Kojable Blog reference entry
How to Build Public Evidence for Claims AI Systems Miss
Public evidence for AI answers is current, accessible and appropriately substantiated information that allows a material company claim to be checked in the public information environment, using evidence appropriate to that particular claim.
Category AI Search Guides
Also known as public evidence for AI answers, AI answer evidence remediation, AI evidence building, evidence for wrong AI answers, correcting outdated company information in AI answers
Direct answer: Build public evidence only after a material claim has been verified and evidence remediation is justified. Start with the claim, define what would substantiate it, and use the appropriate evidence owner. Correct an authoritative existing source when possible, create a new asset only when no adequate owner exists, and use legitimate third-party correction routes for factual records you do not control. Put the claim and its proof in visible, accessible content, keep structural information consistent with that evidence, then retest comparable AI questions.
| In brief | Practical decision |
|---|---|
| Start with the claim | Do not start by deciding to write another page |
| Match evidence to the claim | Product, pricing, security and reputation claims require different proof |
| Use the right evidence owner | Update, create or correct according to the actual evidence condition |
| Evidence before optimisation | Schema, citations and formatting cannot replace substantiation |
| Verify afterwards | Publication is an intervention; comparable retesting shows what changed |
A material claim is a factual proposition that could meaningfully affect how a buyer understands, evaluates or compares a company. It might concern a product capability, target audience, integration, price, security credential, customer outcome or category position.
The objective is not to make the public web repeat a preferred marketing message. It is to make a supportable company truth sufficiently clear and evidenced that a buyer can verify it, then observe whether AI representation changes afterwards.
01 · Definition
What does it mean to build public evidence for a claim AI systems miss?
Building public evidence means creating, correcting or strengthening information that appropriately substantiates a verified material claim. The unit of work is the claim, not the URL, citation count or amount of content published.
Suppose an AI answer describes a software company as serving small businesses when the company now serves mid-market and enterprise teams. “Publish an enterprise article” is not yet an evidence specification.
The useful questions are more precise. What is currently true? Which part of the AI answer is materially inconsistent with that truth? What evidence would allow a buyer to verify the current audience? Does that evidence already exist? Is it absent, stale, contradictory or too vague?
This claim-level approach matters because there is no single source type that owns every final AI citation portfolio. Kojable research found overlapping Brand-owned, Comparable-vendor, community, official-documentation and other third-party sources, with the mix changing by platform and buyer-question type. The research also explicitly warns that source frequency does not establish authority, quality, causal influence or support for a specific claim.
The better starting rule is:
Build the evidence the claim requires, rather than trying to reproduce a generic “AI citation profile”.
02 · Diagnosis
When should evidence-building begin?
Evidence-building should begin after the material claim and current company truth have been verified, and the diagnosis shows that missing, weak, outdated or conflicting public evidence actually requires remediation.
A bad AI answer is not, by itself, proof that more public evidence is needed.
A citation miss is not proof either. Kojable's source-availability research found that more than three-quarters of observed citation misses had a contemporaneous citation elsewhere for the same question. That weakens broad source absence as a default explanation, but it does not reveal what the non-citing platform discovered, accessed, ranked or rejected.
| Observed condition | What the team should determine |
|---|---|
| Important claim is absent from an AI answer | Whether evidence is genuinely missing or another diagnostic condition applies |
| Correct page exists but the proposition is vague | Whether existing evidence needs strengthening |
| Current and old controlled pages disagree | Whether the owned evidence environment needs reconciliation |
| Third-party profile carries an obsolete fact | Whether the record is legitimately correctable |
| Correct evidence already exists | Whether new evidence is actually required |
| One AI system does not cite the evidence | Whether this is an evidence problem at all rather than assuming universal source absence |
The point of diagnosis is to avoid turning every AI representation problem into a publishing request.
If the evidence condition is not yet established, use the upstream diagnostic process first.
03 · Evidence specification
What evidence does the claim actually require?
The evidence burden follows the claim. Ask what a reasonable buyer would need to verify the proposition, then identify the source best placed to provide that proof.
