citation optimization and relevance-led link building at scale. His team studies how buyers use AI tools to shortlist purchases and deploy campaigns designed to increase client citations in recommendations.

He also built Xofu, a platform that tracks brand visibility across AI-generated recommendations, benchmarks competitors, and surfaces the pages AI references. And he leads ZipSprout, which builds sponsorship links by connecting businesses with nonprofits, events, and local organizations.

Garrett’s current explorations focus on decision efficiency and AI response behavior: how buyers decide, how AI systems “decide,” and how comparison assets influence what is cited for high-intent selection prompts.