Gemini rewrites your question into seven searches. ChatGPT runs it almost verbatim. Same query, two different machines.
We mapped the hidden sub-questions behind 20 buyer queries, on both engines. Gemini explodes each buyer question into about seven improvised searches that change every time. ChatGPT searches your question close to word for word and ranks the results. If your GEO strategy treats "AI" as one thing, you are optimizing for the wrong machine.
Original study · measured 2026-08-10
Every guide about ranking in AI answers tells you to optimize for the question your buyer asks. There is a problem with that advice, and a second, larger one. First, one engine does not search the question your buyer asks, it invents its own. Second, the two big engines do this so differently that a single "AI SEO" strategy cannot serve both.
We ran fanout_map against 20 "best X" buyer queries, three times each, on Gemini and on ChatGPT, and recorded the sub-questions each engine actually searched. The two pictures barely resemble each other.
Gemini turns one buyer question into about seven
On Gemini, each query expanded into a mean of 6.8 sub-questions, from 4 to 11. It does not search "best web scraping api." It searches these instead:
- best web scraping api 2026
- top rated web scraping apis for developers
- features to look for in a web scraping api
- criteria for choosing a web scraping api
- what makes a web scraping api the best
Gemini decomposes the question: it reformulates your query into the angles it thinks a buyer cares about, then searches all of them and assembles the answer from what comes back.
On Gemini, those sub-questions are unstable
Ask three times and they barely repeat. The mean run-to-run overlap was 0.25 (where 1.0 would be identical every time), and for 10 of the 20 queries not one sub-question appeared in all three runs. "best password manager" was the extreme: the sub-questions Gemini expanded into on the three runs had zero in common. A single run also surfaced only 47% of the full set. On Gemini, you are aiming at a moving target, and one look shows you half of it.
ChatGPT does the opposite: it runs your question almost verbatim
To be precise about what "ChatGPT" means here: we measured OpenAI's gpt-4o-mini with web search forced on, not the ChatGPT consumer app, which runs a larger reasoning model. With that noted, it does not fan out. Across the same 20 queries, three runs each, it searched the buyer's exact query, once, and returned the identical search on all three runs for all 20 seeds. No reformulation, no year appended, no angle. It searches "best crm for startups" and ranks whatever that one search returns.
A heavier, reasoning configuration of ChatGPT does a little more, about 4 searches across three runs, but they are a different animal: follow-ups that drill into specific vendors it already found ("Zoho CRM startups review," "Freshsales CRM best for startups"), plus reading two or three pages in full. That is not decomposing the question. Where Gemini explores the question (many ways to ask it), ChatGPT commits to your question and explores the answer (verify the candidates, read the pages). Two opposite strategies for the same buyer query.
The two engines, side by side
| buyer query | Gemini sub-queries | ChatGPT | what ChatGPT searched |
|---|---|---|---|
| notion alternative | 11 | 1 | "notion alternative" |
| best ai coding assistant | 9 | 1 | "best ai coding assistant" |
| best authentication service | 9 | 1 | "best authentication service" |
| best web scraping api | 8 | 1 | "best web scraping api" |
| best vpn | 7 | 1 | "best vpn" |
| best crm for startups | 4 | 1 | "best crm for startups" |
| all 20 (mean) | 6.8 | 1.0 | the query, verbatim |
What Gemini reformulates toward
Because Gemini is the engine that decomposes, its sub-questions tell you what content it hunts for. Across the 135 distinct sub-questions it fired, they cluster:
| reformulation | share of Gemini's sub-questions |
|---|---|
| a "top" or "best" listicle | 59% |
| "for" a use case or audience | 39% |
| stamped with a year (2025, 2026) | 30% |
| a "how to" or tutorial | 19% |
| an "X vs Y" comparison | 7% |
Why this splits your strategy in two
The instability on Gemini connects to the citation study: because Gemini's sub-questions change between runs, the pages it retrieves and cites change too. Citation instability is downstream of fan-out instability. ChatGPT, running one stable query, is correspondingly steadier about who it cites.
That gives you two different jobs, not one:
- For Gemini, cover the sub-questions. It will search "web scraping api for developers" and "criteria for choosing a web scraping api" whether or not you have a page for them. Year-stamped, use-case-segmented, comparison content is what its fan-out is reaching for, and being cited means appearing across many of those sub-query results, not ranking for the head term.
- For ChatGPT, win the head query itself. It runs "best web scraping api" once and ranks the results. There is no fan-out to catch you, so being in the result set of that single, literal query, the classic head term, is the whole game.
Treating "AI visibility" as one target is the mistake. The engine that decomposes and the engine that does not reward different pages. fanout_map(seeds=["best X for Y"], engine="gemini") shows you the sub-questions to cover on Gemini; run it with engine="chatgpt" and you will mostly see your own query handed back, which is the finding.
Method
20 "best X" and category buyer queries, each run through fanout_map three times per engine, on 2026-08-10. fanout_map surfaces the sub-questions an engine expands a search into. For ChatGPT the primary run queried OpenAI's gpt-4o-mini with web search forced on, a fast, tool-safe configuration that is not the same as the ChatGPT consumer app (which runs a larger reasoning model); a heavier reasoning model was sampled on 10 of the seeds and is reported separately (it fired a mean of 4.2 vendor-drilldown searches across three runs and opened about 2.7 pages per run). The finding is about how these engines decompose a query, and holds across both configurations we ran, but do not read it as the exact behavior of the consumer app. Gemini figures: "sub-questions" is the union of distinct sub-questions across three runs; "run-to-run overlap" is the mean pairwise Jaccard; a single-run coverage of 47% is the mean fraction of the union seen in one run. Gemini's expansion is stochastic, which is itself a finding; ChatGPT's single query was stable across all runs. Sub-questions are lowercased and deduplicated. No scoring, no verdict.
A point-in-time measurement. Search results and AI answers change, and grounded models vary between runs, so your own numbers will differ. Verithia measures. The interpretation is yours.