A real API search run, not a browser capture.
Native Fanout observes search actions exposed by a provider API under one fixed, published protocol. It does not inspect chatgpt.com, gemini.google.com, claude.ai, browser network traffic, system prompts, private retrieval traces or chain of thought.
- 01
Input
One topic or short question, 2–60 characters and no more than 160 UTF-8 bytes. URLs, files, line breaks and extra request fields are rejected.
- 02
Human check
Cloudflare Turnstile must pass before quota or budget is reserved.
- 03
Native provider
Choose OpenAI GPT-6 Luna, Gemini 3.8 Flash or Claude Haiku 5.5. All three use native search through OpenRouter, restricted to the selected model provider. OpenAI and Haiku use Responses; Gemini uses Chat Completions. No provider key reaches the browser.
- 04
Real web-search run
The request uses openrouter:web_search with engine native. Native providers may ignore the requested search-context size. Routing metadata must confirm the native engine and selected provider; external-search fallback is rejected. max_tool_calls is eight. The fixed public instruction asks the model to search naturally and answer briefly; max_output_tokens is 500 for Responses; Gemini Chat uses max_tokens 500. Provider-reported usage can include tool overhead beyond the configured text limit.
- 05
Observed output only
The parser returns only query strings and cited source URLs present in the provider response. OpenAI and Haiku sources are linked to a query only when the same exposed search-action object supports that link. Gemini citations remain run-level. The checked Gemini route exposes no query strings; its search counter comes from reported search-request usage, without fabricated actions. If the API exposes no query string, the result says so instead of inventing one.
- 06
Plain response summary
The result repeats only provider-grounded counts: visible query strings, distinct source domains and search actions. It adds no automatic topic, price or comparison labels.
- 07
Usage and cost
The response includes provider-reported input and output tokens plus search-action counts. OpenAI input also includes web-result content passed to the model, so the number is much larger than the keyword and instruction. The request asks for low context, which OpenRouter documents as ignored by native search. The UI prefers OpenRouter-reported cost and otherwise labels a dated list-price estimate; provider billing remains authoritative.
- 08
No raw server archive
The server keeps the topic and provider response transient. A visitor can explicitly save up to 20 results for 30 days. Gemini saves remove Google Grounded Results and Search Suggestions before local storage.
- 09
Hard limits
Twenty attempts per hashed IP bucket in a rolling 24-hour window and 40 site-wide per UTC day. The tool shows remaining runs and the reset time. Native requests reserve EUR 0.10 inside the shared EUR 25 soft and EUR 30 hard monthly limits.
- 10
Separate ideas mode
Search Ideas offers GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5 through OpenRouter to generate ten modelled directions without web search. All three models are also available in Native Fanout, where the provider runs real web searches. It is not called native or controlled fanout.
What “native” proves—and where the claim stops
It proves that a specific API run exposed those query strings and citations at that time. It does not prove that a consumer chat product would produce the same fanout, that the source is correct, or that the query has measurable demand.
That limit is deliberate. A run with zero visible queries remains a valid observation when the provider completed without exposing a query field. A timeout, quota response or malformed provider output is shown as an error instead; missing evidence is never replaced with a plausible guess.
Which conclusion does each field support?
| Field or status | Supported claim | Limit |
|---|---|---|
| Query string from Native Fanout | A search phrase exposed in this API run | No search volume, ranking or complete consumer trace |
| Query/action/run source | A source at the scope supported by the response | No per-query attribution without a corresponding relationship |
| Completed, zero visible queries | The run exposed no query strings | Missing fields are not filled with search ideas |
| Search Ideas output | Ten modelled research directions without web search | No observed provider searches or established demand |
| Usage and cost estimate | Reported usage evaluated with dated list prices | Not the provider's binding invoice |
| Optional local save | Up to 20 runs, at most 30 days, exact browser origin | No account sync; Gemini grounding and search suggestions removed |
Illustrative case: a completed native run without a query field and a failed request are different states. A request error is not a zero-query observation. Neither output automatically justifies a new page.
Method/implementation review: 10 October 2026. Dated OpenAI examples retain their original observation dates.
For a page decision, you can now cluster queries and check the question against Search Console. Open the clustering worksheet.
Turn the observation into one defensible SEO decision
Group observed queries by reader job, then compare those jobs with the strongest page already on the site. Add a section when that page can answer the job well. Create a new URL only when audience, intent, evidence and next step are genuinely distinct.
A practical review ends with a small decision set: improve an existing section, research a claim, create one justified page or do nothing. Search Console, analytics and business relevance decide priority; the number of fanout branches does not.
Who is accountable for the method
Matthias Ramahi owns protocol versions, corrections, reviewer status and the monthly cost envelope. This tool is not an independent provider benchmark, and shared ownership of related SEO products is disclosed rather than presented as outside validation.
When a provider contract, parser or price changes, the method and dated source register must change with it. That makes the public boundary reviewable instead of asking visitors to trust an unexplained result card.