Guide

How does Claude choose which businesses to recommend?

Updated · 9 min read · By the SearchCrow team

Claude recommends businesses by searching the web when it judges the question needs current information, reading the pages that come back and writing an answer with citations to them. Anthropic says Claude searches for "information about specific organizations, people, or products that might have changed", which describes most questions about a local business. Without web search, Claude answers from training data that stops at a knowledge cutoff. Anthropic has not published a ranking formula, so the practical route is to be crawlable, clear and well documented on the sites Claude cites.

Key takeaways

  • Claude does not search for every question. It searches when a question benefits from current information, and a user can tell it to search or not.
  • Anthropic runs three crawlers. Claude-SearchBot indexes pages for search, Claude-User fetches a page when a user asks, and ClaudeBot collects content for training. Only ClaudeBot is safe to block without affecting answers.
  • Every Claude response that uses web search includes citations, so you can see which sites it trusted for your trade.
  • With search off, Claude relies on training data with a stated knowledge cutoff (June 2026 for the current models), so new or recently changed businesses depend on search to be found.
  • The work that gets you recommended by Claude is the same work that helps ChatGPT, Perplexity and Gemini: crawlable pages, clear facts, consistent listings and genuine reviews.

How does Claude answer a question about a business?

If web search is on, Claude decides whether the question needs a search, runs one or more, and writes an answer from the results with citations.

Anthropic's help centre describes web search as letting Claude "search the internet for current information", with answers grounded in live web content, and says Claude invokes search when a question "benefit[s] from current information" (Anthropic, 2026). The same article says the search can use your location, inferred from your IP address, to localise results, which is why "best accountant near me" (an illustrative example) can return businesses in the right town.

Anthropic's developer documentation for the same search tool is more specific about what triggers a search (Anthropic, 2026). Claude searches when the request depends on information that is "current, changing, or outside its training data", and the list of examples includes "current prices, rates, scores, or statistics" and "information about specific organizations, people, or products that might have changed". It answers directly, without searching, for "established facts", creative writing and conversation.

Three details from that documentation matter for a business:

  1. Claude can search more than once. The search step "can repeat multiple times throughout a single request"; Anthropic says single-fact questions typically use one to three searches and comparisons can use ten or more. "Compare the three best boiler installers in Leeds" (an illustrative example) is likely to trigger several searches, each a chance to be found.
  2. Results are localised. The tool accepts a user location with city, region, country and timezone so that results match where the customer is.
  3. Citations are always on. Each citation carries the page's URL, title, when the page was last updated and up to 150 characters of the text it relied on. Claude is quoting a specific passage, so the sentence that answers the question is what gets lifted.

Anthropic's help pages do not say which search index powers Claude's web search. In March 2025 TechCrunch reported that Anthropic had added Brave Search to its list of subprocessors and that a developer found a BraveSearchParams parameter in Claude's search function (TechCrunch, 2025). Treat that as reported rather than documented.

It answers from what it learned in training, which has a fixed end date.

Anthropic says that "each model has a knowledge cutoff" and that "these models may not be aware of events or information that occurred after their respective cutoff dates" (Anthropic, 2026). For the current models (Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2) that cutoff is June 2026; the developer documentation lists the same date as the "reliable knowledge cutoff" (Anthropic, 2026).

For a business, this means two things:

  • If you opened, moved, rebranded or changed prices after the cutoff, only a search can find the new facts. An answer from training alone may describe the old business or not know you at all.
  • What the model remembers is a snapshot of the wider web. Businesses that were written about, reviewed and listed widely before the cutoff are more likely to come up without a search.

Anthropic itself tells users to "carefully scrutinize any high-stakes advice given by Claude", warns that Claude can "hallucinate" and advises reviewing cited sources when web search is used (Anthropic, 2026). The same applies to an owner checking what Claude says about them: look at the citations, not just the answer.

What are ClaudeBot, Claude-SearchBot and Claude-User, and what happens if I block them?

Anthropic documents three crawlers with different jobs, and only one of them is about training.

