The answer engine optimisation industry talks constantly about which sources AI assistants cite and where your brand ranks among them. It talks far less about a more basic number: how often AI answers cite anyone at all. That base rate turns out to be low — low enough that it should reframe how you set AEO expectations and read every visibility dashboard you’re shown.
This post assembles the citation-frequency numbers that can actually be verified at their owners — Similarweb’s panel data, Pew Research Center’s behavioural study of Google’s AI summaries, and the engine makers’ own documentation — and flags one widely circulated statistic that can’t be.
The base rate: citations in single-digit percentages of prompts
The best owner-published data on citation frequency in ChatGPT comes from Similarweb. Its “AI Search Stats 2026” report (published 29 July 2026) states that citation presence in US ChatGPT prompts rose from about 1.6% in June 2025 to roughly 6.8% by May 2026 — more than quadrupling over eleven months, but still leaving the overwhelming majority of prompts producing answers with no citation whatsoever.
The category spread matters as much as the average. In Similarweb’s breakdown, Travel & Hospitality prompts carried a citation about 23% of the time — the highest rate tracked — with Automotive at roughly 20%, while Professional Services sat under 4%. If you sell B2B services, the honest baseline is that the engine you’re optimising for attaches a citation to fewer than one in twenty-five relevant prompts.
Scope notes before you quote these numbers anywhere: this is US panel data on consumer ChatGPT usage, not global and not API traffic; and the metric is citation presence — the share of prompts whose answers include any citation — not how many citations an answer carries, and not whether anyone clicks them.
The number we could not verify (and what we found instead)
Several AEO roundups circulate a claim, attributed to Similarweb, that the share of ChatGPT answers including citations rose from 0.6% in January 2025 to 2.8% in August 2025. We went looking for the owner source. Similarweb’s current AI search statistics page does not contain either figure, and searches for the exact numbers surface only third-party roundups citing each other, not a first-party Similarweb page.
The figure may once have existed in an earlier chart or press briefing, but it currently fails the only test that matters: you cannot read it at the owner’s own page. The verifiable owner data (1.6% to 6.8%) tells a directionally similar story — low base, rising fast — so we’ve used that instead and flagged the substitution. It’s the same discipline we describe in reading 2026 AEO benchmark reports honestly: a statistic that exists only in roundups is unverified, whatever its pedigree appears to be.
Why most answers cite nobody: citations live in retrieval mode
The low base rate isn’t a bug the engines are racing to fix — it’s structural. A large language model answering from its training weights has nothing to cite; there is no retrieved document behind the sentence. Citations appear when the system performs retrieval — a web search, a grounding call, a knowledge-base lookup — and attaches the fetched sources to the generated text. The engine makers’ own documentation is explicit about this split.
| Engine / mode | When citations appear | What the owner’s documentation says |
|---|---|---|
| OpenAI models (developer API), plain generation | No citations — the answer comes from model weights | OpenAI’s web search tool documentation only describes citations in the context of search results; there is no citation mechanism for ungrounded generation |
| OpenAI models (developer API) with web search tool | Inline citations attach by default when search runs | “By default, the model’s response will include inline citations for URLs found in the web search results” — and developers “must” make them “clearly visible and clickable” |
| Gemini models (developer API), plain generation | No citations | Grounding documentation describes citations only for grounded responses |
| Gemini models (developer API) with Grounding with Google Search | Inline citation annotations when grounding succeeds | When “a response is successfully grounded”, output includes annotations where each “url_citation annotation links a text segment… to a source URL” |
| Perplexity | Citations are the default product design | Perplexity’s developer documentation describes “web-grounded answers with built-in citations in one call” |
The practical consequence: citation frequency is really retrieval frequency. Your brand cannot be cited in an answer where retrieval never fired, no matter how well optimised your content is. Which is why the first question to ask of any AEO vendor is not “can you get us cited?” but “what share of the prompts you’re targeting actually trigger retrieval?” — a base rate very few pitch decks lead with.
If you’re building an AI-era pipeline and want partners who show you the denominators, not just the wins, you can Apply For Partnership with Zian.
Even when citations appear, almost nobody clicks them
Citation presence is one filter; click behaviour is another. Pew Research Center’s study of 900 US adults’ real browsing behaviour — 68,879 unique Google searches across March 2025 — is the cleanest owner-published data on what happens after a citation appears in Google’s AI summaries:
- Users who saw an AI summary clicked a traditional result link on 8% of visits, versus 15% when no summary appeared.
- Clicks on the sources cited inside the AI summary occurred in just 1% of all visits.
- Users ended their browsing session entirely on 26% of pages with an AI summary, versus 16% of pages with only traditional results.
So the funnel narrows twice: most AI answers carry no citation, and cited sources are clicked on about one visit in a hundred even when they do. That is not an argument against AEO — being the source an engine trusts still shapes what the answer says about your category, which matters even at zero clicks. It is an argument against traffic-based business cases for citation work.
Which queries actually trigger retrieval
Pew’s data also shows that retrieval isn’t random — query shape predicts it. In the same study, only 8% of one- or two-word Google searches produced an AI summary, rising to 53% for queries of ten or more words, and 60% for queries beginning with a question word. Overall, 18% of all searches in the study produced a summary.
| Query characteristic (Google, March 2025, Pew panel) | Share producing an AI summary |
|---|---|
| All searches in study | 18% |
| One- or two-word queries | 8% |
| Queries of ten or more words | 53% |
| Queries starting with a question word | 60% |
| Full-sentence queries (noun + verb) | 36% |
Combine this with Similarweb’s category spread and a workable heuristic emerges: retrieval fires on long, question-shaped, freshness-sensitive and comparison-type prompts, and stays dormant on short navigational or definitional ones. That’s where your citable pages should aim.
