AI search citations: what do cited sources prove?
An AI search citation shows that a URL was attached to a specific API response or search action. It does not prove endorsement, factual correctness, stable visibility or ranking performance.
Read this after a native run when some sources appear beside a query and others remain at run level.
Inspect queries and cited sourcesRun one bounded provider search and keep query, action and run-level sources separate.Inspect a native resultChatGPT citations and Gemini citations need context
A source link displayed in ChatGPT belongs to that displayed answer. An OpenAI API source annotation belongs to its API response, while Gemini grounding data follows Google's response structure and terms. The word “citation” does not make those records interchangeable.
ai-fanout.com analyses only the evidence scope returned by the selected API run. It cannot turn a consumer citation into a hidden query, a traditional ranking position or proof that the source will appear again.
A query records what a provider searched for. A citation identifies a page surfaced in the returned evidence. Finding both in one response does not, by itself, tell you which query led to which URL.
ai-fanout.com creates a query-to-source link only when the provider supplies both inside one single-query action. Sources from a multi-query action remain attached to that action. Gemini source data is shown only at the scope supplied for the current run.
Worked example: one action, four queries and thirteen source domains
In the published OpenAI observation for “Ahrefs vs Semrush for small business”, one search action exposed four query strings and thirteen normalized source domains. The record supports those counts, the exact input, model, locale, timestamp and response status.
It does not support assigning ahrefs.com, sec.gov or any other domain to one particular query because the action did not return that narrower relationship. Keeping the sources at action level preserves evidence that exists without manufacturing evidence that does not.
Together, URL, provider, model, locale and timestamp support the narrow statement that the page appeared in that named run. Recording the tool and method version makes the observation reviewable later.
The record does not reveal why a page was selected, how much of it was read, whether the answer used its claim correctly or whether the page will appear again. Those are separate questions requiring page review or repeated observations.
- Observed: one search action
- Observed: four query strings
- Observed: thirteen source domains
- Not observed: a per-query source ranking
Review the page before using it as evidence
Open the cited page and locate the exact passage that supports your claim. Check the publisher, publication or update date, geographic scope, product version and commercial interest instead of treating the domain name as proof.
Prefer the closest primary source for product behaviour, specifications, prices and policies. Use a secondary source when it contributes reporting, testing or analysis that the primary source cannot provide, and name that role clearly.
The API response contains no traditional organic position, search volume, click potential or guarantee of recurring AI visibility. A cited URL can be useful evidence for one response without being a strong search result or a commercially important page.
Use Search Console and analytics for human discovery and on-site outcomes. If recurring citation visibility matters, define a fixed sample, locale, provider, model, schedule and missing-result rule before comparing observations over time.
- Does the page support the exact claim?
- Is its date and scope suitable?
- Is a closer primary source available?
- What remains unproven?
Provider rules affect storage and publication
Reviewed OpenAI observations may be normalized into public examples under the site's published protocol. The example record carries inputs, versions, status, queries and source domains while making clear that it is neither a consumer-interface capture nor an independent benchmark.
Google applies additional restrictions to Grounded Results and Search Suggestions. ai-fanout.com therefore removes Gemini grounded source data before local saving and does not publish it as a transcript or corpus. The absence of a public Gemini example is an evidence boundary, not missing proof to fill with invented data.
Questions after the first tool run
These questions come from practical tool use. They are not claimed search-volume data.
- Does an AI citation mean the page ranks number one?
- No. A citation records that a URL appeared in the supported scope of one provider response. It contains no traditional organic rank, search volume or guarantee of recurrence.
- Can every cited source be assigned to one query?
- Only when the provider response supplies that relationship. Sources from a multi-query action or a run-level citation list must stay at that broader scope.
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.
- Dated OpenAI fanout example observationsOpen source
Four owner-run OpenAI API observations record exact inputs, timestamps, exposed query strings, search-action source domains, usage, method versions and response status. Versioned JSON, normalized CSV and a JSON Schema are published. They are not an independent benchmark.