Stop Tracking AI Answers Until Your Source Can Survive This Audit

Before counting AI mentions, citations or screenshots, establish whether the website can support an accurate public answer. If the controlled source is vague, unsupported or stale, a larger prompt sample only measures the defect more efficiently.

The first question is not “does ChatGPT mention us?” It is:

I call this source readiness. It is the permission a page earns before external answer tracking can produce a useful comparison rather than a gallery of unexplained outputs.

Choose one answer that could cost something

Do not begin with every page and every prompt variation. Choose one public answer whose inaccuracy could alter trust, fit, cost, risk or a buying decision. Examples include what the company actually does; who a service suits; where the business operates; what a product includes; which evidence supports a material claim; which condition makes the offer unsuitable; and how the company differs on a criterion that matters.

Write the shortest version the business is prepared to publish. Name the page responsible for maintaining it.

This becomes uncomfortable quickly. Sales may use one category while the homepage uses another. A service page may promise more than the documented process delivers. The strongest proof may live in a private proposal. A service-area claim may be true only under conditions the page omits.

Good. The audit has found the work before a visibility dashboard transformed it into a mystery. My rule is to start with one answer at risk, not an abstract score for the entire site.

Cross-examine the answer

Treat the proposed answer as if a cautious buyer, journalist or procurement team challenged it. Ask:

Do not average these answers into “83% AI ready”. One material contradiction cannot be diluted by good design, fast hosting and five accurate secondary facts.

I use a First-Fracture Rule: identify the earliest failure that prevents the answer being dependable, repair it, then repeat the audit. The result is a decision rather than a score everybody can admire without knowing what to do.

Identity failure often masquerades as flexible positioning

A source cannot support a stable answer when the organisation has not decided what to call itself or its offer. One page sells AI search visibility. Another says answer-engine optimisation. A proposal describes LLM discovery. Those terms may overlap, differ or represent components of one service. If the relationship is never explained, the website delegates the distinction to the reader. Check whether the owning page identifies:

• Accountable business or organisation
• Primary service, product or category
• Intended audience
• Problem or decision addressed
• Relevant geography or market
• Responsible person where expertise matters
• Route for verification or correction

Do not solve inconsistency by stuffing every synonym into the opening paragraph. Choose the primary description. Define adjacent language only when it changes understanding. Make controlled pages tell a compatible story. A brand team may call the inconsistency nuance. A buyer usually experiences it as uncertainty.

Make the claim survive outside the design

Brochure language can look persuasive until it is detached from the layout. “We help ambitious organisations unlock transformational growth.” Who helps whom? Through which work? What changes? Under what conditions? The sentence gives a cautious reader almost nothing to inspect.

A source-ready claim names the provider, work, audience, material output and relevant limitation. Specificity does not make the claim true. It makes it testable. Classify each load-bearing statement as:

• Supported evidence permits the wording.
• Conditional true only with an attached condition.
• Contradicted credible sources disagree.
• Private support exists but cannot currently be published.
• Unknown evidence is missing or unresolved.
• Decorative authority language doing no decision work.

Retain supported wording. Attach conditions. Resolve contradictions. Narrow private claims. Investigate unknown facts. Remove decorative authority. I would rather give a buyer a smaller claim that survives extraction than a grand sentence requiring a salesperson beside it.

Proof storage is not proof placement

Many websites possess evidence while leaving the claim unsupported at the point of doubt. A case study sits in an archive. The current method lives in a sales deck. A credential logo has no holder, issuer, scope or validity date. A product limitation appears only inside a download.

Inspect the page as a stranger. Can the reader understand why the consequential statement should be believed without opening six tabs and reverse-engineering the company’s filing system? Useful support may include a bounded case, product documentation, a documented method, a current policy, original measurement, a valid credential, permissioned customer evidence, named professional judgement or an authoritative external source.

The evidence type must match the claim. A customer quotation supports what that customer reported, not a universal result. A method explains how work is intended to happen, not that every project succeeded.

