ChatGPT citation statistics

What does ChatGPT actually cite?

Independent teardowns of how ChatGPT selects sources, how it differs from Claude and Perplexity and the platform-level citation patterns all describe the same shape without publishing a corpus anyone can recompute. The most-cited domains page holds the per-engine comparison, and the concordance study runs a fixed panel instead of a vendor corpus.

The Wikipedia skew, and the caveat that travels with it

Wikipedia's prominence in ChatGPT citations is one of the most reliably reproduced observations in this field, and one of the most frequently misused. The figure is meaningless without the caveat, so they should never be separated: a 47.9% encyclopedic share is measured across a query mix weighted toward topics an encyclopedia covers well. On broad definitional questions an encyclopedia is close to an ideal retrieval source. On specific, current, practical or data-bearing questions it has nothing to offer at all, and the share on those questions is not 47.9% — it is unmeasured, because nobody has published the query-type breakdown.

Hypothesis Why the skew exists: Wikipedia is broad, structured, neutral in register, densely interlinked, freely licensed and unusually consistent in format. Almost every property you would design into an ideal retrieval source, it has. That is inference from the observed pattern, not a mechanism OpenAI has described.

What it does not mean is that you should try to be Wikipedia, or that editing Wikipedia is a visibility tactic — that is against its rules and it backfires. Read as a map of where not to fight, the skew is genuinely useful. The remaining share is the tail where every non-encyclopedic source competes, and it is reached by being specific and current rather than comprehensive and neutral. The Wikipedia dependency study asks the sharper version of the question — what ChatGPT cites when the encyclopedia has nothing on a topic — and the gaps it finds are where an ordinary publisher can win.

The Bing lineage: ChatGPT's most under-checked input

ChatGPT Search performs live retrieval at query time rather than answering purely from training data, and it has historically drawn in part on a Bing-derived index alongside OpenAI's own crawl infrastructure. The exact current architecture is not fully disclosed and has likely evolved. The lineage is not speculation, though. Seer Interactive found 87% of SearchGPT citations matching Bing's top results, which makes Bing's index and user base far more relevant to ChatGPT visibility than its consumer market share suggests. The Bing and Copilot guide covers that surface on its own terms.

This is the distinctive fact about ChatGPT among the assistants, and the practical one. A team optimising hard for Google while never checking Bing Webmaster Tools is neglecting the input with the tightest reported relationship to its citations. Retrieval also depends on the page being fetchable at all — whether these systems execute JavaScript is covered in the rendering study, and the hygiene checklist is the technical GEO audit.

Why Google ranking barely predicts a ChatGPT citation

At roughly 12%, Google ranking is only weakly associated with a ChatGPT citation — a sharp contrast with AI Overviews at about 38%, itself down sharply year over year. Four reasons make that plausible rather than surprising.

Different index. ChatGPT's retrieval has not historically run on Google's, and two corpora with different crawl priorities disagree about what exists before they disagree about what is good. Different unit. Google ranks pages against a query; retrieval selects passages that support composing an answer, and a page can be an excellent supporting passage and a mediocre standalone result.

Fan-out. Retrieval is often for a sub-question rather than the typed query, so comparing citations against rankings for the original query compares two different questions. The fan-out corpus study measures that directly. And no click competition: Google's ranking has been shaped for years by what people click, and retrieval has no equivalent pressure.

The practical conclusion is uncomfortable and clear: your Google rankings are a weak proxy for your AI visibility, and using one to report the other will mislead you in both directions.

GPTBot vs OAI-SearchBot: the switch that matters

Two crawlers means two separate decisions, and conflating them is the most common error we see. A site that blocks OAI-SearchBot while allowing GPTBot removes itself from ChatGPT Search citations while still contributing to future model training — likely the opposite of its owner's intent.

Both tokens are documented by OpenAI in the platform bots page and the developer docs; the consolidated list across every vendor is the bot user-agent registry. How many sites pull each lever is measured in the robots.txt blocking census and independently in publisher-blocking studies; how much traffic each represents is in the crawler statistics hub and Cloudflare's crawler breakdown.

That lineage matters when reading older statistics or screenshots. A citation pattern documented under the early browsing feature does not describe current ChatGPT Search behaviour, and the figures on this page describe the current-generation product as of their stated dates.

