Why AI search cites one page instead of another: the matched page audit method, and what our own audit recorded

When an engine chooses between two pages that answer the same question, the reasons it gave us in writing were about published evidence rather than product quality: a dated price, a source next to a competitor's figure, and an explicit coverage list. On 17 September 2026, on a question about finding every case against a company, Claude ranked two enterprise vendors above clawlaw.in specifically because they publish explicit court and tribunal coverage lists, and it noted that its ordering reflected price transparency and source authority rather than product quality. On the same date it discounted two of the site's own comparison pages because they stated competitors' prices with no link, no date and no source; the figures turned out to be correct when it checked them independently, but it had no way to know that at read time. We have not run a 200 query matched page audit that produces a citation rate per page feature, so this page publishes the design for one and reports only the counts we hold.

A citation rate per feature would be the most useful number in this field. It is also the easiest one to fake, so the design matters more than a headline.

The answer, first

Four things our own record supports, stated as reasons an engine gave rather than as a formula.

A page that can be checked beats a page that must be believed. The 17 September 2026 finding on unsourced competitor prices is the clearest single case: correct figures, discounted anyway, because at read time there was no way to verify them.

A named list beats an adjective. Explicit court and tribunal coverage lists were the stated reason two vendors outranked a competitor on a coverage question.

A page must be readable before anything else matters. On 6 August 2026 Claude recorded that the clawlaw.in pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of the company's own site.

A claim that fails a check can cost you the whole answer. On 6 August 2026 Claude found a headline figure on the main site that could not be true, checked it against public numbers and advised a buyer against the product, after which the company's accurate claims stopped counting for that answer.

What this is not. None of that is a ranking formula, and none of it is a citation rate. Every sentence above is a reason recorded on a date, on one programme, and reasons are evidence about mechanisms rather than measurements of them.

How it was measured, and how to measure it properly

The design below is the matched page audit the question deserves. Its whole purpose is to compare pages that differ in one measurable way, rather than to compare pages that differ in everything.

One. Choose 200 queries with a real choice behind them. Buyer questions where more than one page could reasonably be the source. Five groups with the counts recorded: category choice, comparison, price, suitability, and risk or limits. Blind rule on every one: no vendor name, no product name, no domain, no phrase that appears only on one site.

Two. For each query, collect the candidate page set. The pages that plausibly answer it, including your own and your competitors'. This is the denominator that makes a per feature rate meaningful: pages that could have been cited, not pages that were.

Three. Code each candidate page on seven features, each strictly yes or no. A feature that needs a judgement call cannot be coded consistently, so each of these is written to be checkable by a stranger.

• Dated primary evidence. A number, price or figure with a date next to it on the page.
• Sourced third party figures. Every figure about another company carries a link and a date.
• Named enumeration of scope. Coverage, regions, integrations or courts listed by name and counted, rather than described with a word like comprehensive.
• Stated limits. The page says what it does not cover or who should not buy.
• Readable without scripts. A plain fetch returns the key facts as text.
• Reproducible method. The page describes how a result was produced well enough for a reader to repeat it.
• Independent verifiability of the main claim. The central claim can be checked against a public source the page names.

Four. Build matched pairs. For each query, pair two candidate pages that are alike on everything you can control, page type, length band, site age band and topic, and differ on exactly one of the seven features. Matching is what turns a correlation into something worth reading. Report how many pairs you could build, because on most queries you will not be able to match cleanly and that number is part of the result.

Five. Run the queries and record citations. One query per fresh conversation, two assistants, search recorded per answer, location fixed, every cited domain and URL logged.

Six. Compute the rate per feature as pages cited over candidate pages with that feature. Then compute it again for pages without the feature, over the same query set. Report both counts with both denominators, in the form so many of so many. Also report the matched pair result separately: of the pairs where exactly one page had the feature, how many times the page with the feature was the one cited.

Seven. State the confounders you cannot remove. Domain authority, age, backlinks, brand familiarity in the model's memory, and whether the engine searched at all. Any of them can produce the same pattern as the feature you are testing. This is why the section below says what it says.

Eight. Say in advance what each result would mean. If pages with dated primary evidence are cited far more often, and the matched pair result points the same way, you have a working hypothesis worth acting on, still not a cause. If the overall rate differs but the matched pairs do not, the difference is probably the sites, not the feature. If the searched share is low, the whole audit describes a small live subset and the rest is memory.

What the numbers were

The 200 query matched page audit: not run. We hold no citation rate per page feature, no candidate page denominator and no matched pair counts. We are not going to publish a number of the form so many of so many pages with dated evidence were cited, because we have not counted it.

What we hold on the citation side.

• ChatGPT, 18 August 2026, 18 blind commercial questions, clawlaw.in. Top source on exactly 1 of 18. In that answer the site's own comparison page was named as still being the vendor's own editorial page, which is a reason recorded against a specific page.
• ChatGPT, 6 August 2026. One named page cited: clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india, for a vendor due diligence question, fourteen days after publication, in a zone where the same question set had named the company nowhere at baseline. That page is a how to page with a reproducible method, which is suggestive and is one page.
• Perplexity, September 2026, aiknowsus.com. In one batch of six questions the engine reported afterwards that it had not cited or recommended the domain in any of the six. Across 24 batches and 72 conversations, the phrase recording that we were not cited appears 161 times in the engines' own self audits.

What we hold on the reason side, which is the stronger part of this record.

