- "Citation" isn't one universal mechanism. Both OpenAI's and Anthropic's own API documentation show a real technical gap between sources a model fetches during retrieval and the smaller subset that actually gets a visible citation in the final answer.
- What counts as a citation differs by product: Perplexity cites by default on nearly every answer; ChatGPT only cites when web search is actively invoked; Claude's API attaches citations only when its web search tool runs; Google AI Overviews shows a source panel that functions more like general attribution than Perplexity's sentence-level citations.
- A Semrush study of nearly 4,000 domain appearances across four engines found roughly 62% were "ghost citations": used as a linked source, never named in the visible text. Being a source and being credited are two different outcomes.
- Citations that do appear aren't reliably accurate. Columbia Journalism Review's Tow Center found AI search engines misidentified a direct quote's actual source more than 60% of the time across 1,600 tests, with results varying sharply by platform (Perplexity 37% wrong, Grok 3 94% wrong).
- None of the major labs have published exactly how their citation-selection algorithm decides which fetched source gets surfaced. The "how" remains genuinely unconfirmed, and this piece is explicit about that gap rather than repeating unsourced pipeline descriptions.
AI visibility strategy already draws one distinction that matters here: citation, mention, and recommendation are three separate outcomes, and a tactic that improves one doesn't automatically improve the others. This piece goes one level deeper into the first of those: what "citation" actually is, mechanically, and why being read by a model is not the same thing as being credited by it.
What's the real technical split: fetched versus cited?
Both OpenAI and Anthropic document this distinction directly in their own API schemas, which is stronger evidence than an inferred pattern: it's the labs' own data structure making the split explicit.
OpenAI's Responses API, when its web-search tool is used, returns two different fields: a set of url_citation annotations (a curated subset described as "the most relevant references") and a separate sources field listing the complete set of URLs the model actually consulted, which is documented as often larger than the citation set. A page can be in the second list and never make it into the first.
Anthropic's API shows the identical pattern from a different angle: every page fetched during a web search returns as a web_search_result block, but only the specific pages the final answer text actually references get a citations entry (web_search_result_location) attached to a span of that text. A fetched page with no citation entry pointing to it was read and not credited, visible in the API's own JSON, not inferred from behavior.
Does citation mean something different depending on the product?
Perplexity cites by default on nearly every response. Live retrieval is core to what the product does, so a source list is close to a permanent feature of its answers.
ChatGPT only shows citations when its web-search capability is actively invoked for a given answer. A response generated purely from the model's training data, with no live search triggered, carries no citations at all, because nothing was fetched to cite.
Claude, per Anthropic's own documentation, follows the same pattern as ChatGPT: citations appear automatically whenever the web search tool runs, and don't appear when Claude answers from what it already knows without searching.
Google AI Overviews shows a source panel whenever an Overview appears, but the attribution is closer to "these pages informed this answer" than to a guaranteed sentence-level citation the way Perplexity's inline references work. An Overview citing a source list doesn't necessarily mean every claim in the generated text traces to one specific page in that list.
The practical implication: a monitoring setup built around "did we get cited" needs to account for the fact that the underlying mechanism, and the odds of a citation even being technically possible for a given query, differs by which product generated the answer, not just by how well-optimized the source content is.
What's the academic version of this split?
An independent 2026 preprint frames this fetched-versus-cited gap as a two-stage process worth naming precisely: citation selection (does the platform decide to search, and which sources does it pick to surface) and citation absorption (how much a cited page's actual content shapes what the final answer says). Running 602 controlled prompts across ChatGPT, Google, and Perplexity (21,143 valid citations, 18,151 fetched pages), the researchers found meaningfully different patterns per platform: ChatGPT surfaced fewer citations per prompt on average than Google or Perplexity, but leaned far more heavily on each one it did cite. Google and Perplexity cited more sources per answer, but each individual source carried comparatively less measured influence on the final text.
This is a preprint by independent researchers, not a peer-reviewed institutional study, so it's worth treating as a serious, methodologically-disclosed data point rather than settled fact. Still, it gives a genuinely useful vocabulary for a distinction that's otherwise easy to gesture at vaguely.