When ChatGPT does not search at all

Not every answer involves retrieval. The model can answer from training, and when it does there is no source list and no citation opportunity for anyone. That matters more than it first appears. A citation rate computed over a query set including many non-searching answers measures two things at once, and two studies with different query mixes will disagree for reasons unrelated to either engine or either publisher.

It also bounds what optimisation can achieve. On questions answered from training, no on-page work reaches the user, because the page is never fetched. Leverage concentrates on questions specific, current or contested enough to trigger a search. When you build your own panel, mark which answers searched — it turns a noisy average into two clean ones.

Cited, linked, named: three different outcomes

Almost every citation-rate statistic collapses three outcomes worth very different amounts. Retrieved means your page was fetched and considered — invisible to the user, invisible to you, and producing nothing on its own. Linked means your page appears in the source list; a user who looks at sources might click, and most do not. Named means your brand is mentioned inside the answer text, which is the outcome with real value because it reaches the reader who never opens a source list.

Almost no published study separates linked from named, and named is the one a marketing team actually wants. When you run your own measurement, record them in separate columns. It costs nothing and it is the single biggest improvement you can make over the public figures.

Referral volume, meanwhile, stays a small fraction of total traffic for most sites even where conversion quality is reported higher. The conversion benchmarks page and the referral traffic page handle those claims properly, since the published multiples vary too widely to restate as one number.

Measuring your own ChatGPT presence

Twenty to fifty real questions in customer wording, from support tickets and sales calls rather than internal phrasing. Each run three times, because variance between runs is real and one run tells you nothing. Five fields recorded per run: did it search, were you linked, were you named, who else was cited, and what did it get wrong. Conditions held constant and written down — same account state, same region. Repeated quarterly with the raw records kept, because the value is entirely in the trend. A single snapshot is a curiosity; four quarters is evidence, and the reporting conventions are in the measurement standard.

When you report upward, lead with the trend rather than the level. Show the competitor column: who else got cited on your questions is far more persuasive than your own count. And name what you are not claiming — attributed revenue, above all. Propose a review date rather than a decision, since this channel does not yet support confident budget calls.

What your server logs can and cannot tell you

Because OpenAI runs separately named crawlers, log analysis is more informative here than for Google's surfaces — the structural reason Gemini research is so much thinner. It still has hard limits.

Logs can tell you that a fetch happened, which agent made it, which URL it wanted and what status you returned. That is enough to catch the two most common real problems: blocking something you did not mean to block, and returning errors to a crawler while serving humans fine. Logs cannot tell you whether the fetch produced a citation, whether you were named, or how many people read the answer. There is no path from a request line to an outcome.

And verify user agents properly. A user-agent string is self-reported and trivially forged. Confirm against published address ranges before treating a log line as evidence, and be sceptical of any crawler-traffic figure that skipped this step.

Five mistakes in reading these numbers

Treating a source-mix study as a target list. Knowing which domains dominate tells you where the competition is, not where to publish.

Comparing figures across studies. Different query panels, dates, regions and counting rules produce numbers that are not comparable, however similar they look.

Reading correlation as instruction. Cited pages sharing a trait does not mean the trait caused the citation — the standing problem across the tactic evidence scoreboard.

Assuming stability. These products change without announcement. Our working rule: source-mix figures stay directionally useful for roughly six to twelve months, and specific percentages should be treated as historical after that. Anything about crawler names or tokens should be checked against primary documentation before you rely on it, because that layer changes fastest and breaks most silently.

Ignoring the denominator. A citation rate over queries where the engine never searched is measuring something else entirely.

Open questions

Open question How often does a query trigger retrieval? The denominator under most citation statistics, and nobody has published it.

Open question What is the naming rate versus the linking rate? Measurable with a modest panel. Not published anywhere we have found.

Open question How stable are citations for the same query over time? The citation half-life study is the attempt to measure it, and the latency study asks it from the other end. Our first-party zero-to-cited log study is where both started.

Open question Does interface matter? Web, mobile and API access may retrieve differently. Registered as open, because we have no evidence either way.

Verification status

The ranking-overlap figure traces to Ahrefs with a disclosed sample — Traceable . The source-skew figure is vendor-reported from a non-public corpus — Partial — and should never be quoted without the query-mix caveat above. Conversion claims are treated as Broken chain per the provenance audit until a reconciled, like-for-like figure exists. The same scheme runs on GEO statistics and AEO statistics; the standards behind it are in the method notes.