Kojable Research · Methods
AI Source Research Measurement Framework
A proposed method for tracking the sources researchers discover, the citations AI answers display, and what the evidence supports.
Two evidence tracks
Observed AI answers
Capture the answer and any citations it displays.
Independent source research
Find relevant sources beyond the captured answer.
1. Why a source list is not an opportunity list
When an AI system retrieves a page, what have we learned?
Perhaps that the page was encountered under particular search conditions. Perhaps it appeared in a pool of sources available to a system. Perhaps it was referenced in the final answer. Those are different observations, and not every interface exposes all of them.
They tell us little, on their own, about whether a publisher is relevant to a buyer's question, whether a claim is supported, whether an author should be contacted or whether a useful information intervention is possible.
Yet source research often compresses discovery records, citations, publishers, creators and proposed actions into one list of URLs. A company can then mistake hundreds of discovered pages for hundreds of opportunities. A publisher cited in an answer may be labelled a prospective partner without a suitability review. A content gap may be presented as proof that a new article will improve future AI visibility.
None of those conclusions follows automatically.
A discovered source is not necessarily an observed citation. An observed citation is not necessarily a qualified opportunity. A qualified opportunity is not a confirmed placement or measurable result.
The research question is therefore not simply which sources appeared? It is:
How can source researchers move from recorded observations to reviewable shortlists, verification decisions, information gaps and possible interventions without claiming more than the evidence establishes?
At Kojable, this is part of AI answer alignment: understanding how companies are represented in AI-mediated information environments, where those descriptions are grounded, what is missing and what can responsibly be improved. The paper develops a proposed measurement language for the source-research portion of that broader problem. It is a methodological contribution, not a model leaderboard or an outreach playbook.
2. Study scope: eleven study deliverables, nine recoverable datasets
The research programme has eleven study deliverables—nine spreadsheets and two research documents—now structurally inspected in the supplied files. The studies span business-facing source and opportunity research and two independent investigations in different subject areas. Their designs and intended outputs were not identical. The deliverables are not identified by name in this paper.
A later archive audit located recoverable underlying datasets for nine project studies: seven client-project candidates and two independent studies. For the other two client-study deliverables, the backing run-level records were not located in the reviewed archive. These deliverables remain documentary sources; the missing backing data limit what can be reconstructed or independently checked from them.
This distinction matters. A workbook can preserve a shortlist, a rationale or a proposed gap even when the underlying raw model run is unavailable. It can be described as a workbook record after its contents and rights are checked. It cannot, by itself, establish the complete retrieval process, the number of unobserved candidates, independent reproducibility or a provider-level rate.
| Scope | What is available or reported | Appropriate inference |
|---|---|---|
| Eleven study deliverables | All eleven supplied files opened: nine spreadsheets and two documents; individual records require field-level review | Documentary scope of the research programme, subject to study-level provenance checks |
| Nine recoverable project datasets | Seven client-project candidates and two independent investigations, according to the archive audit | Boundary for retrospective run-backed record-coverage analysis |
| Two deliverable-only client studies | Deliverables exist, but their underlying run datasets were not located in the reviewed archive | Documented observations may be described with provenance and permission; do not treat them as run-reproduced evidence |
Method and design boundary. The archive review was retrospective. It assessed the availability and integrity of existing research outputs and supporting records; it did not prospectively apply the proposed framework in a controlled evaluation. Nor did the eleven study deliverables share a unified sampling strategy, topic, provider configuration or opportunity-qualification protocol.
The archive audit reports 19 completed ten-task provider runs across the nine recoverable datasets, representing 190 task–run instances. These are process and record-coverage counts—not 190 distinct research questions, unique publications, citations or independent studies. The two independent investigations include aligned ten-task runs from two AI systems, potentially allowing within-study comparison when their source definitions and URL handling are harmonised. No overlap statistic from those runs is reported here.
The resulting contribution is intentionally limited:
A proposed taxonomy that distinguishes different research objects.
A measurement contract for source identity, denominators, missingness, verification and decision status.
A retrospective coverage audit indicating which kinds of supporting records are recoverable for further mapping.
