AI search citation statistics

What we counted

Every line here is from our own corpus. Quote any of them with the sample size attached.

• 53%115 of 218 cited page-and-question pairs, 203 distinct pages

  53% of the pages ChatGPT and Perplexity cited for a buying question do not appear anywhere in the first 100 organic results for that same question, according to CueScout.

  Ranking checked against Brave's index, 100 results deep. These are pages that earned a citation without earning a position.

• 20%43 of 218 cited page-and-question pairs

  Only 20% of cited pages also rank in the top 10 for the question they were cited for, according to CueScout.

• 28%60 of 218 cited page-and-question pairs

  A further 28% of cited pages rank somewhere between position 11 and position 100, according to CueScout.

  Buried but present. This is the group ordinary SEO can still move.

• 10 vs 3362 Perplexity answers and 104 ChatGPT answers

  Perplexity cites a median of 10 sources per answer. ChatGPT cites 3, according to CueScout.

  Medians across all grounded answers, including the ones that cited nothing.

• 73% vs 74%266 of 362 Perplexity answers, 77 of 104 ChatGPT answers

  The two engines cite at almost exactly the same rate: 73% of Perplexity answers and 74% of ChatGPT answers carried at least one source, according to CueScout.

  The gap between these engines is how many sources they cite, not how often they bother.

For comparison

Other people's numbers, credited. They are here because ours mean more next to them, not as filler.

• 36%465,823 AI Overview citations, German health queries

  36% of AI Overview citations also ranked in Google's top 10 for the same query, and 74% in the top 100, in a December 2025 study of 465,823 citations.

  Our top-10 overlap is 20% against their 36%, and our top-100 overlap is 47% against their 74%. Different engines, different index, one vertical against ten, so read it as a direction rather than as a contradiction.

  Source: SE Ranking, December 2025

How this was counted

• The corpus is every grounded answer CueScout logged from ChatGPT and Perplexity for a tracked buying question between 16 June and 17 August 2026: 466 answers, 98 distinct questions, 10 brands. Grounded means the engine retrieved and cited. Answers a model produced from its weights are excluded, because a URL a model remembers is not a citation.
• Two engines, and only two. We run ChatGPT and Perplexity. We do not read Google AI Overviews, Gemini, or Copilot, so nothing here is a claim about them.
• For the rank figures, each cited URL is searched against the exact question it was cited for, 100 organic results deep, and matched on normalised host and path. 218 page-and-question pairs resolved, covering 203 distinct pages. "Not in the first 100" means we searched that deep and it was not there.
• The ranking index is Brave, not Google. Brave runs its own index rather than reselling Google's, so these positions are Brave positions. We use it because it is the index we can query at depth through an API, and because a cited page that no independent index ranks is the finding either way.
• Numbers are recomputed by scripts/statistics/corpus-stats.sql. They will move as the corpus grows, which is why each one carries its own denominator and why the date at the top is the date of the run.

What this does not show

• Ten brands is a small corpus, and they are not a random sample. They are brands that came to CueScout because somebody already suspected they were invisible in AI answers, which skews toward smaller and newer companies than the web at large.
• Brave returned no forum cluster on any of these 218 lookups, so Reddit and similar threads are counted in the organic figures rather than separately. Reddit is the most-cited domain in the corpus, so that treatment matters: some share of the pages counted as unranked are threads a searcher would still find.
• We are not reporting how often these answers name a brand, or how often a cited page names the brand it was cited about, even though the query computes both. Our page reader is a static fetch with no JS rendering, and the most-cited domains here render their text in the browser, so both figures are biased downward by an amount we cannot measure. A number we cannot defend is worse than no number.
• This is one snapshot of one two-month window. It says what is cited now, not how anything got there.

Run these checks on your own site