“Publish more authoritative content” is not an evidence specification.
A product feature is verified differently from a security certification. A current price requires a different evidence owner from a customer outcome. A company's own documentation may be the appropriate authority for an implementation detail, while some reputation or market claims reasonably require independent evidence.
Claim-to-evidence specification
| Claim type | What needs to be substantiated | Evidence to evaluate | Common failure |
|---|---|---|---|
| Product capability | What the product does, current scope and material limitations | Current product page or technical documentation | Capability implied in promotional copy but never stated precisely |
| Integration | Whether the integration exists, is supported and under what conditions | Official integration or product documentation; relevant partner evidence where appropriate | Historic directory entry or vague “integrates with everything” wording |
| Pricing or packaging | The current commercial fact | Current authoritative pricing or offer source | Old article or directory carrying retired pricing |
| Security or compliance | The exact certification, standard, audit or control | Formal security documentation, certification evidence or appropriate official record | “Enterprise-grade security” with no named substantiation |
| Customer outcome | The result, attribution, analytical scope and material conditions | Documented case evidence or research | Percentage with no denominator, method, date or attribution |
| Audience or use-case fit | Evidence that the company genuinely serves the stated audience or problem | Current positioning and use-case evidence, with proof appropriate to the claim | Generic ICP assertion without product or customer context |
| Category or differentiation | What makes the classification or distinction defensible | Clear first-party definition plus appropriate comparative or independent evidence | Brand slogan presented as an objective market fact |
| Company fact | Current verified company truth | Appropriate official or authoritative company source | Conflicting addresses, product names, personnel or company descriptions |
The next step is to turn that judgement into an explicit evidence specification before anyone writes or edits a page.
Evidence Specification
| Evidence-specification field | What the team records |
|---|---|
| Material claim | The exact proposition that needs support |
| Verified truth | What is currently true according to approved evidence |
| Buyer relevance | Which buyer decision the claim materially affects |
| Evidence owner | The source or team that should authoritatively own the evidence |
| Required substantiation | The proof needed to make the claim defensible |
| Scope and qualifiers | Conditions, exclusions, audience, product tier or other boundaries |
| Currency | Relevant date and the event that should trigger review |
| Public evidence surface | Where the claim and proof should actually appear |
| Corroboration | Independent or official evidence where the claim reasonably requires it |
| Verification signal | What later AI-answer observation should be retested |
This changes the team's starting question from:
“What content should we publish?”
to:
“What does this claim require before we can reasonably expect a buyer to verify it?”
That distinction prevents arbitrary source quotas. There is no defensible rule that every company claim needs three pages, two third-party mentions or a directory profile.
Kojable's source-composition research helps explain why. In the observed cohort, product-capability questions showed strong Brand-owned presence, documentation questions leaned more towards official documentation, and vendor-evaluation questions exposed broader comparative portfolios. The supported conclusion is not that one source family wins. It is that the evidence environment changes with the information need.
04 · Evidence owner
Should you update an existing source, create new evidence or correct a third-party record?
Strengthen or correct the appropriate existing evidence owner when one exists. Create a new asset when the required evidence has no adequate home. Correct external factual records only where a legitimate correction or profile-management route exists.
Creating another page is attractive because the company controls the work. That does not make it the appropriate intervention.
| Evidence condition | Appropriate action |
|---|---|
| Authoritative source exists but the proof is incomplete | Strengthen the existing source |
| Authoritative source contains an incorrect or stale fact | Correct the existing source |
| Multiple controlled pages conflict | Reconcile the controlled sources |
| Required evidence has no suitable public owner | Create the appropriate evidence asset |
| Managed third-party profile contains a factual error | Use its legitimate correction process |
| Independent historical source was accurate when published | Do not try to rewrite history; establish current evidence elsewhere where justified |
| A source merely appeared in one AI citation | Do not change it solely because it was cited |
| Correct and sufficient evidence already exists | Do not automatically publish another page |
A visible citation can identify a source worth investigating, but it is not hidden-model telemetry. Kojable's content-access research shows that an emitted citation does not establish whether the platform fetched the live page, used another representation, extracted its text or selected it causally over other candidates.