The archive audit does not establish that the taxonomy has been independently validated, that the framework improves decisions, or that any proposed intervention changed later AI answers. Those questions require separate evaluation.
3. Seven research objects and two evidence tracks
The framework distinguishes seven objects that are often merged in source research.
| Research object | Operational meaning | Minimum evidence to retain | What it does not establish |
|---|---|---|---|
| Discovered source | A page, document, publication, author or other source identified through a recorded research process | Question or task, method, source identifier and provenance | That a model cited the source in its final answer |
| Unique source | A distinct object after justified identity resolution at a declared level | Raw identifier, resolution rule and page/domain/publisher/person unit | That every record is a different publisher or person |
| Observed AI citation | A source visibly referenced in a captured model answer | Answer text, question, provider/surface, date and citation or support record | That the source caused the answer or will appear again |
| Shortlisted opportunity | A candidate selected for further assessment against explicit relevance criteria | Candidate, reviewed question, selection rationale and review status | That outreach is appropriate, accepted or likely to yield citation |
| Verified evidence | A source, identity or specific claim checked against a stated verification requirement | Check type, evidence inspected, reviewer/status and unresolved limits | That a URL check establishes factual truth, or that action is eligible |
| Content gap | An information need insufficiently addressed within the reviewed evidence for a specific question | Question, evidence scope, observed deficiency and confidence | That relevant information is absent across the entire web |
| Proposed intervention | A content, evidence or distribution action suggested in response to a documented gap | Gap link, proposed action, owner and decision state | That the action occurred or improved a future answer |
These objects may be related, but they are not interchangeable. An independently researched specialist publication can be worth examining without ever appearing in a measured model answer. Conversely, an institution cited in an answer may be an important reference but not an appropriate commercial contact.
Two distinct tracks
Observed-answer track. A defined question is submitted under recorded model and retrieval conditions. The full answer and any exposed citations or source supports are captured. This track supports statements about that particular answer, at that time and under those conditions.
Independent-research track. Researchers examine the wider information environment: relevant publications, authors, experts, creators, communities, documentation and technical sources, whether or not those sources appeared in the measured answer. This track supports a researched candidate inventory, not a claim about hidden model retrieval or training.
Both tracks may inform a content-gap review, but independent discovery cannot be relabelled as an observed citation, and an observed citation cannot automatically be relabelled as an outreach opportunity.
4. Measurement units, source identity and denominators
Before reporting a count or percentage, a source-research study needs to say what is being counted.
One page can occur in several answers. One domain may host several publishers. A named author may publish across multiple sites. A model's exposed research process may contain the same URL repeatedly. Counts at those levels answer different questions.
| Unit | Example question it can answer | Counting discipline |
|---|---|---|
| Research question or task | How many distinct questions were examined? | Count unique task IDs within the declared study |
| Model run or task–run instance | How many provider/task observations completed? | Keep provider, model/surface, task, timestamp and run distinct |
| Source record | How many source appearances were logged? | Count appearances, including repeats, at the raw-record level |
| Unique page or document | How many distinct resolved source items were found? | Apply declared page-level identity and deduplication rules |
| Domain or publisher | How many distinct hostnames or publishing organisations occurred? | Do not assume domain and publisher are equivalent |
| Author or creator | How many evidence-supported people or creator identities occurred? | Resolve person identity independently of page and publisher |
| Shortlisted candidate | How many items passed a stated review threshold? | Count only within a declared reviewed-candidate population |
| Content gap | How many question-linked information deficiencies were coded? | Apply a disclosed coding rule and preserve uncertainty |
| Intervention | How many actions were proposed, approved, executed or measured? | Report each status separately; do not collapse the stages |
For example, one page appearing in five task records can yield five task–source observations but remain one unique page. A shortlist containing twenty pages published by five organisations is twenty page-level candidates and five publisher entities—not twenty different publishers.
Source identity is a measurement decision
Source records often include redirects, tracking parameters, changed titles, syndicated copies and multiple addresses for related material. Deduplication should be conservative.
Retain the original identifier alongside any resolved URL. Record whether the resolution concerns an exact page, a domain, a publisher or a person. Use a documented rule for removing tracking parameters or following redirects. When evidence is insufficient to resolve a precise page or author, mark the identity unresolved instead of inventing a canonical identity.
