Query fan-out: how an unbranded question becomes a branded search

The opening prompt and the search string are different objects

OpenAI describes rewriting a user request into one or more search queries. A later search can also depend on earlier results. The exact wording and count vary, and not every answer uses web search.

This matters when the user asks about a category without naming a company. The search string can introduce vendor names or constraints before the final answer is written. Inspect that wording rather than assuming the assistant searched the literal prompt.

What the public study adds

Radyant analysed 3,842 searched runs of 615 unbranded questions in August 2026. Its tracked vendor names appeared in 28.6% of opening queries; 18.1% of all runs named at least two vendors. Dividing those rates gives the conditional 63.3% figure above.

The denominator matters: this is the share among opening queries that named a tracked vendor, not 63.3% of all questions. The tracked vocabulary and mostly German sample limit generalisation. A brand appearing in the search string does not establish that the final recommendation was predetermined.

Read a fan-out as a set of added requirements

Suppose the user asks for a CRM for five people with Gmail support. A captured search might ask about an integration limit or compare named vendors. Those changes point to evidence the assistant is seeking, not automatically to a new article for every wording variant.

Illustrative prompt-to-query changes, not captured study records

What the query adds	What to investigate
Named competing products	The relevant comparison and each product's documented fit.
A price or team-size restriction	A current plan table with limits attached to the price.
An integration requirement	The exact supported behaviour and any setup or plan restrictions.
A review or reliability term	Independent evidence that actually addresses that requirement.

Capture queries before analysing them

Use the DevTools extraction tutorial for the response fields and troubleshooting. Preserve the conversation turn, model, date and available context. Where query metadata is absent, record it as unavailable instead of asking the model to generate a plausible replacement and labelling it observed.

Capture repeated runs and deduplicate exact strings for vocabulary analysis. Do not count duplicate metadata as new search executions, and do not assign all conversation-level queries to the most recent user message.

Compare the captured queries with the actual cited pages

Keep the query strings and cited URLs as separate records. A page's appearance in a search result is not the same as a citation, and a citation is not necessarily a product recommendation. Inspect the final answer before deciding which gap the content should address.

Update an existing page when it can answer the missing requirement accurately. If the answer relies on an independent comparison that omits a relevant product, check its criteria and offer verifiable evidence for editorial review. Do not manufacture a review or an unsupported product claim to match the query.

Use fan-outs as evidence, not a hidden ranking formula

A query log helps reveal the wording exposed by the tool. It does not reveal every server-side retrieval step or prove which source caused the final recommendation. A fresh Google search for the same string is another observation, not a reconstruction of the assistant's historical result set.

For category research, Aiso's conversation and fan-out analysis keeps the task grounded in observed questions. Check the available model, market and date coverage in the methodology before treating a sample as representative demand.