That is why “fix the cited source” is not a complete methodology.
The broader decision between evidence remediation and other interventions belongs in the parent remediation framework.
05 · Reconciliation
How should you correct outdated or conflicting company information?
Establish the verified current fact, correct authoritative sources you control, reconcile contradictions across controlled surfaces, then address legitimately correctable external records where they materially conflict with current reality.
Outdated information is a particular evidence problem because the correct claim may already exist somewhere while older versions remain public.
Imagine a B2B software company that has moved from serving small teams to enterprise organisations. Its homepage now reflects that position, but an old product page still describes small-business workflows and a managed software profile carries the previous category description.
Publishing another “enterprise software” article would add another public surface without resolving the contradiction.
A better sequence is to establish the verified current position, identify which controlled pages remain inconsistent with it, update those sources according to their actual role, then review external factual profiles that the company is legitimately entitled to maintain or correct.
The distinction between outdated and historical matters. A live vendor profile claiming a discontinued capability may be factually stale. A five-year-old news article accurately describing what the company offered at the time is historical evidence. It should not be treated like an editable company profile simply because the business later changed.
“External source” is also not synonymous with “source we control”. A company may be able to edit a managed profile or request correction of a factual error through an established process. That does not give the company editorial control over an independent publication.
The objective is not to erase the past. It is to make the current evidence environment accurate, coherent and appropriately dated.
Whether AI answers subsequently change is a separate verification question.
06 · Evidence surface
How should the page carrying the evidence be structured?
Put the material claim and its substantiation in visible content. Use titles, headings, metadata and appropriate structured information to describe that evidence accurately, not to replace it.
Evidence and optimisation are often conflated.
A page can have excellent structured data while making an unsupported proposition. A page can also contain the correct evidence but bury it inside vague marketing language. Neither makes a strong evidence surface.
A 2026 KDD paper, SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization, provides useful controlled evidence here. Its benchmark used 2,700 queries across nine domains and a corpus of 171,003 web documents, with retrieval, reranking and generation evaluated as separate stages. In that experimental environment, body-text-only optimisation frequently harmed upstream visibility, while structural information and substantive body content played complementary roles. The paper's citation-source analysis also found that most cited material originated in body text rather than the structural fields.
The limitation belongs next to the finding. SAGEO Arena is a controlled benchmark, not a reproduction of every commercial AI search system, and the authors explicitly identify proprietary ranking and user-behaviour signals as outside the environment.
The practical lesson is therefore narrower than “schema improves AI visibility”.
| Page layer | Its job |
|---|---|
| Visible body content | State and substantiate the material claim |
| Title and headings | Make the page subject and evidence structure clear |
| Entity references | Remove avoidable ambiguity about who or what the claim concerns |
| Metadata | Accurately summarise the page |
| Structured data | Describe supported visible facts where appropriate |
| Technical accessibility | Make the public evidence available to relevant search or retrieval systems |
| Internal relationships | Connect evidence to relevant product, research, trust or application context |
Google's current guidance is useful because it sets a clear boundary. To be eligible as a supporting link in AI Overviews or AI Mode, a page must satisfy normal Google Search eligibility. Google says there are no additional technical requirements, no special schema.org markup is required, important information should be available in textual form, and structured data should match the visible text. Meeting those conditions still does not guarantee crawling, indexing or serving.
Google: AI features and your website
OpenAI likewise says public websites can appear in ChatGPT Search and advises publishers not to block OAI-SearchBot when they want page content to be eligible for inclusion in summaries and snippets. That describes an accessibility condition, not a guaranteed citation outcome.
OpenAI Publishers and Developers FAQ
Make the proposition explicit
Good evidence copy should allow a reader to identify the subject, claim, scope, substantiation and material qualification without reconstructing the proposition from several vague paragraphs.
“Supports enterprise deployment” is weak evidence if the page never defines what enterprise deployment means.