Crucially, resolving a URL proves an identity or access fact only to the extent of the check performed. It does not verify every factual statement on the destination page. Kojable's separate source-identity research examines this problem in a different empirical context; those findings are not pooled into the present retrospective archive.
For every published metric, report at least its numerator, denominator, analytical unit, eligibility rules, exclusions, deduplication and relevant provider/date context. Without those fields, a percentage may be arithmetically correct yet substantively uninterpretable.
5. What the retrospective audit can—and cannot—measure
The user-supplied archive audit reports that all nine recoverable datasets contain some discovery, citation, shortlist and source/URL-identity verification records. That is evidence of technical record availability in the archived projects. It is not proof that every record was complete, that the seven objects were consistently classified across projects, or that every measurement layer was operationalised.
In particular, overlap eligibility, question-linked content gaps and intervention states still require field-level mapping. The source definitions and URL-handling rules of the two independent model comparisons must be fixed before an overlap calculation can be reported. A full nine-study-by-seven-layer coverage matrix therefore remains a pending research asset, not a result to fill by assumption.
Integrity checks are not full reproduction
The audit summary reports the following checks. The reported figures were reconciled against private evidence for this release.
| Audit item | Reported outcome | Interpretation boundary |
|---|---|---|
| Recoverable project datasets | Nine | Run-backed coverage denominator; not the number of study deliverables |
| Completed full provider runs | Nineteen | Completed archived process units, not nineteen independent studies |
| Task–run instances | 190 | Nineteen ten-task provider runs, not 190 unique sources |
| Raw/normalised file hashes | 380 of 380 passed applicable checks | File integrity under those checks; not proof that all analytic outputs reproduce |
| Policy-snapshot hashes | 19 of 19 passed applicable checks | Snapshot integrity under the reported policy; not validation of the taxonomy |
| Full-run aggregate manifests | Missing for 17 of 19 runs | Material limitation to end-to-end reproduction |
| Working-tree status | One independent study recorded a modified working tree | Environment/provenance caveat requiring reconciliation |
These results support a narrower statement than “the framework was validated across eleven studies”: the archive contains substantial recoverable research records that can be mapped to the proposed framework, subject to further reconciliation.
Missing is not zero
At every layer, the reporting system should distinguish:
| State | Meaning | Example |
|---|---|---|
| Zero | The measurement was actually made; no qualifying event was observed | A completed citation check captured no references under its declared rules |
| Not measured | The measurement was not performed or retained | No final-answer capture was part of a particular workbook's research process |
| Unknown | Available material cannot resolve the status | A citation field is blank and the relevant run record cannot be recovered |
The existence of a curated shortlist does not establish the number of all discovered sources. A missing citation field does not imply that the answer had zero citations. An unfilled verification field is not a failed check.
For the nine run-backed datasets, a publishable coverage register should record each layer as present, partial, not measured or unknown, with the supporting record location and applicable check type. The two deliverable-only studies may be described separately at the documentary level once reviewed, but they must not be used to enlarge a denominator that represents recoverable raw-run coverage.
6. Shortlisting and qualification are review decisions
Discovery is intentionally broad. Qualification is selective.
In this framework, a shortlisted opportunity is a source, publisher, author or creator selected for closer review because it may help address a relevant question or evidence need. It is not a forecast of publication, partnership, acceptance or AI citation.
The reason for including each candidate should be reviewable. We propose seven dimensions:
Question fit. Does the candidate address the actual buyer or research question?
Topical relevance. Is the publication, author or person active in the relevant subject?
Evidence fit. Is there a substantive information need the source could credibly help address?
Editorial independence. Are ownership, perspective and commercial interests understood?
Credibility. Is there support for identity, expertise and the claims material to this decision?
Practical eligibility. Would further research or a contribution be appropriate under the relevant rules?
Contribution value. Is there something genuinely useful to add beyond promoting a brand?
These are proposed qualification criteria, not validated predictors of editorial acceptance, engagement or citation. Different projects may require different thresholds and review owners.
Report the screening population
A shortlist count should ideally be accompanied by the number of candidates actually examined, the selection rule and the reasons for exclusion or deferral.
