Sources retrieved vs sources cited: reading AI citations correctly

A visibility report gives a domain a "citation rate". Before you act on it, you need to know three things: what was counted as a citation, out of what, and for which engine. The same fact can be reported as 0%, 2.6% or 42.3% without anyone making an error, because the numbers count different objects. This post separates the objects, shows the October 2026 numbers for two engines, and ends with a checklist you can run on any report.

What are the three things people call a source?

Most confusion comes from using one word for three layers.

Research drawing the same line has a name for each end. A preprint of April 2026, analyzing 602 controlled prompts across three engines, separates citation selection (which pages the platform chooses) from citation absorption, where a cited page contributes language, evidence, structure or facts to the final answer. Its central descriptive finding is that breadth and depth diverge: Perplexity and Google cite more sources on average, while ChatGPT cites fewer but shows higher average influence among the pages it fetched. A count of links is therefore not a count of influence.

A second preprint, from October 2026, sent its queries to each engine's public API with web search on and replayed the searches of an OpenAI model (gpt-5.4 nano, which the paper labels ChatGPT) and of Gemini. It reports that every page that OpenAI model cited was one its own search retrieved; Perplexity was not replayed. Cited pages are a subset of retrieved pages, at least in that setup.

What do OpenAI and Perplexity actually return?

As of October 2026, their documentation describes two different lists.

OpenAI's web search tool includes inline citations for URLs found in the search by default, and each citation annotation carries a URL, title and location. The API can also return a separate sources list: the complete set of URLs the model consulted, which OpenAI says is often larger than the set of inline citations.

Perplexity's Agent API returns a list of search results with each answer. Whether the answer text carries numbered markers depends on the prompt, according to its documentation, and it says to treat the id and URL of each search result as the source of truth for citations.

So the "sources" an OpenAI integration shows by default are the linked layer, and the list a Perplexity integration holds is mostly the retrieved layer. That is also how Dopbot's October 2026 index records them: OpenAI's linked pages and Perplexity's retrieved pages, kept as separate lists for each engine. It is also why we call Perplexity's lists retrieved sources, not citations.

How different are the two lists in practice?

Very. Our October 2026 sample has 800 answers from OpenAI's API with web search (gpt-6-luna) and 787 from Perplexity's Agent API with the Sonar model (perplexity/sonar), both with web search on:

Perplexity's lists hold about twelve times as many URLs and about eight times as many distinct domains. The second preprint, which queried an OpenAI model through its API with other questions, measured 2.0 distinct linked pages per answer, close to our 2.2 links per answer, although its questions were about home internet service and business software, not shopping. For Perplexity its definition counts only URLs linked in the answer, and it reports that 96% of Perplexity answers had none; it does not say which Perplexity endpoint it called. That was another question set and a stricter definition, so it is not a like-for-like comparison with our 27.3 retrieved URLs, but it shows how far the number moves with the definition.

Why do pooled citation rates mislead?

Because a pooled rate is a weighted average of two very different engines, and the weights are the sample mix. Here is one real fact from our data, reported three ways:

A figure that pools the two columns would fall between them and describe neither engine, and it would move with the mix of the sample. The same distortion hits averages: OpenAI's median answer has 2 links and Perplexity's has 29 retrieved URLs, and a pooled average would describe neither. Perplexity's lists are about twelve times longer, so a ranking that combines the two engines mostly reflects Perplexity's lists. For the Wirecutter case in full, see Wirecutter in AI shopping answers.

How should you read a visibility report?

1. Ask which layer was counted. Linked in the answer, retrieved by the search, or both? Which API field or screen element does the number come from?
2. Ask for each engine on its own, with its own number of answers. A figure that mixes engines is an artifact of the mix.
3. Check the unit and the denominator. Share of answers containing the domain, share of URLs, or a raw count? Out of all answers or only answers with at least one source?
4. Ask how many answers stand behind it, and whether there is an interval. A rate from 10 answers can move a lot on one answer; see how many answers you need before trusting a change.
5. Ask whether it is the API or the app, and which country and prompt wording. Prompt wording can change whether an answer carries source markers; Perplexity's documentation says so for its web search tool. See API answers versus app answers.
6. Check how "own site" is matched. A strict rule (a domain whose name equals the brand's) misses country domains and parent companies. See your own website in AI shopping answers.
7. Do not read presence as influence. If a report says a domain drives recommendations, ask what test shows it.
8. Ask for the raw answers and source lists, so that you can recount with your own rules.

What does this mean for a brand?

Being retrieved looks like a precondition for being linked, at least for the OpenAI model in the second preprint's replay. That makes it worth knowing which pages the search finds for questions in your category, and that is a different question from which pages the answer links. The preprint also found that, for ChatGPT, retrieval by the search explained far more of the differences in citation counts between pages of the same website than page properties did, 32% against 2.8% of the variation. Its questions were about home internet service and business software, not shopping, so this describes citation, not recommendation, and may not carry over to product categories. None of this says that any page change will lead to a recommendation, and no number in a visibility report does either.

For a first look at your own category with both lists, our free manual method records them separately, and the websites inventory shows what that looks like for 160 categories. The same logic applies to forum and video domains: see Reddit and YouTube in AI shopping answers.

Frequently asked questions

Why does Dopbot say retrieved for Perplexity instead of cited?

Perplexity's API returns the pages its search retrieved. Unless the answer text carries markers, it does not say which of them the answer points to, and even markers show a link, not influence. The October 2026 index therefore records the whole list and calls it retrieved sources. OpenAI's linked pages are recorded as links.

Can you get both lists for the same answer?

For OpenAI, the API can return the inline citations and the full sources list together. For Perplexity, you get the search results, and you can see which results the answer text points to only when it carries markers. Our October sample recorded one list per engine, so it has no answer with both.