Query fan-out decides AI citation before a page is ever read

Two measurements published nine days apart describe the same mechanism, and neither one names the other. Read together they locate where AI citation is actually decided — and it is not on the page.

The engine writes its own queries, and what it writes predicts the answer. An independent analysis of raw traffic from a ChatGPT Plus account found the model injecting brand names the user never typed into its own background search queries, in 21 of 27 tested conversations. Brands ChatGPT named in its own query reached the final answer 68.9% of the time. Brands that surfaced only on the pages it retrieved reached it 2.1% of the time — a roughly 33x gap, reported by search consultant Suganthan Mohanadasan.

Retrieval is not citation, and the gap is wide. In the same analysis, 110 of 3,554 retrieved pages — 3.1% — earned a citation at all. The rate fell from 5.2% at the first-listed domain to 0.3% from the sixth position on. A page can be found, fetched and read, and still be one of the roughly 97 in 100 that go uncited.

The query-writing step changes without notice. Promptwatch, measuring the collapse in Reddit’s ChatGPT citations, attributed the break to an August 8 change in how ChatGPT assembles those background queries: use of the site: operator in fan-out queries rose from 0.37% to 16.8%, roughly forty-six-fold. It reads that as the engine asking named sites directly rather than searching the open web. The company calls the finding provisional and cannot rule out a collection issue on its own end.

Put in sequence, the three findings say one thing. The engine decides which brands and sites to look for; that decision predicts the final answer far better than anything on the pages it then fetches; and the decision rule can change overnight without announcement. Reddit’s citation share on ChatGPT fell 86.4% in four days without Reddit changing anything.

What this does not settle. Both measurements are single-source and both are ChatGPT-only. Google describes a fan-out technique behind AI Mode, but no comparable measurement of its query-writing step has been published, so nothing here transfers to AI Overviews or AI Mode on evidence. Nor does any of it show what moves an engine to name a brand in its own query in the first place — the record has a separate finding that models search for brands they already know 3.2 times more often than unfamiliar ones, which is suggestive and not the same measurement.

The practical consequence is narrower than “optimization does not work,” and sharper. An AEO program that measures only page-level signals is measuring downstream of the step that decides most of the outcome, on a layer no on-page work reaches.