My test is proof at the point of doubt: place enough evidence or explanation beside the claim for a careful stranger to understand its basis. Deeper records can carry the detail later.

Make the caveat travel

AI answers, search snippets and hurried readers compress pages. A qualification near the footer may vanish while the confident sentence travels alone. Keep material conditions beside the statements they limit: geography, availability, eligibility, product version, price basis, sample period, causes a method does not address and outcomes the business cannot control. For example:

Improving how clearly a website states and supports a public claim cannot guarantee that a search or answer product will cite, rank, mention or recommend it.

That boundary does not weaken the article. It makes the sentence reusable without becoming a larger promise after extraction.

Contradictions matter more than missing keywords

Compare the owning page with the homepage, related services, company profiles, product documentation, policies and relevant external records. Pay attention to names, locations, service scope, prices, credentials, dates and guarantees.

Repetition is not corroboration. Ten directories may repeat the same unsupported description because they copied one old company profile. One current issuer record may carry more weight for a credential than a wall of badges.

State what each source is allowed to establish. A business can describe its process as a first-party statement. An issuer can confirm its credential record. A customer can report an experience. An official register can establish the fact recorded within that register’s scope. The objective is not a wall of matching sentences. It is an answer that does not collapse when the reader checks another credible source.

Freshness requires rechecking reality

Changing the year in a footer is not factual maintenance. For each unstable answer, record:

A page can pass today and become misleading tomorrow without anybody editing it. Maintenance is part of source readiness because no answer system can preserve a correction the business never publishes.

Test access after the answer is worth accessing

Technical eligibility still matters, but it should not become a substitute for the source audit.

For Google AI Overviews and AI Mode, check ordinary Search foundations: crawl access, indexability, canonical handling, rendered text and snippet eligibility. Google says no additional technical requirements or special schema are needed for these features, and eligibility does not guarantee inclusion (Google’s AI-features guidance).

For ChatGPT search, OpenAI advises publishers who want content included in summaries and snippets to allow OAI-SearchBot, which it distinguishes from GPTBot used in relation to potential training controls (OpenAI: publishers and developers FAQ).

Check each product’s documented access route. Do not treat every crawler, training control and answer experience as the same system. Access testing can prove that a barrier exists or was removed. It cannot prove future citation or selection.

Build a repair register rather than another dashboard

For each failed answer, record the answer at risk; the owning page; the first fracture (identity, claim, proof, caveat, contradiction, maintenance or access); the exact source or observation permitting the finding; the smallest repair that resolves the fracture; the claim ceiling the repair must not promise beyond; the owner who approves and maintains the fact; and the reopen trigger that demands another review.

Fix identity before adding an answer block. Publish evidence before amplifying a claim. Correct the sentence before adding markup. Repairing the first fracture stops the team commissioning work around a defect rather than resolving it.

Sample only after you have a baseline

Once the controlled source is defensible, sample external answers against it. Record the exact question, product, mode, date, relevant location or language, complete response and visible sources. Classify the material description as accurate, incomplete, misleading, wrong or unresolved.

OpenAI notes that ChatGPT search responses may include citations and a Sources panel while results can still be incomplete, outdated or incorrect (OpenAI: ChatGPT search). Treat each output as a dated observation, not an exposed account of every retrieval and generation step.

If the source is wrong, repair it. If an authoritative profile is stale, correct it. If the controlled source is accurate and one answer varies, preserve the observation without vandalising a sound page to chase one output. One changed answer after one edit is another sample. It is not automatic proof of causation.

Audit the object before tracking its shadow

Choose one answer that matters. Name its page. Cross-examine identity, claim, scope, evidence, caveat, contradiction, owner and access. Repair the first material fracture. Then compare external answers with the approved baseline.

This audit cannot guarantee crawling, citation, mention, clicks or sales. It can establish something smaller and more useful: the page now gives a careful reader an identifiable, supported, bounded and maintainable answer.

Stop counting shadows until the source can survive daylight.