Google's AI cites pages Google doesn't rank: 140 of 594 cited domains were on page one

How we measured it

The questions are the 40 fixed English buyer questions of our AI Source Index, in four shapes: 16 "best X" questions, 8 "alternatives to X", 8 head-to-head comparisons and 8 "how do I buy" questions. The AI side is every domain that Google AI Overviews and Google AI Mode cited for those questions in the weekly runs of 28 September and 5 October 2026. The organic side is Google's first results page for the same question, fetched once per question on 6 October 2026 through DataForSEO, for Google US in English, the same location the AI checks use.

• Unit: one domain in one answer. A domain is the lowercase host with a leading www. removed, so blog.x.com and x.com stay different. A domain counts once per answer however many of its URLs were cited, and its organic rank is its best rank on the page.
• Google's own shopping links are dropped, as the public index drops them. They are never an organic result.
• 38 of the 40 results pages were usable. The other two returned a provider error and those questions are left out of every rate, not counted as zero.
• The two runs are pooled. Every rate below is a count out of a count with a 95% Wilson interval.

Three of every four cited domains were not on page one

The two surfaces were close: AI Overviews 90 of 363 (24.8%), AI Mode 50 of 231 (21.6%). Neither one is a copy of the results page above or below it.

Much of the overlap that does exist is three platforms. reddit.com accounts for 39 of the 140 matches, youtube.com for 15 and facebook.com for 9. Take those three out and the overlap falls to 77 of 479 cited domains, 16.1%. For an ordinary company site, being on page one and being in the AI answer were mostly two separate outcomes in this panel.

Question shape: a split we cannot call

We expected the shape of the question to matter, and the pooled numbers seem to say it does: "best" questions 37 of 187 (19.8%), alternatives 31 of 129 (24.0%), comparisons 36 of 151 (23.8%), how-to-buy 36 of 127 (28.3%). We are not going to publish that as a finding.

The reason is the week-to-week movement inside a single surface. AI Overviews went from 35 of 185 (18.9%) on 28 September to 55 of 178 (30.9%) on 5 October. AI Mode went the other way, from 29 of 102 (28.4%) to 21 of 129 (16.3%). A swing of twelve points between two runs of the same surface is larger than the eight and a half points between the lowest and highest shape. With two runs, the shape split is inside the noise, and a third run could reorder it.

How this squares with Ahrefs' 38%

Our AI Overviews figure on its own is 90 of 363 cited domains, 24.8% [20.6–29.5%], and its interval sits below their 37.1%. We matched on domains, which is looser than their URL match and should push our number up, not down, so the gap is not a matching artefact. Three differences could explain it, and our data cannot tell them apart. Our 38 inputs are long buyer questions, not keywords, and a long question gives Google more room to answer from its own follow-up searches, which Ahrefs names as the reason its number halved. Our page one held 6 to 9 organic results rather than ten ranks. And our answers are seven months newer, with the same downward trend possibly still running. Read the two numbers together: whichever way you measure it, most of what Google's AI cites is not what Google ranks for the query you typed.

What could make these numbers wrong

• Page one is not ten results. Each results page held 6 to 9 organic results (8.0 on average); the rest of the page was AI Overview, People also ask, forums, video and products. "Page one" here means the organic results the page actually returned.
• The dates do not line up. The results pages are from 6 October and the answers from 28 September and 5 October, so the organic side is 1 to 8 days newer. The per-run table in the published summary shows how much that could matter.
• AI Mode answered without sources. In 76 checks it gave an answer with no sources 22 times and showed no answer 4 times; AI Overviews gave 8 source-less answers in 75. An answer with no sources has no domains to match.
• One sample each. One results page per question and one answer per surface per run. Google personalises and rotates both.
• Domain, not page. A match is the same domain, not the same URL: a domain cited for one article and ranking for another counts as overlap. That makes 23.6% an upper bound for page-level overlap.
• Two sensitivity checks barely move it: collapsing subdomains into their parent domain gives 23.9%, and keeping Google's shopping links as a cited domain gives 22.8%.

What to do with it

• Do not use your Google rank as a stand-in for your AI Overview presence. In this panel the two lists shared about a quarter of the AI side and about one in seven of the organic side. Measure them separately, per question.
• Read the sources Google's AI actually cited on your own questions. Most of them are not the page-one results you already watch, and that list tells you which sites are being read for your topic.
• Expect week-to-week movement. A single run can show a twelve-point swing on the same surface, so compare several runs before you decide anything changed.
• If you want to check our arithmetic, the rows are public: one line per question, surface, run and cited domain, and the full results page for each question. If you count it differently, we would like to know.

For the broader picture of what the answer engines cite, see what AI engines actually cite, and for how the two disciplines differ, GEO vs SEO.

Frequently asked questions