Claude decides which businesses to recommend by searching the live web when a question calls for it, then naming whichever businesses its search results and training data describe clearly, consistently, and with independent evidence behind them. It does not rank businesses the way Google does. It synthesizes an answer from whatever it retrieves, and Anthropic has published more detail about that mechanism than any other major AI lab.
That matters because most AEO advice is written for ChatGPT and applied to every engine unchanged. Claude works differently in three specific, documented places: how it searches, how it cites what it finds, and how willing it is to crown a single winner.
Anthropic's own developer documentation for the web search tool and crawler policy, plus published Anthropic research on how Claude is trained to avoid unwarranted agreement. Where a claim comes from third-party testing rather than Anthropic itself, we say so.
How does Claude actually decide which businesses to recommend?
It decides whether the question needs current information, and if so, calls a web search tool, reads what comes back, and writes an answer with citations attached.
Anthropic's documentation for the web search tool describes the sequence directly: Claude determines when to search based on the prompt, the API runs the search and returns results, and this can repeat multiple times in a single turn before Claude gives its final, cited answer (Anthropic). Claude searches for recent events, current prices or statistics, and information about specific organizations, people, or products that might have changed. It answers directly, without searching, for established facts, math, and anything already in the conversation (Anthropic).
That second category is the one worth sitting with. Unless your business is already well enough known to count as "established," Claude's training data alone is not enough to get you named. It needs to search, and it needs to find you when it does.
The volume of searching scales with the question. Anthropic's own guidance says single-fact queries typically use one to three searches, while "comparative or multientity research can use 10 or more" in a single turn (Anthropic). A question like "who's the best AEO agency for a mid-size law firm" is exactly that kind of multi-entity research. Claude may run several separate searches to cover it, generating its own queries rather than reusing the customer's exact words, which means you are competing for each generated query separately, not for one phrase.
There is also a limit on how much searching a single answer can contain, and it is set by whoever built the product asking Claude the question, not by the person typing. The max_uses setting caps the number of searches a developer allows in one turn, and a search tool call that fails still counts against that cap (Anthropic). A cost-conscious integration that caps Claude at two or three searches per answer gives your business far fewer chances to surface in a crowded, comparative query than one that allows ten. That ceiling is invisible to the end user and to you, which is one more reason the same question can get a different answer depending on where it is asked.
Does Claude need web search turned on to find your business?
Yes, for anything current, and that toggle is not always on.
On claude.ai, a person enables web search from the chat input, and on Team or Enterprise plans an organization owner has to turn it on first before members can use it at all (Anthropic support). On the API, it is not on by default: a developer has to explicitly add the web_search tool to the request. Anthropic prices it separately from ordinary usage, at $10 per 1,000 searches plus standard token costs (Anthropic).
The practical effect is that a product built on Claude without that tool enabled is answering purely from training data, frozen at a cutoff date, with no way to find your business if it opened, rebranded, or earned a new review after that point. A lot of AEO advice assumes every AI answer is a live search. For Claude specifically, that assumption needs checking on a case-by-case basis, because the search step is optional infrastructure, not a given.
Which crawlers does Claude actually use, and why does blocking the wrong one backfire?
Three separate ones, each doing a different job, and Anthropic documents all three by name.
Its support article lists ClaudeBot, which "collects public web content that may be used to train Anthropic's AI models," Claude-User, which "fetches pages when a Claude user's question requires it," and Claude-SearchBot, which "indexes content to improve the relevance and accuracy of search responses" (Anthropic support). Training, live retrieval, and search indexing are three different systems with three different robots.txt rules.
Blocking ClaudeBot alone only opts you out of future training data. It does nothing to your search visibility. If a security preset or a blanket "block all AI bots" rule also catches Claude-User and Claude-SearchBot, Anthropic's own wording is that doing so "may reduce visibility in Claude's search results." Check all three tokens in your robots.txt and your WAF rules, not just the one with the recognizable name.
