Retrieved sources: definition, and why citations are only a subset

A retrieved source is a URL that an AI engine pulled into the model's context while answering a query. The retrieved set is everything the model consulted. The cited set is the part it chose to credit in the answer, as an AI citation or an inline citation.

In the four-stage model of AI visibility, retrieval is stage 2 (does the page enter the context for this query) and citation is stage 3 (which retrieved pages the model uses and credits). Retrieval has to succeed before citation can happen. For the full comparison, read the guide on cited vs retrieved sources.

When an engine decides to search, it rarely sends your exact prompt. OpenAI's help documentation says ChatGPT search typically rewrites the query into one or more targeted queries. The results of those queries form the retrieved set. The model then writes an answer and attaches citations to some of those results.

Three APIs expose the two sets separately, as of October 2026:

• OpenAI Responses API: a sources field lists every URL the model consulted, separate from the citations. OpenAI notes the sources list is often longer.
• Anthropic web search tool: search results return a URL, title and page_age. Citations point to text spans, with cited_text up to 150 characters.
• Perplexity Agent API: search_results lists each result with an id, URL, title and snippet. Answers cite whole sources by integer index. Web search is opt-in on this API.

Because the units differ (spans versus whole sources), retrieved and cited counts from different engines don't add up to one comparable number.

If you count only citations, you see the end of the process and miss the middle. A brand page can be retrieved for a query, read by the model, and left uncredited. That page still passed the access and retrieval gates, which tells you crawling and indexing work for that engine.

Citations are also a noisy signal of influence. A mechanistic study of attribution (arXiv:2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources. Citation output can be separated from what the model actually drew on.

The next move: log both sets for the queries you care about. The playbook on logging retrieved vs cited sources shows how to do this through the APIs. Treat the results as a proxy, because API behavior is not identical to the consumer apps.

• "If I'm not cited, the AI didn't read me." Wrong. Retrieval and citation are separate steps. See whether AI can use your content without citing you.
• "A citation means my page shaped the answer." Not necessarily. Citation emission is partly decoupled from source use.
• "I can see the retrieved set in ChatGPT." Not for a real user's query. No consumer surface exposes it as of October 2026. Only the API shows it.
• "Rewriting a retrieved page will turn it into a citation." The evidence doesn't support this. A 2026 critical survey of GEO studies (arXiv:2607.14035) warns that citation-oriented rewrites can impair retrieval.
• "Being retrieved guarantees anything." It doesn't. Retrieval is per query and at the engine's discretion.

• AI citation: a source the answer credits.
• Inline citation: a citation placed next to the claim it supports.
• Citation rate: how often a source is cited across a set of answers.
• Seeing what ChatGPT retrieved but didn't cite: the practical limits of looking inside one engine.

Is retrieved sources the same as sources consulted? Yes, in practice. OpenAI describes its `sources` field as the complete list of URLs the model consulted. Other vendors use different names: Anthropic returns search results and Perplexity returns `search_results`. All three describe the set of pages that entered the model's context, before the model chose which ones to cite.

Why would an engine retrieve a page and not cite it? Nobody outside the engine can say for a given answer. Retrieval pulls in more pages than the answer needs, and the model picks which to credit. Research on already-retrieved content finds topical relevance and position in the context list are the most reproducible drivers of citation. Publishers do not control context position.

Does a bigger retrieved set mean better AI visibility? It means your pages reached the model more often for the queries you tested. That is useful evidence that access and indexing work for that engine. It does not promise citations, traffic or accurate answers. Measure each engine separately, because their retrieval systems and citation formats differ.

Can I measure retrieved sources without an API? Not reliably. Consumer apps show citations, not the full retrieval set, and they may not search at all for a given prompt. Server logs can show that an answer-time fetcher such as ChatGPT-User requested a page, which hints at retrieval. API logging gives a cleaner, repeatable record, with the caveat that it approximates the apps.