Why does ChatGPT cite competitors that rank below you?

Google ranks pages. ChatGPT constructs an answer.

Google ranks pages against a query. ChatGPT constructs an answer, and it does not necessarily search for the query you were optimising for. OpenAI states that ChatGPT search typically rewrites a prompt into one or more targeted queries, and that after reviewing the first results it may send additional, more specific queries.[3]

1. 01

   Prompt

   The question the user actually types.

3. 02

   Rewrite

   One or more targeted searches, which may never include the original wording.

5. 03

   Retrieve

   A candidate set wider than any single results page.

7. 04

   Synthesise

   Claims assembled into one answer.

9. 05

   Cite

   Supporting sources attached to those claims.

A broad CRM prompt can become separate searches for manufacturing fit, implementation effort, pricing, integrations and customer evidence. Each of those has its own best source.

The domain often survives. The ranking page often does not.

Ahrefs ran 3,311 short-tail head terms through ChatGPT, Perplexity and Google top 100 results, then measured how often an AI citation matched something Google ranked. For ChatGPT the gap between the two levels of matching is the finding.[1]

Google may rank competitor.com/product while ChatGPT cites competitor.com/research/buyer-guide. The domain stays relevant; the chosen evidence changes. Your competitor may not be beating you on the visible keyword at all — it may be winning a search generated behind the answer.

Overlap falls as the question narrows.

Across a separate set of 15,000 long-tail prompts, an average of 12% of the URLs cited by ChatGPT, Gemini, Copilot and Perplexity ranked in Google top 10 for that prompt, and roughly 80% did not rank for it at all.[2] ChatGPT on its own sits below that four-assistant average, and it falls further as the query gets closer to what the system is really searching for.

The pattern runs the wrong way for anyone hoping to reverse-engineer the retrieval path. Even when the measured query is the one the system generated itself, overlap with the ranked results is at its lowest.

Accessible is not the same as cited.

AI visibility becomes diagnosable once it is split into stages. Each stage answers a different question and fails for a different reason, so a single visibility score tells you nothing about which one broke.

1. 01

   Eligibility

   Documented

   Can ChatGPT access and surface the site at all? OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links.[4]

   Blocked crawlersIP range blocksInaccessible pages

   MeasureOAI-SearchBot access in robots.txt and at the CDN

2. 02

   Retrieval

   Inferred

   Does the page surface for the searches the system generates, rather than the one the user typed?

   Weak query alignmentIndexation gapsThin topical coverage

   MeasurePresence across fan-out style query variations

3. 03

   Selection

   Inferred

   Among the candidates retrieved, is this the most useful source for the claim being made?

   Insufficient evidenceIntent mismatchWrong format

   MeasureSource-fit comparison against the cited page

4. 04

   Citation

   Observable

   Is the page attached to a claim in the visible answer? OpenAI documents a larger set of consulted sources behind a smaller set of inline citations, and notes the source count is often greater than the citation count.[5]

   Consulted but not citedAnother source proves it better

   MeasureCited URLs per tracked prompt

5. 05

   Recommendation

   Observable

   Is the brand named in the answer, with or without a link back?

   Weak relevanceMissing proofNo third-party validation

   MeasureBrand mention and recommendation rate

Access is an eligibility condition, not a citation guarantee. A page can rank, be retrieved and even be consulted without ever receiving a visible citation.

Why can the lower-ranking page be a better source?

OpenAI publishes no list of organic citation factors, and states plainly that ChatGPT ranks search results using multiple factors and that placement is not guaranteed.[3] What follows are practical audit criteria and testable hypotheses — not confirmed ranking signals.

Different questions need different sources.

A single page cannot be the best source for every question in a category. Working out which question a citation was answering is usually faster than guessing at the page.

Where your library has no page in the right format, no amount of optimisation on the pages you do have will close the gap.

Directness is not the same as brevity.

Both passages below are the same length. Only one of them can be lifted into an answer without the model having to work out what is being claimed.

Analyse the claim–source relationship, not only the page.

Plot every tracked prompt against two axes: whether you rank, and whether you are cited. The four cells are four different problems, and only one of them is a ranking problem.

Run the citation audit in six passes.

The work is comparative. You are not scoring your page against a checklist; you are working out what the cited page gave the answer that yours did not.

1. 01

   Build prompts around decisions

   Cover problem, solution, category, comparison and validation questions. One broad keyword tells you almost nothing about retrieval.

   RecordA fixed prompt set you can re-run unchanged

2. 02

   Record both surfaces separately

   Capture ranking URLs and cited URLs as different objects, alongside brand mentions, date, location and test conditions.

   RecordPrompt, retrieved page and cited page as three fields

3. 03

   Compare domains and exact URLs

   Establish whether the citation went to the ranking page, another page on the same domain, documentation, research or a comparison.

   RecordDomain overlap and URL overlap, tracked apart

4. 04

   Reconstruct the evidence need

   For each citation, identify the definition, fact, comparison or recommendation it was supporting.

   RecordThe claim each citation is attached to

5. 05

   Run a source-fit comparison

   Intent, direct answer, evidence, buyer context, freshness, crawl access and internal linking, side by side.

   RecordA row per dimension with a named action

6. 06

   Improve the right asset

   Update an existing page where the intent already matches. Create a new one only where the decision, the format or the evidence genuinely differs.

   RecordOne decision per gap, with the reason written down

Compare your page with the cited page.

The comparison that produces a work item rather than an opinion. Four dimensions is usually enough to find the difference.

Where every row comes back even, the gap is probably eligibility or retrieval rather than selection — go back to the five stages.

Track source performance beside search performance.

Rankings remain useful, but they describe one part of visibility. Microsoft makes the same point about its own reporting, noting that citation activity in Bing AI Performance does not represent ranking, authority or importance.[6] Treat citation counts as observation, not as a scoreboard.

A high ranking signals competitiveness. A citation signals usefulness to one answer.

A high ranking signals that a page is competitive for a query. A citation signals that it was the most useful source for one claim in one answer. The two are connected, but they are not interchangeable, and the evidence says the connection is weaker than most reporting assumes.

Your ranking may well improve discoverability. ChatGPT can still rewrite the prompt, retrieve a wider candidate set, consult more sources than it ends up citing, and select another page with stronger intent, passage or evidence fit.

The practical response is to stop treating a citation as an extension of a keyword position. Track search performance and source performance together, then go and study the lower-ranking pages that keep winning citations — they are telling you what the answer needed and your page did not supply.