A stronger evidence surface might state the relevant deployment model, supported authentication or administration requirements, applicable security evidence, important product scope and known limitations.
That does not require turning every sentence into an “AI-optimised” formula. It requires being precise enough for the proposition to be checked.
Kojable's cited-page research reinforces the need for restraint. Lists, tables and structured markers were common among pages already cited in its benchmark, but the study explicitly states that this descriptive prevalence does not establish that any of those features increased retrieval or citation probability. It also does not identify an ideal page length.
Evidence clarity is the objective. Imitating the average cited page is not.
07 · Sufficiency
How much public evidence is enough?
There is no universal number of pages, citations or third-party references required. Evidence is sufficient when the material claim has appropriate, current and credible substantiation for the buyer decision it supports.
In some cases, one authoritative source may be the natural evidence owner.
If the question is the current price of a company's plan, its current pricing source may be the appropriate first authority.
If the claim concerns a formal certification, the relevant certification or security record may be required.
If the proposition is that customers in regulated enterprises consistently achieve a particular outcome, the evidence burden is much greater. An unsupported product-page assertion is not equivalent to documented outcome evidence.
First-party versus third-party is therefore the wrong first question.
Ask instead:
What makes this particular claim verifiable?
Kojable's cross-provider research also argues against assuming every system will converge on the same few pages. In a fixed nine-question B2B benchmark, average within-question exact-URL Jaccard overlap ranged from 0.93% to 2.02% across provider pairs. Domain and publisher overlap were slightly higher but remained below 5% in that panel. The study used one observed run per provider-question cell, so those numbers should not be interpreted as stable provider rankings.
The practical implication is not “publish more URLs”.
It is:
Build defensible evidence without assuming one page must become the universal AI source.
08 · Anti-patterns
What evidence-building tactics should you avoid?
Avoid tactics that make content look more authoritative without making the underlying claim more supportable.
| Weak tactic | Why it fails the evidence test | Better action |
|---|---|---|
| Publish before diagnosing the gap | The new page may solve a problem that was never established | Verify the claim and evidence condition first |
| Set an arbitrary source quota | Three weak sources do not necessarily substantiate a claim better than one appropriate authority | Define claim-specific evidence requirements |
| Add citations simply to increase citation count | Citation quantity and evidentiary support are different measurements | Cite sources because they support the material proposition |
| Treat a recurring source as the cause | Final citations do not expose the complete retrieval or ranking process | Treat recurrence as diagnostic evidence, not causal proof |
| Use structured data instead of visible proof | Machine-readable description does not make an unsupported proposition true | Put the evidence in visible content, then describe it accurately |
| Add statistics for authority | A number without source, denominator or method can weaken the claim | Publish reproducible metrics with the required context |
| Make language more technical for GEO | Added terminology can reduce clarity or shift relevance | Use the terminology the proposition actually requires |
| Create a page for every prompt variation | One information need can be expressed in many ways | Build around the underlying entity, claim and buyer decision |
| Assume one platform's citation behaviour applies everywhere | Search and citation environments differ | Verify across the surfaces that matter |
| Promise a fixed update period | Crawl, retrieval and AI-answer behaviour do not follow one guaranteed schedule | Use predefined measurement checkpoints |
Kojable's citation-quality research directly supports the distinction between citation quantity and evidentiary quality. In its matched panel, citation volume, material-claim citation coverage and reviewed source support were separate measurements; a larger marker count could not establish whether the attached source supported the important claim or whether the overall answer was accurate.
Do not optimise for one exact prompt
A related mistake is treating the precise wording of one AI prompt as the permanent content brief.
SAGEO Arena's supplementary analysis found that its tested optimisation methods showed limited robustness when the evaluation queries were reformulated, especially as the wording moved further from the original phrasing. That is evidence from one controlled benchmark, not a universal commercial-engine mechanism.
Official platform documentation gives a second reason to avoid prompt overfitting. Google says AI Overviews and AI Mode may use query fan-out across related searches and subtopics. OpenAI says ChatGPT Search typically rewrites a request into one or more targeted queries when using search partners and may issue further queries after reviewing initial results.
