Claude vs Gemini citations

What this page compares

This is a comparison of citation mechanics, not of citation frequency. Both Anthropic and Google document what a grounded answer hands back when a web source is used. Neither documents how a source is chosen. That boundary is where most writing on this subject quietly crosses from reporting to invention, and this page stops at it.

The practical reason to care: if your brand is mentioned inside a product built on one of these APIs, whether a reader can see and click your link is decided by the citation contract between the vendor and the developer. That contract is published on both sides, and the two versions are meaningfully different.

Claude: citations as a property of the tool

Anthropic's web search documentation says citations are always enabled for web search.[^claude-citations-always] There is no flag to turn them on. A developer using the tool receives them whether or not they planned for them.

Each citation carries four things: the URL of the cited source, its title, an encrypted index used to continue multi-turn conversations, and up to 150 characters of the cited content.[^claude-citation-fields] That last field is the interesting one. It is not a link to your page; it is a fragment of your page, returned to the developer, marking the sentence that supported the statement.

Anthropic also removes a small piece of friction that would otherwise push developers to strip citations: it states that the cited_text, title and url fields "do not count toward input or output token usage."[^claude-citation-tokens] Citing you is free at the token layer.

On display, the documentation says that when API outputs are shown directly to end users, "citations must be included to the original source."[^claude-display-guidance] It is a requirement placed on the developer, not a rendering the vendor performs.

The search results themselves carry a page_age field described as when the site was last updated, along with url, title and encrypted_content.[^claude-result-fields] A freshness signal being present in the result payload is worth noting, though nothing published says what the model does with it.

Two further controls sit on the request rather than the response. A caller can pass allowed_domains or blocked_domains — not both, which returns a 400 error[^claude-domain-filtering] — meaning a developer can decide in advance which slice of the web their product is allowed to cite. And Anthropic documents when Claude searches at all: on requests that depend on information that is current, changing, or outside training data, rather than on stable knowledge.[^claude-when-searches]

Gemini: grounding as an inspectable search step

Google frames grounding with Google Search as connecting the model to real-time web content so it can "provide more accurate answers and cite verifiable sources beyond its knowledge cutoff."[^gemini-grounding-purpose]

What comes back is shaped less like a footnote and more like a search report. The response contains the search queries the model executed.[^gemini-queries-returned] It contains inline url_citation annotations on the text content block, with start and end indices marking which span of generated text came from which source URL.[^gemini-url-citations] And it contains search_suggestions, "an HTML snippet for rendering search suggestions in your UI."[^gemini-search-suggestions]

Google's own framing of the annotations is that they give the developer "complete control over how you display sources in your user interface."[^gemini-display-control] Control, rather than obligation, is the posture — with the qualification that Search suggestions come with their own usage requirements, which Google's documentation does not restate but points at its terms of service for.[^gemini-usage-requirements]

That is an honest thing to record and an awkward one to build on: the requirement exists, and the developer-facing page does not tell you what it is.

The publisher-side control is the clearest part of the Google picture. Google-Extended is documented as the token publishers use to manage whether crawled content "may be used for training future generations of Gemini models that power Gemini Apps and Vertex AI API for Gemini and for grounding."[^gemini-training-control] Grounding is named explicitly. A publisher who blocks Google-Extended is making a decision about this exact feature.

Reading the two side by side

The table above lines the two up row by row. Three of its cells say "not documented", and those are as informative as the rest.

The most substantive difference is what a citation contains. Anthropic returns a piece of the source. Google returns a pointer into the answer. If you are building a product where a reader should be able to see why a claim was made, Anthropic's shape gives you the material without a second fetch. If you are building a product where the reader should be able to continue searching, Google's shape — queries plus a suggestions widget — is closer to what you need.

The second difference is who is responsible for the interface. Anthropic states a requirement and supplies no widget. Google supplies a widget and refers the requirement to its terms. A developer in a hurry will render Google's snippet and satisfy something without reading it; a developer in a hurry on Claude has to write the citation UI themselves and will therefore have thought about it. Neither approach is obviously better, and both have a predictable failure mode.

The third is cost transparency. Anthropic publishes a figure — ten US dollars per one thousand searches, as of 3 September 2026, plus standard token costs[^claude-search-pricing] — and says the citation fields are free at the token layer. Google's grounding page publishes no equivalent figure, so the cost of a grounded call has to be looked up elsewhere.

