How should you compare a competitor's AI-cited pages?

By Abiot Y. Derbie - Founder, RankEcho - Updated 2026-09-06

What question does a same-domain comparison answer?

It answers a descriptive question: among pages sampled from one competitor domain, which measured traits differ between URLs visibly cited in a registered answer panel and a matched comparison pool not observed cited in that panel? It does not answer why the system selected any URL.

A direct your-site-versus-rival comparison mixes page treatment with site history, brand recognition, link profile, source coverage, topic portfolio, and other domain-level differences. Using one domain for both sets reduces that bundle. It does not hold page topic, age, template, crawl history, indexing, freshness, or query relevance constant unless you match those separately.

How should you build the observed-cited set?

Start with a fixed prompt-engine-repeat panel and retain every eligible answer. A competitor URL enters the observed-cited set only when a visible citation in one of those answers resolves to the registered competitor domain. Preserve the exact answer, prompt, engine, run time, URL, and citation placement when available.

Deduplicate URLs only for the page-level trait comparison; retain every cell-level appearance separately for frequency reporting. A cited URL is an observed attribution for that answer, not proof that the provider indexed it through a particular crawler, used it for every claim, ranks it highly, or will cite it again.

• Report eligible cells and unavailable/error cells for the panel.
• Normalise redirects and canonical variants without discarding the raw URL.
• Verify domain ownership before assigning subdomains or syndicated copies.
• Keep repeated citations as cell-level observations even when page traits are measured once.

How should you choose the comparison pool?

Call the second set 'not observed cited in this panel,' not 'ignored by AI.' A crawl or sitemap can identify candidate pages, but it cannot prove that an engine saw and rejected them. Record the collection date, discovery method, fetch status, inclusion criteria, and exclusions.

Match pages as closely as the available inventory allows: documentation with documentation, comparison pages with comparison pages, similar topics with similar topics, and recent pages with similar publication windows. Select the rule before measuring traits so the pool is not chosen to manufacture a large difference.

• Define the eligible content types and language before sampling.
• Exclude login walls, redirects, duplicate canonicals, and failed fetches transparently.
• Use all eligible pages or a reproducible random/matched sample.
• Publish both set sizes and flag thin or badly imbalanced samples.

Which page traits are useful hypotheses?

Measure traits that can be defined consistently from public HTML: a self-contained opening answer, question-shaped headings, a table, visible supporting links, paragraph length, update date, or valid structured data that matches the page. These are page descriptions, not universal AI-selection factors.

Do not turn correlation into provider preference. Google explicitly says its AI Search features need no special AI schema, and normal Search eligibility still does not guarantee appearance. A structured-data gap may justify a quality review; it does not show that adding markup will create a citation.

• Write a detection rule for every trait before comparing sets.
• Inspect false positives such as navigation links counted as sources or layout tables counted as data tables.
• Report numerator, denominator, and percentage for both sets.
• Describe a gap as associated with the observed set until a separate validation test exists.

What does a worked synthetic comparison show?

Suppose a fixed panel yields eight distinct observed-cited competitor URLs. A pre-registered matched sample supplies 20 same-domain pages not observed cited in that panel. Six of eight observed-cited pages have a self-contained opening answer (75.0%); seven of 20 comparison pages do (35.0%). The descriptive gap is 40.0 percentage points.

That 6/8 versus 7/20 result is a useful hypothesis, not an effect estimate. The observed-cited set might contain more evergreen guides while the comparison pool contains product announcements. After matching only guide pages of similar age, the gap could shrink or disappear. Report both analyses instead of keeping only the stronger number.

Which checks can disconfirm the page-trait hypothesis?

First test whether the difference survives matching by topic, page type, age, template, language, and fetch status. Then inspect whether the trait was present before the observed citation and whether pages without it are still cited in other cells. A trait that disappears after matching or has many counterexamples is a weak implementation target.

Also test rival hypotheses: the prompt may map more closely to the cited page, the cited URL may carry unique first-party data, an external source may corroborate it, or ordinary answer variability may select different sources on repeat runs. The method should be allowed to return 'no actionable separator.'

• Does the gap persist within the same content type and topic?
• Was the trait present at the time of the citation observation?
• Do repeated runs select the same URL or different pages?
• Are there observed-cited counterexamples without the trait?
• Does the page offer distinct evidence rather than only a formatting difference?

How does this analysis move from Find to Fix to Prove?

Find produces the two reproducible URL sets, trait counts, matching rules, counterexamples, and alternative explanations. Fix applies one bounded, truthful change to an appropriate owned page only when the surviving hypothesis fits that page. Copying a competitor's prose, claims, or design is not part of the method.

Prove repeats the exact registered prompt-engine cells after the change window and retains unchanged and negative results. If an owned URL becomes cited, report the observed before/after movement and its competing explanations; do not say the competitor analysis caused the selection.

Which false positives and limitations matter most?

A citation parser can misread redirectors, footnotes, navigation URLs, syndicated copies, or domains whose names overlap. A crawler can mistake hidden, templated, or stale HTML for the page a user saw. A small observed-cited set can swing sharply when one page is added, and many traits tested at once increase the chance of a coincidental gap.

The method cannot observe undisclosed provider inputs, prove that a not-observed page was considered, isolate domain reputation completely, or guarantee that a copied trait will transfer across sites. If no competitor URL is visibly cited in the panel, there is no observed-cited set and this specific comparison should not run.

• Publish sample sizes, missing pages, fetch failures, and all traits tested.
• Use percentage points and raw counts; avoid significance language on tiny descriptive samples.
• Keep the analysis date because pages and answer systems change.
• Prefer 'not observed cited' to 'uncited' or 'ignored.'

What do official provider sources allow you to conclude?

Google says normal Search eligibility and snippet controls govern supporting-link eligibility in AI Overviews and AI Mode, with no additional technical requirements or special AI schema; eligibility does not guarantee selection. Bing's February 2026 AI Performance announcement says citation count is not placement, ranking, authority, or page importance, and its grounding-query data is sampled.

OpenAI documents OAI-SearchBot, GPTBot, and ChatGPT-User as separate controls for search, potential training, and user-triggered actions. Bot permission can be checked as one access condition, but it does not show which pathway produced a citation or guarantee a future answer. The claim-level official-source records below were checked on 2026-09-01.

Sources reviewed

Provider eligibility and measurement claims below were checked against primary documentation. These records do not establish a universal selection formula, causation, or a guaranteed ranking, impression, recommendation, or citation.

Frequently asked questions