Independent testing adds one more wrinkle worth noting cautiously. One analysis of live Claude search results found 86.7% overlap with Brave Search's top non-sponsored results, against only 20 to 27% overlap with Bing or Google across the same queries, and concluded Claude's web search likely draws on Brave's index for at least some live lookups (Profound). Anthropic has not confirmed this publicly, so treat it as a plausible pattern from third-party testing, not a documented fact the way the three crawlers are.
How does Claude decide what to cite once it has search results?
It attaches a citation to the specific sentence that used it, not to the page as a whole, and that citation is short.
Each web_search_result_location in Claude's response carries the source url, its title, and cited_text, which is capped at "up to 150 characters of the cited content" (Anthropic). Results also carry a page_age field recording when the source was last updated. Both details point the same direction: Claude is matching a narrow, recent fragment to a narrow part of its answer, not crediting a whole article for a general impression.
This lines up with what has already been measured across AI search generally. Fresher content gets cited more: AI platforms cite content 25.7% fresher than what ranks organically in traditional search (Ahrefs). A page where the answer to one specific question sits in a clean, self-contained paragraph under its own heading gives Claude exactly the 150-character fragment it needs. A page where that same answer is spread across three paragraphs of context does not.
Does Claude recommend local businesses the way Google or ChatGPT does?
Partly. Claude can localize a search, but it has no equivalent to Google's dedicated business-listing database behind it.
The web search tool accepts a user_location parameter, approximate city, region, country, or timezone, which Anthropic says is used "to localize search results" (Anthropic). So a question with a local dimension does get location-aware search queries. What Claude does not have is anything like Gemini's separate Grounding with Google Maps, a second data source built from business-profile data rather than web pages. We cover that mechanism in how Gemini recommends businesses, and it has no parallel here.
For Claude, a local business is found or missed entirely on the strength of its public web footprint: the website, the directory listings, the review platforms, and local press coverage that Claude-User and Claude-SearchBot can actually reach. There is no profile database filling the gap when that footprint is thin.
Why does Claude sometimes refuse to just name one business?
Because Anthropic has specifically studied, and trained against, the tendency of chat models to tell people what they want to hear.
Anthropic's own 2023 research paper, Towards Understanding Sycophancy in Language Models, found that sycophancy, agreeing with or flattering a user rather than giving the most accurate answer, is a general behavior that emerges from the standard way these models are trained, not a quirk unique to one model. Anthropic's published constitution states the priority explicitly: Claude is meant to be "exceptionally helpful while also being honest, thoughtful, and caring about the world," with safety and honesty ranked above simple helpfulness when they conflict.
The practical effect shows up exactly where a business owner notices it most: ask Claude to recommend "the best" option in a category with thin or conflicting evidence, and it is more likely than some other engines to list several businesses with caveats, or to say plainly that it does not have enough information to pick one confidently, rather than confidently crown a single winner. That caution is a trained behavior, not a bug.
This connects to a gap we have written about separately: being read by an AI engine and being recommended by it are two different outcomes, and the distance between them is often wider than people assume. Our breakdown in the retrieval gap between cited and recommended businesses covers why a business can supply the content behind an answer and still not be the business that gets named in it.
What can Claude not see about a business?
Everything outside what its own crawlers and search tool can retrieve, which is a narrower window than most owners assume.
Claude has no access to a company's CRM, its private sales data, its reputation inside a closed community, or anything a happy customer said out loud to a colleague rather than in writing somewhere public. It only knows what ClaudeBot picked up during training, plus whatever Claude-User and Claude-SearchBot can fetch and index live. A genuinely excellent business with almost no public footprint and a mediocre one with a complete website, active listings, and frequent independent mentions will not score the same, even though quality says the opposite should happen. AI visibility measures public, retrievable evidence. It does not measure quality directly, and treating the two as the same thing is the most common mistake in this space.