For example, a report might separately give the numbers reviewed, shortlisted, pending verification, reference-only and excluded. The example is a suggested reporting format, not an observed result from the eleven study deliverables.
If a workbook records only a curated list of twenty prospects, it can substantiate twenty recorded candidates at that unit. Without a complete discovery or screening ledger, it cannot substantiate that those twenty represented a particular proportion of all available sources.
An independent publication can be a credible reference without being an appropriate outreach target. Conversely, a relevant commercial publication may be a research candidate but not yet suitable for a contribution. Qualification must make that distinction explicit.
7. Verification is different from action eligibility
The word verified often conceals several checks. Resolving a URL, confirming a publisher's identity, checking the factual support for a claim and approving a candidate for outreach do not mean the same thing.
We propose two separate axes, together with a record of the specific check type.
| Axis | Question | Proposed statuses | What must be recorded |
|---|---|---|---|
| Evidence verification | What identity, source or claim was actually checked, and to what standard? | Verified; partially verified; pending review; failed; not assessed | Check type, supporting record, reviewer, date and unresolved limits |
| Opportunity eligibility | Is further action appropriate in this research context? | Eligible; conditional; reference-only; excluded; not assessed | Eligibility rule, decision maker, rationale and conditions |
This separation allows a source to be identity-verified yet factually unassessed, or credible and reference-only. For example, an authoritative public institution may be necessary to understand a topic but inappropriate to approach as a promotional opportunity.
The archive audit's reference to “verification” concerns source or URL identity checks. It does not establish that all source claims were fact-checked, that scientific claims were validated or that proposed outreach targets were independently assessed for ethics and rights.
Verification standards are purpose-dependent. A basic publisher-identity check may suffice for an inventory; a claim about patient safety would require a different substantive evidentiary and specialist-review process. No public-health conclusion is inferred from the mere inclusion of an independent monitoring workbook.
At minimum, a report should not use the bare label “verified” without naming what was verified.
8. Comparing source sets without conflating discovery and citation
Different systems can surface different sources while investigating similar questions. That observation can be informative, but only if the compared sets represent the same kind of event.
We distinguish:
Discovery overlap: intersection of sources found through defined model-specific research processes.
Citation overlap: intersection of sources explicitly referenced in captured final answers.
Shortlist overlap: intersection of candidates selected under defined qualification rules.
These measures need not agree. A model can discover a URL without citing it; a team can independently shortlist a publication that never appeared in the captured answer.
For two sets of comparable, normalised source identities, the Jaccard index is:
Here the numerator is the count of shared unique identities, and the denominator is the count of identities present in at least one set. A zero union requires an explicit no-data or both-empty rule; it should not silently become a meaningful overlap score.
Before computing or interpreting an overlap statistic, disclose:
Matched research questions or tasks and any excluded cells.
Providers, model/interface versions where available, retrieval conditions and collection dates.
Whether the sets represent discovery, final citations or shortlists.
Whether the identity level is exact URL, document, domain, publisher or author.
Redirect, normalisation, deduplication and unresolved-identity handling.
The unit and aggregation method: per task, per run or pooled, with appropriate denominators.
According to the archive audit, the two independent studies have aligned ten-task runs from two AI systems. They may support within-study source-set comparisons after their source definitions and URL-resolution rules are reconciled. The current material does not provide a calculated Jaccard value for those studies. The remaining client projects must not be presented as a matched provider-comparison benchmark simply because their deliverables contain citation or opportunity fields.
Low overlap, if subsequently observed, would describe the recorded sets under the specified conditions. It would not, by itself, prove model inaccuracy, permanent source preferences or that one provider is better.
Kojable has separately published a cross-model citation-overlap analysis with its own matched-panel design and findings. Those published results answer a different empirical question and are not results of this eleven-study programme.
9. Question-linked content gaps and proposed interventions
A content gap is not simply a company missing from a citation list. It is an information need that the reviewed evidence does not adequately address for a specified question.
Imagine a hypothetical buyer asking how a software category handles sensitive-data migration. The research might find broad comparison pages, product documentation and community discussion, but little accessible evidence for a particular security process. The potential gap may be a missing, supportable technical explanation—not a lack of promotional articles.