Setting AEO expectations against frequency, not just rank
Three planning rules fall out of the base rates:
1. Multiply before you celebrate
A “top-3 citation” for a tracked prompt is worth: (probability that real users’ versions of that prompt trigger retrieval) × (probability your source survives that day’s retrieval) × (~1% click-through if you’re counting traffic). Rank is the last multiplier, not the first. And because engines re-roll their retrieval frequently, a citation held today can vanish tomorrow — we’ve measured this directly in our work on AI answer volatility.
2. Expect concentration inside the small cited slice
The minority of answers that do cite sources draw disproportionately from a familiar shortlist — in Pew’s data, Wikipedia, YouTube and Reddit were the most frequently cited sites in Google’s AI summaries, and government domains appeared markedly more often there than in standard results. We’ve covered who wins that concentrated game, and how a challenger brand gets onto the list, in our analysis of AI citation source concentration; this post’s job is only to establish how rarely the game is played at all.
3. Be present when retrieval fires
Since you can’t force retrieval, position for the prompts where it’s likely: publish pages that directly answer long, question-form queries; keep dates, entities and figures current; and make claims specific and sourced, because systems attaching citations favour pages that look like citable evidence. In under-retrieved categories like professional services, fewer competitors are trying — a below-4% citation rate cuts both ways.
Zian applies the same base-rate honesty to sales outreach: our SmartReach AI™ and PrecisionPitch AI™ agents are optimised against real success outcomes, not vanity metrics. If that’s the operating style you want in a partner, Apply For Partnership.
Sources & ownership
| Claim | Owner (organisation) | Owner URL | Date checked |
|---|---|---|---|
| Citation presence in US ChatGPT prompts rose from ~1.6% (June 2025) to ~6.8% (May 2026); Travel & Hospitality ~23%, Automotive ~20%, Professional Services under 4% | Similarweb | https://aisearch.similarweb.com/blog/gen-ai-stats/ | 2026-08-22 |
| Circulating “0.6% to 2.8%, Jan–Aug 2025” figure attributed to Similarweb | Not verifiable at owner — absent from Similarweb’s current page; found only in third-party roundups | (no owner URL exists; superseded by the Similarweb page above) | 2026-08-22 |
| 8% vs 15% link-click rate with/without AI summary; 1% clicked cited sources; 18% of searches produced summaries; 8%/53%/60%/36% trigger rates by query shape; 26% vs 16% session-end rates; 900 adults, 68,879 searches, March 2025; Wikipedia/.gov/Reddit/YouTube among top-cited | Pew Research Center | https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/ | 2026-08-22 |
| OpenAI web search tool: inline citations included by default; must be visible and clickable | OpenAI | https://developers.openai.com/api/docs/guides/tools-web-search | 2026-08-22 |
| Gemini API: citation annotations returned when Grounding with Google Search succeeds | https://ai.google.dev/gemini-api/docs/grounding | 2026-08-22 | |
| Perplexity: “web-grounded answers with built-in citations in one call” | Perplexity | https://docs.perplexity.ai/getting-started/overview | 2026-08-22 |
FAQ
What percentage of ChatGPT answers include citations?
Per Similarweb’s published US panel data, citation presence in ChatGPT prompts rose from about 1.6% in June 2025 to roughly 6.8% by May 2026. Rates vary sharply by category — around 23% for travel and hospitality prompts but under 4% for professional services — so most answers, in most categories, cite nobody.
Why don’t most AI answers include citations?
Because citations only attach when the system retrieves external content. Answers generated purely from model weights have no source documents behind them. OpenAI’s and Google’s developer documentation both describe citations exclusively as a feature of web-search or grounding modes, so citation frequency is effectively retrieval frequency.
Do people click the sources AI answers cite?
Rarely. Pew Research Center’s March 2025 study of 900 US adults and 68,879 real Google searches found users clicked a source cited inside an AI summary in just 1% of visits, and clicked traditional results on only 8% of visits when a summary appeared, versus 15% without one.
Which queries are most likely to trigger citations?
Long, question-shaped queries. In Pew’s data, only 8% of one- or two-word Google searches produced an AI summary, versus 53% of queries with ten or more words and 60% of queries starting with a question word. Comparison, compliance and freshness-sensitive prompts are the ones most likely to fire retrieval — and therefore citations.
Is the “0.6% to 2.8%” ChatGPT citation statistic real?
We could not verify it. The figure circulates in AEO roundups attributed to Similarweb, but it does not appear on Similarweb’s current AI search statistics page, and searches surface only third parties citing each other. The verifiable owner data shows 1.6% to 6.8% between June 2025 and May 2026 — similar direction, different numbers. Treat the original claim as unverified.
Does a low citation rate mean AEO is a waste of time?
No — it means AEO business cases should be built on answer influence, not traffic. Being a trusted source still shapes what engines say about your category even when nobody clicks, and in low-citation verticals fewer competitors are contesting the retrieval slots that do exist. It does mean you should discount any pitch that quotes citation rank without citation frequency.