Is being cited the same as being credited?
A Semrush study (115 prompts across 14 countries, run on ChatGPT, Google AI Overviews, Gemini, and Google AI Mode, yielding 3,981 domain appearances) found that roughly 62% of those appearances were what the study calls "ghost citations": the AI used the page as a linked source, but never actually named the brand anywhere in the visible answer text. A source can be doing real work behind an answer and still be functionally invisible to the person reading it, and to any brand tracking mentions by scanning visible text rather than checking the underlying source list.
Is being cited the same as being cited correctly?
A citation appearing at all is a weaker guarantee than it looks like. Columbia Journalism Review's Tow Center for Digital Journalism ran 1,600 test queries across eight AI search engines (ChatGPT Search, Perplexity (standard and Pro), Gemini, DeepSeek Search, and Grok 2/3 among them), giving each engine a direct quote from a real news article and asking it to identify the correct source. The overall failure rate was above 60%. Perplexity performed best at 37% wrong; Grok 3 performed worst at 94% wrong. A visible, confidently-presented citation carries no guarantee the model actually traced the claim back to the right place.
What's still genuinely unconfirmed?
No major lab has published an official technical account of exactly how its citation-selection algorithm decides which fetched source, out of everything retrieved, actually earns a visible citation. Pipeline descriptions that circulate widely (a specific candidate-source count narrowed down to a specific cited count, for instance) are reverse-engineered guesses from SEO analysis, not confirmed specifications from Google, OpenAI, Anthropic, or Perplexity. The technical split between fetching and citing is real and documented in the labs' own API schemas; the exact selection logic underneath it isn't public anywhere, and any post that states it with confidence is speculating past what's actually known.
Frequently asked questions
What's the technical difference between a model "reading" a source and "citing" it?
Both OpenAI's and Anthropic's own API documentation distinguish the two at the schema level. A model can fetch and use a page during retrieval without it ever appearing as a visible citation in the final answer. OpenAI's API returns a full list of consulted sources separately from the smaller, curated set that gets an actual citation annotation, and Anthropic's API attaches citations only to the specific pages the final text actually references, distinct from every page it fetched.
Does "citation" mean the same thing on ChatGPT, Perplexity, and Google?
No. Perplexity cites by default on essentially every answer, since live retrieval is core to the product. ChatGPT only shows citations when its web-search mode is actively invoked: an answer generated purely from training data carries none. Google AI Overviews shows a source panel whenever an Overview appears, but its citation UI functions more like "these pages informed this" than a sentence-level attribution the way Perplexity's does.
How often does a source get used without ever being credited?
More often than most brands assume. A Semrush study of 3,981 domain appearances across four AI engines found roughly 62% were "ghost citations": the AI used the page as a source, but never named the brand in the visible answer text. Being fetched is not the same as being credited, and most of the time a source gets used, it doesn't get named.
Are AI citations actually accurate?
Frequently not, per one of the more rigorous studies available. Columbia Journalism Review's Tow Center tested 1,600 queries across eight AI search engines, giving each a direct quote and asking it to identify the source. The overall failure rate was above 60%. Perplexity had the lowest failure rate at 37%; Grok 3 was wrong 94% of the time. A visible citation isn't a guarantee the model actually identified the right source.
Sources referenced
- Anthropic: Web search tool documentation
- Yang, "News Source Citing Patterns in AI Search Systems" (arXiv 2507.05301, Jul 2025)
- Columbia Journalism Review / Tow Center: AI search engines and news citation accuracy (Mar 2025, 1,600 queries)
- Nieman Lab coverage of the Tow Center study
This piece extends the citation/mention/recommendation distinction made in AI visibility strategy. For what actually makes a source more likely to be selected in the first place, see how to write content that AI models can actually quote. For the technical pipeline behind "citation" in the first place, see what is a generative engine, technically? For how this plays out on a model-by-model basis, see how Claude cites sources and, for the newest exposure point, AI browsers and citation.