We propose six initial categories, subject to future independent validation:
| Proposed gap type | What the researcher examines | Possible response—not an observed outcome |
|---|---|---|
| Factual evidence | A material claim lacks accessible, adequate support | Clarify or publish substantiated evidence |
| Question coverage | Available sources do not directly answer the relevant question | Develop a focused explanation or comparison |
| Independent context | External discussion is limited or lacks relevant nuance | Investigate legitimate educational contribution opportunities |
| Freshness | Relevant material may be outdated | Review dates, update evidence and disclose changes |
| Identity or attribution | Source, author, organisation or supporting claim cannot be resolved | Improve attribution, provenance or documentation |
| Format or accessibility | Useful information exists but is difficult to find, understand or reuse | Restructure documentation or create an accessible explanation |
Each gap record should contain the originating question, the reviewed source set, the claimed deficiency, scope and confidence, and a proposed next action. A gap is always relative to the evidence reviewed; it is not proof that no relevant material exists elsewhere on the internet.
The gap classifications already present in archival workbooks do not establish that this six-part taxonomy has been independently validated. The archive audit also reported that gap and intervention fields still need further mapping to determine comparable coverage.
The companion study B2B Technology Content Gaps: Findings from Nine Audits reports seven recurring evidence themes in historical B2B technology audit records.
Keep proposals, actions and effects separate
Not every gap calls for a new article. Appropriate next steps could include clearer documentation, a better-supported claim, a direct comparison, an educational contribution through a legitimate independent publication, expert research, or no action if the evidence is too weak.
Record each stage separately:
| Stage | What can be reported | What cannot yet be inferred |
|---|---|---|
| Proposed | A possible intervention linked to a defined gap | That anyone approved or implemented it |
| Approved | A decision-maker accepted a specific action | That it was executed |
| Executed | The evidence, content or outreach action occurred | That a publisher accepted it or a model used it |
| Observed after execution | A later captured answer or source set changed or did not change | That the action caused the change |
| Effect evaluated | A specified evaluation assessed outcomes under recorded controls/conditions | Broad generalisation beyond the tested setting |
A submitted contribution is not a placement. A publication is not a guaranteed AI citation. A later answer change is not, on its own, a causal estimate.
For any follow-up, retain the questions, surfaces, dates, retrieval conditions, original answer evidence, subsequent answer evidence and competing explanations. Where practical, use repeat observations and controls. None of the eleven-study evidence supplied to this drafting process establishes a demonstrated intervention effect.
10. Minimum reporting and reproducibility standard
The purpose of a measurement framework is to make important decisions traceable, even when source collection is incomplete. We propose a six-part reporting contract.
| Report component | Minimum disclosure |
|---|---|
| Scope | Research question, subject, population, timeframe, provider/interface and collection method |
| Coverage | Which of the seven measurement layers were present, partial, not measured or unknown |
| Units and denominators | Numerator, denominator, resolved identity level, eligibility, exclusions, deduplication and aggregation |
| Qualification | Selection criteria, reviewed-candidate population, status, exclusions and rationale |
| Verification | Check type, evidence inspected, outcome, reviewer/date where appropriate and unresolved limitations |
| Actions and limits | Gap-to-intervention link, proposal/approval/execution/evaluation state, rights and uncertainty |
Such a report should let a reviewer trace an important recommendation back to a specific question and the evidence that supports it.
For this research programme, three provenance layers should also remain visible:
Documentary evidence: the eleven spreadsheet/document deliverables, with version, study scope, table/field definitions and authorship where available.
Recoverable run-level evidence: the nine datasets' source records, policies, provider/task captures and resolution outputs, with field mappings and integrity checks.
Unresolved evidence: the two studies lacking located backing datasets; incomplete aggregate manifests; unmapped gap/intervention fields; and unresolved source identities.
These layers should not be silently combined into one denominator. A workbook can support a description of what its author recorded; a reproducible model-observation rate requires the relevant underlying observations, eligibility rules and exclusions.
The proposed reporting standard is a methodological recommendation. Its effect on decision quality or AI-answer alignment has not been experimentally demonstrated by this retrospective audit.
