Claude searches only when a question needs current information, cites every source it uses, and attaches up to 150 characters of the exact passage it drew from. That last detail makes it the engine that shows you not just which page was used, but which sentence.
All three behaviours are documented by Anthropic rather than inferred, which is unusual enough in this category to be worth building an audit around.
Most writing about how a given engine “prefers” certain sources is guesswork dressed as analysis. This page sticks to what is documented, because for Claude an unusual amount is, and the documented parts turn out to be more useful than the speculation.
When does Claude search at all?
Not always, and the rule is published. Anthropic states that “Claude determines when to search based on the prompt”, and sets out both sides of that decision explicitly.
Recent events and announcements, current prices and statistics, information about specific organisations or products “that might have changed”, or an explicit request to look something up.
Established facts, maths and science fundamentals, coding concepts, creative writing, or analysis of material already in the conversation.
The practical read is that vendor questions are close to the paradigm case for triggering a search. Ask about a company’s current pricing or how it compares to another, and you are asking about something that might have changed, which is one of the stated triggers.
What does a citation actually contain?
More than anywhere else, and this is the genuinely distinctive part.
Anthropic documents that “citations are always enabled for web search”, and that each one carries the URL, the title of the source, and cited_text, which is “up to 150 characters of the cited content”.
Which page. Every engine that cites gives you this much.
Useful for grouping sources by type when you are counting domains.
Up to 150 characters of the text actually used. This is the part almost nothing else gives you.
That is a meaningful difference for anyone trying to work out why a page got cited. A URL tells you a document was relevant. A quoted passage tells you which two sentences of it did the work, which is directly actionable in a way a link never is.
Why does that make it good for diagnosis?
Because it closes the loop between what you published and what got used.
If a competitor’s comparison page is cited, you can see the exact claim that was lifted. If your own documentation is cited, you can see which paragraph earned it and write more like that one. If a stale third-party listing is cited for your pricing, you get the offending sentence rather than a page you then have to search.
Every other engine tells you that a page mattered. This one tells you which sentence did.
What is the retrieval architecture underneath?
The same pattern as the rest of the live-search family, with one API-level detail worth knowing.
Anthropic documents that the API “runs the searches and provides Claude with the results”, and that “this process can repeat multiple times throughout a single request”. So a single answer can involve several rounds of searching rather than one, which is closer to how a person researches than to a single lookup. Simple factual questions are noted as typically using one to three searches, while comparative or multi-entity research “can use 10 or more”.
That last figure is the interesting one for this category, because a buyer comparing four vendors is precisely a multi-entity research question. More searches means a wider retrieved set, which means more opportunities to be included and more competitors in the pool alongside you.
Are there controls that affect what it sees?
Yes, at the API level, and they are worth knowing because they explain why two people can get different answers.
The web search tool accepts allowed_domains and blocked_domains, so an application built on Claude can restrict which sources are eligible entirely. Anthropic notes that administrators can also restrict which domains it searches from the console.
The consequence for you is that “Claude” is not one uniform environment. A product built on the API with a domain allowlist is a different retrieval surface from the consumer assistant, and being absent from one says nothing about the other. That is a caution about generalising, not something you can act on.
What should you actually do about it?
The same work that helps everywhere else, with one emphasis that follows directly from the citation format.
How does this compare with the other engines?
Claude sits in the live-search family alongside ChatGPT and Perplexity, and differs from the grounded family in where its corpus comes from.
The live-search engines run their own retrieval. OpenAI documents a dedicated crawler for the index behind ChatGPT’s search features, kept separate from its training crawler, and Perplexity documents the same split. Google’s surfaces are grounded in the Search index instead, where a page qualifies when it is “indexed and eligible to be shown in Google Search with a snippet”.
The architecture underneath all of them is the same retrieval-augmented pattern set out in Lewis et al.: fetch documents at query time, generate from what came back. What differs between engines is which documents are reachable and how much of the working is shown.
Cited passage alongside the URL and title, so you can see the sentence that was used rather than only the page.
No sources at all, because none were fetched. A real mention with nothing behind it to inspect or fix.
What is not documented?
How sources are ranked or preferred, and it is worth saying plainly because plenty of writing claims otherwise.
Anthropic documents when Claude searches, what a citation contains, how many searches a request tends to use, and what controls exist. It does not publish how candidate sources are scored against each other, which types of domain are favoured, or how the cut-off is decided. Anyone telling you that this engine prefers primary sources over aggregators is describing an impression rather than a documented behaviour.
That gap is not a flaw in the documentation. It is the same gap every engine has, and the honest response is to measure rather than to theorise.
How the engines split into retrieval families is in the engine guide, the architecture underneath all of them is in the RAG explainer, and what a citation is and is not is in why an AI citation is not a backlink.