We asked AI 20 buyer questions three times. Four in five sources it cited do not rank on Google.

When someone asks Gemini or ChatGPT "what is the best X," the model names sources. The common assumption is that those sources are the pages that rank on Google. We tested it, and we tested it carefully, because the last time we ran this we made two mistakes worth admitting: we compared against only Google's top three, which inflates the gap, and we ran each query once, which for a grounded model is a coin flip.

So we ran citation_domains against 20 "best X" buyer questions, three times each, and for every query compared the domains Gemini cited to the domains in Google's top ten. Sixty grounded runs, one honest baseline.

The headline: a 79% gap, measured generously

Across the 20 queries and three runs each, Gemini cited 365 distinct domains. 76 of them also sit in Google's top ten for the same query. That is a 21% overlap, a 79% gap.

We used the top ten, not the top three, on purpose. Comparing against only the top three would report a 91% gap and flatter the point we are making. Even measured generously, against a full page of Google results, four of every five sources AI cites are not the pages that rank. Ranking on Google and being cited by AI are mostly different games.

The overlap that exists is Reddit and YouTube, not you

Look at what those 76 overlapping domains actually are. 27 of them are user platforms, Reddit and YouTube above all. Reddit appears in the overlap for 13 of 20 queries, YouTube for 12. Strip the platforms out and the overlap between "ranks on Google" and "cited by AI" is 49 domains out of 365, about 13%. If you are not Reddit or YouTube, the two channels barely touch.

AI does not answer the same way twice

This is the finding that changed how we measure. We asked each question three times because a single run lies, and it does. 52% of the domains Gemini cited appeared in only one of the three runs. A single run surfaced about 10 domains; three runs surfaced 18. Measure once and you see a little over half the picture, and half of what you see will not be there tomorrow.

Two queries prove the cost of measuring once. "notion alternative" and "best web analytics tool" returned no citations at all on an earlier single run, which looked like "there is nothing to win here, only brand memory." Across these three runs they cited 21 and 24 sources. There is rarely no citation to win. There is a citation you did not catch.

AI cites small sites Google will not rank, and cites them consistently

Best CRM for startups is the clearest case. Google's top ten is what you would expect: a Reddit thread, Salesforce, HubSpot, Zoho, Zendesk, Quickbase. Gemini cites those names too, but only sometimes. The domains it cited on at least two of the three runs include ahoy.ai, baserow.io, withallo.com, lightfield.app, and cloudtalk.io, small tools that rank nowhere near the first page. The giants flicker in and out of the answer; several unknown tools are cited every time.

The pattern repeats: nitrosend.com and agentmail.to for transactional email, focusreactive.com for headless CMS, scrapfly.io for scraping. A page that cannot crack Google's top ten can still be a stable AI source, because the engine retrieves the passage that answers the sub-question, not the domain with the most backlinks. The citation layer is winnable by sites that cannot win the ranking layer.

The honest denominator, query by query

query	AI-cited (3 runs)	in Google top 10	in Google top 3
best crm for startups	23	6	3
notion alternative	21	6	3
best password manager	19	5	3
best vpn	19	5	3
best vector database	15	5	3
best web scraping api	18	4	1
stripe alternative	16	4	2
best data observability tool	21	3	0
best ai coding assistant	16	2	0
best api documentation tool	16	0	0
all 20 queries	365	76 (21%)	34 (9%)

The top-three column is why we do not report it as the headline. It makes the gap look larger than a fair test supports.

What this means

The practical takeaway is the one the whole product is built on, with one correction. You do not get into AI answers by out-ranking the incumbents. You get in by being the page the engine retrieves and cites for a specific sub-question, which is a different and more winnable target. But because the citation layer is unstable, you cannot measure it once and call it done. You measure it repeatedly and read the domains that show up every time, because those are the durable slots.

citation_domains shows you exactly which domains hold the answer for your queries. Run it several times and watch which names are stable. Those are the ones you are really competing with, and often they are smaller and more beatable than the brands that own Google.

Method

20 "best X" and category buyer queries, each run through citation_domains with engine gemini three times, on 2026-08-10. Gemini grounding is stochastic, which is the reason for three runs and for reporting the union across them. Google's baseline is the top-ten organic domains from one Grounder SERP call per query. Overlap is the intersection of the AI-cited union with Google's top-ten domains; the top-three column is shown for contrast only. Every query grounded on all three runs, so none are excluded. Domains are normalized (www stripped, lowercased) and deduplicated. No scoring, no verdict, raw domains only.

How to run it

citation_domains(query="best X for Y", engine="gemini") returns the cited domains for one query. Run it three times and keep the domains that recur, then run serp_profile on the same query to see how far those cited domains sit from who actually ranks.