11. How the framework relates to Kojable's research capabilities
Kojable is an AI answer alignment platform. The practical objective is not merely to count mentions or citations, but to help a company understand the evidence and context behind how AI systems describe it, decide what information is incomplete or misaligned, make an appropriate improvement and check what changed.
Its AI Citation Monitoring capability focuses on sources appearing in captured AI answers and the surrounding answer context. Those observations can help a team investigate source coverage, claim support and missing information. A visible citation does not prove that the source caused an answer or will be reused in another run.
Creator & Publisher Intelligence examines a related but broader research space: relevant publications, authors, experts and creators, including candidates that were not observed in a measured answer. It supports research, assessment and preparation of possible educational contributions for human review. Identifying a relevant publisher does not confirm availability, acceptance, partnership or future model citation.
The proposed measurement framework supports Kojable’s See → Align → Prove → Compound process.
See includes observing answer claims and exposed citations under defined conditions and diagnosing which claims or identities the recorded evidence supports, leaves out or cannot resolve.
Align means acting where an evidence-backed information need supports a responsible change.
Prove means documenting what changed and comparing later observations. Establishing causation requires additional evidence.
Compound uses the evidence record to choose which unresolved claims or research questions merit further work.
This distinction also protects the research page's role. The paper explains a measurement method. The product pages explain current capabilities. Separate practical articles can address implementation choices without duplicating this paper's definitions and evidence requirements.
12. Limitations, disclosure and research agenda
The framework is intentionally conservative because the evidence environment is heterogeneous.
Eleven study deliverables are not eleven independently reproduced runs. All eleven supplied workbooks/documents can be opened, but underlying archived data were recoverable for nine. The two deliverable-only studies require full field-level inspection of their contents, authorship, selection criteria and disclosure status. Their missing backing files cannot be treated as zero observations or as proof that the deliverables are invalid. They also cannot substantiate run-level completeness or reproducibility.
The nine run-backed projects have different purposes. Seven are client-project candidates and two are independent studies. Their prompts, source types, research methods, model settings and selection rules were not designed as one pooled benchmark. Differences between their records cannot automatically be interpreted as model performance differences.
The archive audit establishes technical availability, not usefulness. Its reported discovery, citation, shortlist and identity records provide a basis for further mapping. They do not show that the framework improved researcher decisions, that the seven objects were independently validated or that the six gap categories were reliably coded by independent reviewers.
Source-identity checks are not substantive fact checks. A resolved URL or confirmed organisation does not validate the accuracy of the page's claims. The correct verification standard depends on the subject, intended use and potential consequences.
Source discovery is incomplete by design. A URL absent from a captured answer or research ledger may still exist, be relevant or be retrievable under different circumstances. Model interfaces and exposed retrieval metadata vary and can change.
Qualification introduces judgement and selection bias. A shortlist reflects the research question, available evidence, reviewer and eligibility rules. Unless the full screened population is recorded, it should not be converted into a universal conversion rate.
Overlap is observational. Two recorded source sets may differ because of task wording, provider, retrieval exposure, date or identity rules. Without a matched design, repeated measurement and appropriate controls, a low or high score should not be generalised into a stable provider preference.
Proposed intervention is not demonstrated effect. No source-to-revenue effect, guaranteed placement, stable model preference or causal citation increase is established by these materials.
Reproduction remains incomplete. The user-reported archive hashes passed the applicable checks, but most full runs lack manifested aggregate files, and one independent study recorded a modified working tree. Integrity of available files is different from end-to-end rerunnability.
Publication and confidentiality. Kojable approved publication of the de-identified aggregate findings. Client and project identities and the underlying private research records remain confidential.
What would turn this proposal into stronger evidence?
The next research steps are concrete:
Complete the eleven-file register. All supplied files passed structural opening; next record version, project role, source definitions, table/worksheet field meanings, missing backing files and restrictions. Distinguish deliverable-only claims from run-backed claims.
Finish the nine-dataset coverage map. For each recoverable project, classify the seven layers as present, partial, not measured or unknown, with reproducible evidence pointers and check types.
Reconcile run provenance. Inspect the archive-audit JSON/TXT, missing aggregate manifests and the modified working tree; distinguish passing file hashes from reproducible aggregate outputs.
