A citation from an assistant still has to be opened

Start here

Some assistants answer entirely from what they learned during training; others reach out to the live web mid-conversation and come back with links. A reader cannot always tell which happened just by looking at a confident paragraph, which is why the vendors document when browsing triggers and what a citation represents. Anthropic's web search help page states the trigger plainly: when a person asks about topics that benefit from current information, Claude invokes a search tool to ground its response in content from the live web, and a person can also ask for it explicitly.

What the documents say

All three vendors document search as an added tool layered onto the base model, not a constant background process. OpenAI's web search help page states ChatGPT can search automatically when a question would benefit from current information, and that it rewrites a question into targeted queries sent to search partners rather than passing the original text through unchanged. Google's Gemini Apps privacy documentation states more narrowly that some responses are grounded on Search results, without the trigger detail the other two describe. On citations, Anthropic's page states every response includes citations so a person can verify sources themselves, and OpenAI's page instructs a reader to select Sources to view cited links. Both then qualify that promise: OpenAI's states results and citations can be incomplete, outdated, or incorrect, and Anthropic's lists limitations including that website links may occasionally not function.

Check this

The concrete check both pages point to is the same one: open the cited source rather than trusting the citation's presence as proof. OpenAI's page adds a second step, checking when a page was published or updated, since a linked page can be real and still stale for a fast-moving topic. A reader can also test whether a response actually used live search at all, since a citation-free answer to a time-sensitive question is itself informative: it suggests the assistant answered from training data rather than the web.

What holds and what fails

The claim that a citation is checkable evidence, not decoration, holds as far as these vendors document their own citation features: a link is supposed to point to a real source actually consulted. It fails wherever a link is broken, outdated, or the underlying document has changed since the assistant read it, all of which the vendors name as known limitations rather than rare edge cases. Whether a given assistant even has a browsing mode at all is worth checking first, since answers from training data alone still describe plenty of chat interfaces and older model configurations.

• Open a cited source rather than accepting that a citation exists as proof of accuracy.
• Check the publication or update date on a cited page for any time-sensitive question.
• Confirm whether the specific assistant and mode in use actually searches the web before trusting current-events answers.

A citation is the vendors' own invitation to check their work, not a substitute for checking it. The documentation describing how search and citation work is also the documentation that says, in each vendor's own words, why the citation still needs opening.