How can you compare OpenAI and Gemini fanout?
Run the same topic once with each provider, keep locale fixed, save both results and compare them locally. The browser shows exact query overlaps and each run side by side without another API request.
Use this before comparing providers or the same provider on different dates.
Keep the comparison setup fixed
Use the same topic, language and country, run observations close together and record each exact model. The tool sends one provider request at a time so cost and failures remain visible.
Select at least two saved runs and choose Compare. The browser calculates exact query matches and shared source domains without a provider request. History stays on this device for up to 30 days.
Compare exposed query strings, user jobs and sources at the scope supplied. Exact overlap is objective but narrow; semantic grouping finds paraphrases but introduces judgment.
Google restricts reuse of Grounded Results and Search Suggestions. The site removes that data before local saving and does not publish a retained Gemini transcript.
A benchmark needs a declared sample, repetitions, missing-result rules, stable versions, cost reporting and a method chosen before results. This view compares selected observations only.
- Query strings
- User jobs
- Sources at supported scope
Zero queries can be an honest result
The provider may decide the topic needs no web search, or may search without exposing a reusable query string. The native result should remain zero; filling it with likely searches would turn observation into guesswork.
A completed zero-query run may still include provider status or usage. Timeout, CAPTCHA failure, quota stop, invalid response and provider errors should appear explicitly, not as empty success.
Use one topic up to 60 characters. Avoid URLs, multiple questions or instructions about how many searches to perform. A narrow fact may legitimately require only one search or none.
For brainstorming, switch to Search Ideas. It generates research directions without web search. Keep that label: modelled ideas do not reveal what the native provider searched.
Record provider, model, time, locale, visible status and public error code. Never send an API key or private raw response. Those details help distinguish normal zero behavior from a parser regression.
Sources used
These sources support the functions and limits described here. They do not prove claims beyond that scope.
- OpenAI web search tool guideOpen source
Documents web_search_call output and search actions that can include the query or queries searched. It does not expose chain of thought or guarantee that every route returns query strings.
- Gemini grounding with Google SearchOpen source
Documents Gemini 3.7 Flash Google Search, google_search_call arguments.queries, cited URL annotations and per-query billing.
- Gemini API Additional TermsOpen source
Restricts caching, syndication, analysis and reuse of Google Grounded Results and Search Suggestions; drives the Gemini local-save redaction and public-example boundary.
- Direct OpenAI and Gemini native-fanout implementationOwner record
One direct provider request, strict provider allowlist, provider-exposed query/source parsing, honest zero-query output, CAPTCHA-first shared atomic quota and budget accounting, no raw server archive and separate public flag. Bounded live production fixtures passed for OpenAI and Gemini on 2026-08-26.
- Opt-in browser-local result historyOwner record
Explicit save action, IndexedDB-only storage, maximum 20 entries, 30-day expiry, individual deletion, clear-all and local JSON/CSV export without accounts or shared response caching.