Conduct a defined overlap analysis where eligible. Fix task matching, URL resolution and identity levels for the two independent studies, then calculate and report actual within-study values rather than implying results.
Assess taxonomy reliability. Test whether independent reviewers can consistently distinguish the seven objects, six gap types and verification axes on a disclosed sample.
Evaluate utility separately. Determine whether applying the framework changes auditability, reviewer agreement or decision quality relative to a defined baseline.
Study interventions prospectively. Where ethical and practical, use comparable before/after observations, repetitions or controls, and record downstream actions and confounders.
Until those steps are complete, the framework remains proposed and the archive component remains a retrospective technical-coverage audit.
13. Frequently asked questions
Is a discovered source the same as an AI citation?
No. A source may be discovered through independent research or exposed retrieval without being referenced in the final answer. An observed AI citation requires a captured answer that actually identifies the source. The capture's date, provider and conditions matter.
Why does the paper refer to eleven studies if underlying data were recovered for only nine?
The research programme contains eleven study workbooks and documents. The retrospective archive audit located backing run-level datasets for nine. The other two deliverables exist, but their underlying run data were not found in the reviewed archive. We preserve all eleven as the documentary scope while restricting run-backed coverage and reproduction claims to nine.
Does the figure of 190 mean there were 190 separate studies or 190 citations?
No. The audit reports 19 completed ten-task provider runs, or 190 task–run instances. A task–run instance is a process observation under a particular run, not a distinct study, unique URL, citation or publisher.
Can a workbook be useful without the original raw model runs?
Yes, as a documented research deliverable. It may preserve the selection rationale, a shortlist or a classification. But a workbook-only record does not independently reconstruct unseen source discovery, provider responses or a complete sampling denominator. Its claims need explicit provenance and appropriate review.
What makes a publisher or creator a qualified opportunity?
It must merit closer examination for a defined question or evidence need, under recorded criteria for relevance, credibility, independence, eligibility and contribution value. Qualified means shortlisted for review here; it does not imply contact, consent, acceptance, partnership or a later AI citation.
Does a verified source URL mean that the page's claims are correct?
No. URL resolution and source identity checks do not amount to substantive fact-checking. A report should name the check performed and separately record any claim-level verification and action-eligibility decision.
How should source overlap between two AI systems be measured?
Compare matched tasks and like-for-like source sets—discovered sources with discoveries, final citations with final citations, and shortlists with shortlists. Resolve identities consistently and declare the unit, time, interface and missing-data rules. Jaccard overlap is one appropriate descriptive measure. No numeric overlap result from the two independent archived studies is asserted here.
Does identifying a content gap mean the company will appear in future AI answers?
No. A content gap identifies an information need relative to the reviewed evidence. Improving a source may or may not affect discoverability, retrieval, selection, synthesis, attribution or recommendation. A later changed answer is an observation, not automatic proof that the intervention caused it.
How is this paper different from Creator & Publisher Intelligence?
The paper sets out proposed measurement definitions, evidence requirements and reporting limits. Kojable's Creator & Publisher Intelligence is a research capability for investigating independent publishers, authors, experts and creators and preparing potential contributions for human review. The framework explains how to label and assess those outputs without confusing them with measured AI citations.
Conclusion: measure the evidence before recommending the action
AI-mediated discovery creates a complicated information environment. A business may be described through its own material, independent publications, technical documents, communities, experts and creators. Different systems may expose different parts of that landscape, and a single observed answer rarely reveals the full chain from source discovery to generated claim.
The challenge is therefore not just to collect more URLs. It is to establish what the records represent, what has actually been observed, which identities are distinct, which candidates deserve review, what evidence has been checked and whether a proposed next step is justified.
The eleven-study research programme motivates that distinction; the nine recoverable datasets provide a bounded basis for further technical coverage analysis. Neither foundation supports an unqualified claim of validated effectiveness.
A source-to-opportunity measurement framework can nevertheless make research decisions more intelligible. Its value as a proposed discipline is to insist on traceable questions, explicit units, reviewable qualification, honest missingness and a clear line between a recommendation and an outcome.
For the related research capability, see Kojable Creator & Publisher Intelligence.