How ChatGPT searches for software recommendations: a query fan-out study

Query fan-out is the set of web searches an answer engine runs to answer one prompt. On September 12, 2026, StatWharf sent the same 100 software category prompts to ChatGPT, Gemini and Google AI Overview, and recorded what each engine exposed. This page reports the method and the counts.

Method

The panel used 50 software categories that StatWharf covers. Each category had two prompts. The baseline prompt was best {category}, for example best accounting software. The year prompt added the year, for example best inventory management software 2026. That gives 100 prompts per engine.

StatWharf captured the network traffic of each answer as it was generated. For ChatGPT, the searches appear in the live response stream. The saved conversation record returns the search call with blank arguments, so tools that read only the saved record see no queries.

Only ChatGPT exposed its searches

Google runs its fan-out on the server. A full AI Mode page load fired 94 requests, and every request carried only one distinct q= value, the prompt itself. The 1.5 MB of captured response bodies held no sub-queries. Gemini behaved the same way. Its stream carried a status frame labeled “Searching the web” with no query text.

Two strings in the Gemini stream look like fan-out but are not. The <Elicitation> and <FollowUp> tags carry a query attribute that holds suggested next prompts for the user. The shopping carousel embeds google.com/search?q= advertisement links. The study excludes both.

Some third-party tools advertise Google fan-out data. In this study, Google did not send those sub-queries to the browser. Such tools therefore predict the sub-queries with their own model. No predicted query appears in this dataset.

What ChatGPT searched

ChatGPT never searched the prompt exactly as typed. The median prompt was 5 words long and the median fan-out query was 11 words long. The median prompt produced 2 searches, and the largest number for one prompt was 8.

The expansion followed one pattern. ChatGPT kept the category phrase and added vendor names and a year. For best board portal software, it searched best board portal software Diligent Boards OnBoard BoardEffect Nasdaq Boardvantage iBabs features 2026.

The baseline prompts contained no year. They produced 102 queries, and 63 of those queries (62%) added a year anyway.

Citations

ChatGPT cited StatWharf in 32 of 100 runs: 17 baseline runs and 15 year runs. In 30 of those 32 runs, the cited page was the StatWharf page for the category in the prompt.

The runner captured 2,863 ChatGPT source rows on 720 hosts. The five most frequent hosts were:

These counts depend on the panel. ChatGPT cited 20 distinct StatWharf pages, and 8 of them are SAP governance pages. SAP-focused hosts rank high for the same reason.

Gemini cited StatWharf in 0 of 100 answers. Google AI Overview cited StatWharf in 0 of 100 answers. StatWharf appeared in the ordinary organic results on 8 of the 100 Google pages, and seven of those were SAP categories.

What this means for software research pages

ChatGPT kept the category phrase in 68% of its searches. A page whose address and heading use that phrase stays reachable for those searches. ChatGPT also added vendor names and the current year that the user did not type. A page that names the vendors in the category and states the year matches more of the search text than the phrase alone.

Questions about the method can go to editorial@statwharf.com.

Factual corrections are free: editorial@statwharf.com.