Google’s AI Overview does not cite the pages Google ranks.

An AI Overview appeared on 8 of 8 high-demand questions we put through it. Its cited sources overlap the organic results on the same page, at the same moment, by an average of 35%. Two-thirds of what it cites does not rank beneath it — and on one question, the two layers shared no domain at all.

Ipalibo Da-Wariboko · CULT+MATH · October 29, 2026 · 8 questions · one run each · United States · public data only

The decision this is for

Whether answer-engine work is a discipline of its own, or a second name for search ranking. The buyer is a CMO or Head of Search deciding whether to fund a separate workstream, or to tell an agency that the one they already have covers it.

Opposite answers, opposite actions. If the overview cites the pages that already rank, there is nothing separate to buy and the honest advice is to do SEO. If it cites different pages, the ranking a brand has paid for does not carry it onto the surface above, and that gap is the work. This measurement was registered with the first outcome as a live possibility, which is the only condition under which the second one means anything.

Every question, and what it cited

Lunch ideas for kids did not return: the call timed out mid-run. It is listed unrun rather than as an absence.

On “grocery list,” the two layers share nothing

Zero of 6. The ranked results and the generative answer above them have no domain in common.

A brand that wins that ranking wins nothing above it. The question is a shopping task rather than a recipe, and it is the shape where the two layers diverge furthest.

What follows for a brand

Sort the questions by shape before sorting them by volume. On named, specific, one-correct-answer questions the overview mostly repeats the ranking, and a separate workstream buys little. On open-ended and task-shaped questions it does not, and that is the addressable set.

An impression in an overview is not an impression you earned by ranking. They are different lists, produced at the same moment, on the same page. A report that counts only the first tells you nothing about the second.

And nobody can price it yet, because the click figure does not exist — not in Search Console, and not recoverable from analytics, where an overview click carries the same referrer as an organic one.

A correction, and it ran in our favour

The first pass reported 22% mean overlap. The engine was counting things that are not citations — the data provider’s own endpoint, Google’s thumbnail service, favicon URLs — 22 of 62 references.

Those inflate the denominator, and overlap is shared over cited. Every false reference lowered the measured overlap and made the overview look more independent of the ranking than it is. The defect pointed at the conclusion we were hoping for, which is the kind that has to be found rather than noticed.

Corrected, the figure is 35%, and that is the number used throughout. The cleaning rules now run at the source.

Limits

One category, one day, one run per question. Enough to show that the two layers differ, not how stable the difference is.

The organic comparator is the same response, which makes the comparison exact in time and means both layers come from one operator. A second generative search surface with its own index would make this independent, and has not been run.

Nothing here is a revenue claim. The data connecting a citation to a visit sits with Google, and the field is not published.

Method

Questions were drawn from measured US search demand rather than written by hand. Each was run once through the Google SERP advanced endpoint, which returns the AI Overview alongside the organic results; the overview’s cited domains and the organic domains were read from the same response.

A positive control ran first, alone — a query confirmed in a browser to carry an overview — because an earlier attempt used an endpoint that cannot return the field at all and would have reported nine absences as a finding. Overlap is recomputed from the cleaned citation list; the engine’s own percentage is never quoted.

The companion question — whether chat assistants read the web before answering these same questions — is at the retrieval signature study, and the category findings are in the CPG study.

The next one

Re-run when Google changes the surface, and on a second category to find out whether the question-shape split holds anywhere else. The working goes out first on Sound Decisions.