OpenAI: Searching the web with ChatGPT
Build around the entity, material claim and buyer information need, not one screenshot of one prompt.
09 · Verification
How should you verify what changed?
Record the evidence intervention, preserve the buyer question and baseline conditions, then retest comparable questions at predefined checkpoints. A changed answer is an observation, not automatic proof that the evidence change caused it.
Evidence-building is incomplete if the team stops at publication.
A useful intervention record should preserve the original representation gap, the evidence condition that justified action, the exact source changed or created, the date of the intervention and the buyer questions that will be retested.
Later checks should use the same or comparable conditions wherever possible. If the material claim begins to appear accurately, record that movement. If it remains absent or wrong, revisit the diagnosis without jumping to the conclusion that one particular source “failed to influence” the system.
There is no universal propagation deadline. Google states that satisfying Search and AI-feature requirements does not guarantee that content will be crawled, indexed or served.
Use predefined measurement checkpoints because they make comparison possible, not because a platform is expected to update by a particular day.
The full methodology for preserving a baseline, retesting comparable prompts and classifying observed movement belongs in the downstream verification guide.
The operating loop is:
Monitor → Diagnose → Improve → Verify → Monitor again.
FAQ
Frequently asked questions
Do AI systems need third-party evidence before they will reflect a company claim?
No universal rule requires third-party evidence for every company claim. The appropriate evidence depends on what is being asserted. A current product capability or controlled commercial fact may have an appropriate first-party authority, while a reputation, comparative or market-outcome claim may reasonably require independent evidence.
Is my company's website enough public evidence?
For claims the company is the natural authority on, such as current functionality, documentation or controlled commercial facts, an accurate first-party page may be the right evidence owner. Other propositions may require external verification or corroboration. The important question is not whether evidence is first-party. It is whether it appropriately substantiates the claim.
How many public sources should support a company claim?
There is no defensible universal number. Use enough appropriate evidence to make the proposition checkable and supportable. A simple current company fact may have one clear authority. A strong comparative, reputation or outcome claim may require a broader evidentiary basis.
Does structured data help AI systems use company information?
Structured data can make page semantics explicit for systems that use it, but it must describe facts supported by the visible page. It does not substitute for the evidence itself. For Google AI Overviews and AI Mode specifically, Google says structured data should match visible text and that there is no special schema.org structured data required for those features. Meeting Search requirements also does not guarantee that the page will be crawled, indexed or served.
Does an AI citation prove that the cited page caused the answer?
No. An emitted citation shows that a source reference reached the final answer interface. It does not, by itself, establish whether the platform fetched the live page, used another representation, extracted particular text, ranked the source above alternatives or caused the associated claim. Use citations as diagnostic evidence, not hidden-model telemetry.
How long should I wait before retesting an AI answer after changing the evidence?
Use predefined checkpoints rather than treating one interval as a guaranteed propagation period. A T+7, T+14 or T+30 result is an observation at that checkpoint. It is not proof that an AI system should have updated by then. The important methodological requirement is to preserve the baseline, retest comparable questions consistently and record material changes in the testing environment.
The practical takeaway
When an AI system misses or repeats the wrong version of an important company claim, the solution is not automatically more content.
Start with the material claim. Verify what is actually true. Define what would substantiate it. Identify the appropriate evidence owner. Strengthen an existing authoritative source where possible, create new evidence when a genuine gap exists, and use legitimate correction routes for external factual records when appropriate.
Put the evidence itself in visible content. Use structure and machine-readable information to describe that evidence accurately rather than replacing it. Do not measure the work by the number of pages published, citation markers acquired or supposed GEO tactics applied.
Measure whether the public evidence is stronger and whether comparable AI answers subsequently represent the material claim more accurately.
Kojable treats this as part of the wider AI answer alignment process:
Monitor → Diagnose → Improve → Verify.
If the team has not yet established which company claims are inaccurate, missing or weakly evidenced in AI answers, start with the current representation baseline before deciding what evidence to build.