There is a second blind spot worth naming. Because Claude is steerable through a system prompt, a business built on top of Claude (a support bot, a shopping assistant, a research tool) can be configured to search more or less aggressively, or to skip search entirely for speed and cost (Anthropic). Two products running the identical Claude model can behave completely differently toward your business depending on choices a third-party developer made that have nothing to do with you. Testing "what does Claude say about us" only in claude.ai itself will not tell you how a Claude-powered product elsewhere treats the same question.
How is optimizing for Claude different from optimizing for ChatGPT or Perplexity?
The underlying signals overlap heavily. The retrieval and citation mechanics do not.
| Claude | ChatGPT | Perplexity | |
|---|---|---|---|
| Retrieval | Own web_search tool, decides per prompt | Web search plus training data | Own exabyte-scale index |
| Crawlers | ClaudeBot, Claude-User, Claude-SearchBot | Not separately documented | PerplexityBot, documented |
| Citation unit | 150-character cited fragment | Passage level | Fine-grained document fragment |
| Local data | user_location parameter, no listings database | Web sources and listings | Web sources |
| Paid placement | None documented | None in organic results; ads introduced separately | None documented |
| Stated caution | Trained against sycophancy (Anthropic) | Not published | Not published |
The overlap is still the more useful half of the story for anyone deciding where to spend effort. Yext's analysis of 6.8 million AI citations found 86% come from brand-managed sources, your own site plus the directories and listings you control, and that holds regardless of which engine is asking (Yext). Ahrefs' finding that branded mentions correlate with AI visibility more strongly than any other measured factor, at 0.664, is not platform-specific either (Ahrefs). Build for fragment-level retrieval, keep pages current, and earn independent mentions, and you have covered most of what moves all three engines, including Claude. Our comparison of how AI search platforms actually differ goes deeper on where the remaining differences sit.
What is genuinely Claude-specific is narrower: three named crawlers instead of one, a 150-character citation fragment rather than a passage, no local-listings database, and a documented bias toward hedged, multi-option answers over a single confident pick when the evidence is thin. For a direct look at how ChatGPT handles that same decision, see how ChatGPT recommends businesses; for Perplexity's published scoring model, see how Perplexity chooses sources.
What should a business actually do about Claude specifically?
Five things, in order of how directly they follow from the mechanism above.
Check all three crawler tokens, not one. Audit robots.txt and your WAF or bot-management rules for ClaudeBot, Claude-User, and Claude-SearchBot separately. Blocking the first without realizing you also blocked the other two is the single most common way a business becomes invisible to Claude without anyone deciding that on purpose.
Keep your entity description identical everywhere. Name, services, location, and positioning need to match across your website, your directory listings, and your review profiles. Claude is deciding whether evidence about a business is consistent enough to cite confidently, and contradictory descriptions read as weaker evidence, not more of it.
Write each section as a complete, self-contained answer. Since a citation is capped at 150 characters of matched text, the paragraph that answers a specific question needs to do it on its own, under a clear heading, rather than depending on three paragraphs of setup before it.
Earn independent, branded mentions. Reviews, press, directories, industry roundups. This is the strongest measured signal for AI visibility generally, and it matters more for Claude specifically because it is also the kind of independent evidence that makes a hedging model comfortable naming you rather than listing you as one of several unverified options.
Measure whether you are named at all, across phrasings, before worrying about being named first. Given Claude's documented caution about picking a single winner, the useful metric is usually whether Claude mentions the business across a realistic set of buyer questions, not whether it calls the business the best.
See what Claude actually says about you
A free AI visibility audit across Claude, ChatGPT, Perplexity, Gemini, and Google AI Overviews, run on the real questions your buyers ask.
Book your audit
Ashur Homa
Built and scaled a digital brand to $100M+ in sales with zero ad spend. Has helped businesses generate millions through AI go-to-market strategy. Leads growth at Omni Eclipse.
Connect on LinkedIn