What neither of them tells you

Neither document explains which retrieved source becomes a citation.[^no-selection-logic-published] Both describe retrieval and both describe the citation payload. The step in between — the selection — is unpublished on both sides.

This matters because it is precisely the step that a brand would want to influence, and it is the step about which the most confident advice circulates. Any claim that a particular structure, schema or phrasing increases citation rate on Claude or Gemini specifically is an inference from sampling, not a reading of the documentation. Inferences from sampling can still be useful, but they should be labelled as what they are, along with the sample size, the date and the prompt set.

The behaviour is also non-deterministic. Anthropic's documentation notes that simple factual queries typically use a small number of searches while comparative or multi-entity research can use many more, which means the same question asked twice can retrieve differently. A single observation of a citation is an anecdote; a repeated, dated sample is a measurement.

Practical implications for a site owner

1. Decide the Google-Extended question deliberately. It is the one

documented publisher control on this page that names grounding explicitly. Blocking it is a real decision about Gemini grounding, not a generic AI-training gesture.

1. Expect to be quoted, not just linked, on the Claude side. Up to 150

characters of your page can be returned to a developer as the supporting excerpt. Pages whose key claims are stated in complete, standalone sentences are easier to excerpt correctly than pages whose meaning depends on the paragraph above.

1. Do not model these two as one channel. The developer obligations, the

payload shapes and the publisher controls all differ. A dashboard that merges them without saying which is which is producing a number nobody can check.

1. Treat any citation-rate claim about either platform as a sample. Ask for

the date, the prompt set and the sampling depth before acting on it.

A note on multi-turn behaviour

One detail that looks like plumbing has a real consequence for how citations survive a conversation. Anthropic documents an encryptedcontent field on every search result and an encryptedindex on every citation, and states that a continuing conversation must send the assistant's content blocks back exactly as received, including that encrypted content, or the request fails with a validation error.

The practical effect is that a developer cannot quietly strip the citation scaffolding and keep the conversation going. The citation payload is load-bearing rather than decorative. Whether that was designed as an integrity measure or falls out of how context is restored, the outcome for a publisher is the same: on the Claude side, the link to your page is harder to drop by accident than it would be if citations were an optional garnish.

Google's grounded response is structured around inspectable steps instead, with the executed queries and the annotations returned alongside the generated text. That gives a developer more freedom in how sources are presented, which Google frames as a feature. Freedom cuts both ways for a publisher: more control over presentation also means more room for a product to present sources minimally, and nothing in the documentation prevents that.

There is a second-order effect worth noticing. Because Anthropic returns an excerpt of your page rather than only a pointer to it, the sentence that gets shown is chosen by the model from your text. A page whose central claims are written as complete, self-contained sentences gives the model something usable to excerpt. A page whose meaning is spread across three paragraphs of build-up gives it a fragment that reads as a non sequitur. That is not a ranking trick; it is a consequence of the payload shape, and it is one of the few places where the published mechanics have an obvious editorial implication.

Related comparisons

The retrieval layer, rather than the citation layer, is where most visibility problems actually live. For how two operators gate access to your pages in the first place, see ChatGPT search versus Perplexity; for the difference between an AI answer that inherits eligibility from search and one that gates it behind its own crawler, see Google AI Overviews versus ChatGPT search; and for how Anthropic and OpenAI each split crawling across multiple agents, see ChatGPT versus Claude for brand visibility.

The limit of this page

Everything above is taken from Anthropic's web search tool documentation and Google's grounding documentation, both fetched on 3 September 2026, plus Google's crawler documentation for the Google-Extended control. Prices, field names and tool versions on both sides change without notice, which is why each claim carries the date it was read rather than being written as a permanent fact.

What this page will not give you is a ranking. Saying that one vendor cites more generously than the other would require data neither vendor publishes, and a comparison that manufactures that answer is worth less than one that says where the documentation stops.

[^claude-citations-always]: Anthropic web search tool documentation, retrieved 2026-09-03. [^gemini-grounding-purpose]: Google, grounding with Google Search, retrieved 2026-09-03. [^no-selection-logic-published]: Both vendors' grounded-search documentation, read 2026-09-03.