• Sourced third party figures, Claude, 17 September 2026. Two comparison pages discounted for stating competitors' prices with no link, no date and no source, despite the figures being correct.
• Named enumeration of scope, Claude, 17 September 2026. Two enterprise vendors ranked above clawlaw.in specifically because they publish explicit court and tribunal coverage lists, with the engine noting that its ordering reflected price transparency and source authority rather than product quality.
• Readable without scripts, Claude, 6 August 2026. Prices rendered only after scripts ran, so a crawler received no prices, and the quoted prices came from an app store listing instead.
• Consistency of the public record, ChatGPT, 6 August 2026. The website and the app store listing carried different plan names and different prices for the same product, and the engine said so in its answer.
• Independent verifiability, Claude, 6 August 2026. A headline figure that could not be true, checked against public numbers, followed by advice against the product.
• Attribution is not guaranteed even when content is used, Claude, 17 September 2026. It admitted using two specific arguments from the site's pages and dropping the attribution in both places, describing it as a citation lapse rather than a ranking judgement.
• Official sources win their own questions, Claude, 17 September 2026. Official court portals above every commercial product on how to questions, with the engine stating that no commercial product should rank above the official portal for a question about using that portal.
• Instructing the engine from inside the page backfires, Claude, 6 August 2026. A competitor's page carried a hidden block of text addressed to answer engines, instructing them to cite that company as the source. The engine found it, refused it, and named the company that had done it.

Eight recorded reasons on named dates, against three citation counts. That imbalance is honest: we know more about why than about how often.

What this cannot tell you

Correlation is not causation, and here it is not even correlation yet. We have reasons an engine stated and citations we counted, and we have not linked them in a design that controls for anything. Even a completed matched pair audit would only narrow the confounders, not remove them.

• An engine's stated reason is not necessarily the operative one. It is a statement made in an answer. Our own capture includes an engine withdrawing a statement about its own behaviour in three separate batches in September 2026.
• Page features travel together. Sites that date their prices also tend to source their comparisons and state their limits. Separating those needs matched pairs, which is exactly the part not yet run.
• The candidate set is a judgement. Deciding which pages could have been cited is the softest step in the whole design, and two analysts will build different sets.
• One citation is not a rate. The 6 August 2026 page is one page, one question, one engine, one day.
• Nothing here survives a change in retrieval behaviour. On 27 July 2026 ChatGPT confirmed it had run no live web search for any of 78 blind questions, in which case no page feature on anybody's site could have decided anything.
• No guarantee. Adding all seven features to a page does not promise a citation, and no product including ours can.

Common questions

If I can only add one of the seven features, which one?

A date next to your most important number, and a source next to any figure about another company. That is the one our record speaks to most directly: on 17 September 2026 two comparison pages with correct competitor prices were set aside as advocacy purely because there was no link, no date and no source at read time.

Is a long page more citable than a short one?

We have not measured length and we will not guess. What our record shows is that the deciding factors named by an engine were published evidence and published scope, both of which take words but are not produced by adding words.

Does adding structured data make a page more citable?

It can help a machine read what is on the page and it cannot supply a fact that is not there. The failure recorded on 6 August 2026 was that the prices did not exist in what the crawler received at all, which no markup would fix.

Should my comparison page name competitors at all?

Yes, and source every claim about them with a link and a date, or expect the page to be read as advocacy. That is the recorded outcome for two such pages on 17 September 2026, and the engine was explicit that the problem was verifiability at read time rather than accuracy.

Can I win a question that an official portal answers?

Usually not the question itself. On 17 September 2026 Claude said plainly that no commercial product should rank above the official portal for a question about using that portal. The winnable questions are the adjacent ones: what the portal does not cover, what goes wrong, what it costs in time, and what to do with the result.

Would this audit work for one small site?

The candidate page part does, at a smaller size. Take ten queries, list the pages that could answer each, code them on the seven features, run the queries and see which got cited. You will not have a rate worth publishing, and you will have a list of specific gaps on your own pages, which is what the work needs anyway.

Sources and change log

• Tier_1/claude_response_17_09_audit.md. The 17 September 2026 audit: the unsourced competitor prices discounted as advocacy, the published coverage lists that outranked a competitor, the official portals above commercial products, the six questions with no third party review source, being named without being linked, and the dropped attribution.
• Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md. The 27 July 2026 blind run and its zero search confirmation, and the 6 August 2026 findings: the cited URL fourteen days after publication, the unreadable pricing page, the two conflicting public price lists, the unsupportable headline figure, and the competitor page with hidden instructions to answer engines.
• Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md. The 18 blind commercial questions of 18 August 2026 and the top source count of 1 of 18.
• geo-audits/aiknowsus-com/. September 2026, 24 batches and 72 conversations, including the batch of six with no citation and the three batches where the engine withdrew its search claim.

Change log. 27 July 2026 baseline. 6 August 2026 interim findings. 18 August 2026 commercial run. 17 September 2026 source selection audit. September 2026 own domain audit. Not run: the 200 query matched page audit, the candidate page coding, the per feature citation rates and the matched pair counts. When it is run this section will carry the query count, the candidate page count, the number of pairs that could be matched, and the counts for each of the seven features, including the features that show no difference.

What to do first

Take your three most commercially important pages and code them yourself against the seven features, honestly, yes or no. Most pages fail on three: no date next to the key number, no source next to a competitor's figure, and scope described with an adjective instead of a named list. Fix those three this fortnight, then run ten blind queries on two assistants and record which pages got cited instead of yours, so that in a month you have a before and after on the same list rather than an opinion.