How Claude decides which brands to recommend

Ask Claude "what's the best CRM for a five-person sales team" and — depending on the exact wording — you might get an answer built entirely from what the model already learned in training, or one it just went and checked live, with a numbered citation attached to nearly every claim it makes. Which happens is not random: Claude decides, per request, whether the question needs a live web search, and Anthropic's own documentation says comparative, multi-brand questions are exactly the kind most likely to trigger one. That's the mechanic this post walks through — how Claude's web search tool actually works, what makes it different from ChatGPT's browsing, Gemini's grounding, and Perplexity's always-on citations, and what you can do about it.

Claude decides whether to search, not whether to answer

We've already covered how ChatGPT decides what to recommend (prediction, topped up with retrieval when browsing is on) and how Gemini's dynamic retrieval works (a per-question threshold on whether a Google search is worth running). Claude's version of the same decision is documented in unusually specific terms by Anthropic itself: the model searches when a request "depends on information that is current, changing, or outside its training data" — recent events, current prices, or "information about specific organizations, people, or products that might have changed." It answers directly, without searching, for "established facts" and "analysis of content already provided in the conversation."

Read that list again with a buyer's recommendation question in mind. "What's the best CRM for a five-person sales team" is not a stable, textbook fact — it's a live judgment about specific, competing organizations, which is precisely the category Anthropic names as search-worthy. And the more entities a question asks Claude to weigh, the more it actually searches: Anthropic's own guidance says a simple factual query typically costs 1–3 searches, while "comparative or multientity research can use 10 or more." A "best X for Y" question, by definition, is multi-entity research — you're asking Claude to hold several competitors in its head at once and decide who wins. That's not a one-off browsing feature bolted on; it's the specific shape of question this tool is built to handle with more searches, not fewer.

Every search comes with a citation — that part isn't optional

Here's the detail that makes Claude meaningfully different from the other three engines we track: citations are not conditional. Gemini attaches sources only when grounding fires, and whether it fires is itself a judgment call. Claude's documentation states plainly that citations are always enabled for web search — every claim the model draws from a search result carries a URL, a title, and up to 150 characters of the exact cited text, whether you asked for sources or not. If Claude searched to answer your buyer question at all, there's a citation trail behind it, full stop.

That has a practical consequence worth sitting with: the question isn't just "did Claude search," it's "did Claude search and cite your domain, or did it search and cite three other sites while still mentioning your name in the prose." Those are different outcomes, and only the second half of the split — the actual citation — is the durable, checkable evidence that Claude went and verified something about you specifically, rather than reciting your name from a training-data memory that might be stale, or wrong, or simply absent for a smaller brand.

Where this leaves you, versus the other three

Versus ChatGPT and Gemini: all three tools run the same core loop — the model judges whether a search would help, then folds results back in — but Claude is the only one of the three whose own documentation frames "comparative or multientity research" as the specific trigger for heavier search behavior, and the only one that guarantees a citation the moment it does search. That combination should raise your confidence that a real, sharply worded "X vs Y" or "best X for Y" question is more likely to send Claude out to check the live web than a vague or single-brand question would.

Versus Perplexity: Perplexity cites on almost every answer by design — showing sources is the whole product. Claude's citation guarantee only applies to the subset of answers where it actually searched; an answer it judged "stable knowledge" and answered from memory alone carries no citation at all, because nothing was fetched to cite. The lesson is the same one that runs through every post in this series: a single manual check tells you which mode you landed in this one time, not which mode you'll land in on the next, slightly different phrasing of the same question.

One real caveat worth naming honestly: Claude's web search can be scoped or switched off entirely at the account or organization level — an admin can disable it outright, or restrict it to an explicit allowlist or blocklist of domains. Most consumer and default API usage has it on with the open web available, but if you're evaluating Claude visibility for an enterprise buyer's specific deployment, it's worth knowing the tool itself supports being walled off in ways the other engines we track don't expose the same way.

What actually earns the citation

Given that mechanic, the same three levers that move every engine we've written about apply here, with one Claude-specific emphasis:

• Third-party pages shaped like the comparison, not your own homepage. Claude reaches for evidence it can quote and attribute — the same gap that decides ChatGPT's answers decides which pages Claude's search surfaces and cites.
• A quotable, specific sentence near the top of the page. Citations carry up to 150 characters of the actual cited text — a page that states its category and audience plainly in one sentence gives Claude something clean to lift; a page that buries the same fact under brand-voice copy gives it nothing worth quoting.
• Being reachable at all. None of this works if ClaudeBot can't crawl your site in the first place — check that first with the free AI crawlability checker before investing in content.
• The comparative shape of the question itself is your leverage, not just a variable to accept. Because Anthropic's own docs single out multi-entity research as the search-heaviest case, a real "X vs Y" or "best X for Y" page you publish is aimed at exactly the situation most likely to make Claude go check the live web instead of answering from memory alone.

Why one prompt won't tell you which mode you got

You can't tell from a single Claude reply whether it searched and cited, searched without finding you, or skipped the search step entirely and answered from memory — the same honest limit we've written about for Gemini and Perplexity. Your AI visibility score exists for exactly this reason: one sample from one engine tells you almost nothing reliable about the pattern underneath it.

PingMyBrand runs 25 real buyer questions against Claude specifically — alongside ChatGPT, Gemini, and Perplexity — and shows you, per question, whether you were named, whether Claude cited your domain, and the exact sentence it returned. You can also track Claude on its own on the Claude visibility tracker.

Run the free scan and go straight to the Claude column of your report: no signup, about a minute, and you'll see exactly which of your buyer questions Claude is answering from memory alone — and which ones it actually went and searched, cited sources and all.