All three are described on Anthropic's crawler page, which says its bots honour robots.txt (Anthropic, 2026):

CrawlerWhat it doesIf you block it
Claude-SearchBot"Navigates the web to improve search result quality" by analysing content for relevance and accuracyYour content is not indexed for search, which Anthropic says "may reduce its visibility and accuracy" in search results
Claude-UserFetches a website when an individual asks Claude a question that needs itClaude cannot retrieve your content for that user's query, which may reduce visibility in user-directed search
ClaudeBotCollects web content that may be used to train Anthropic's modelsSignals that your site's future material should be excluded from training datasets. No documented effect on search answers

Three practical points from the same page:

  • Each bot has its own rule. A Disallow: / under User-agent: ClaudeBot does not touch Claude-SearchBot or Claude-User, and the reverse is also true. If you want to opt out of training but stay in answers, block ClaudeBot only.
  • You can slow a bot rather than block it. Anthropic supports the non-standard Crawl-delay directive, for example Crawl-delay: 1.
  • Do not block by IP address. Anthropic warns that blocking its bots' IP addresses "may not work correctly or persistently", because the bots then cannot read your robots.txt to learn your preferences. Use robots.txt instead.

The check takes two minutes: open yourdomain.co.uk/robots.txt and look for Disallow: / under User-agent: Claude-SearchBot or User-agent: Claude-User. Blanket "block all AI bots" rules often catch the search crawlers too. Our AI visibility checklist has a copy-paste example that allows the search crawlers for every major assistant and leaves the training crawlers as your choice.

What kinds of pages does Claude cite?

Anthropic documents the mechanism (search results, short passages quoted with a citation) but not a list of favoured websites.

Two things follow from the documentation (Anthropic, 2026). A page the search cannot find, or that Claude-User cannot fetch, cannot be cited. And because each citation is an extract of up to 150 characters, a sentence such as "we are a Gas Safe registered heating engineer covering Harrogate and Knaresborough" (an illustrative example) is the kind of thing that gets quoted.

What practitioners observe (not confirmed by Anthropic)

People who track Claude's answers commonly report that it cites a mix of the business's own site, review platforms, directories and local or trade publications, with the mix shifting by question. Anthropic has not confirmed this and answers vary between runs. Treat these as places to look, then confirm with the citations from your own five-question check below.

Make your site crawlable by Claude-SearchBot and Claude-User, put the facts in plain text, and be present on the sites Claude cites for your trade.

  1. Allow Claude-SearchBot and Claude-User in robots.txt. Remove any Disallow rules for them, and check your host or CDN for "block AI bots" settings that might catch them. Block ClaudeBot separately if you want to opt out of training (Anthropic, 2026).
  2. Write the answer in one quotable sentence. Because Claude cites short passages, each service page should open with a sentence or two that states what you do, where, for whom and roughly what it costs. Detail goes below, not above.
  3. Put the facts in text, not images. Prices or "from" prices, areas covered, opening hours, qualifications and contact details should be readable in the page source. A crawler cannot cite a price that only exists in a photo of your menu.
  4. Keep your name, address and phone number identical everywhere. Website, Google Business Profile, Bing Places, Apple Maps and the directories for your trade. Anthropic has not documented a consistency check, but an assistant that finds two phone numbers has less reason to trust either.
  5. Get listed and reviewed where Claude looks. Run the check below and note which sites are cited. If a directory, review platform or local publication keeps appearing and you are not on it, that is your next job. Our Google Business Profile and reviews guide covers the review side.
  6. Show you're real and current. An about page with named people, trading history and accreditations with registration numbers gives Claude a trustworthy page to cite when a customer asks "is [business] any good?" (an illustrative example).
  7. Fix what Claude gets wrong at the source. If Claude quotes an old price or a closed branch, the citation tells you which page it came from. Fix that page or listing and the next search picks up the change.

Our guide to how AI assistants choose businesses covers the other assistants, and the AI search optimisation guide goes deeper on writing pages that get cited.

How do I check whether Claude recommends my business?

Ask it five customer questions with web search on, then the same five with it off, and read the citations.

  1. Turn web search on. In a chat, click the "+" button in the lower left, select "Web search" and look for the tick. On Team and Enterprise plans an Owner has to enable it first under Organization settings > Capabilities. In the newer Claude experience there is no toggle; Anthropic says Claude "searches the web when it helps", and you can type "Search the web" in your message to make sure (Anthropic, 2026).
  2. Write five questions in a customer's words. For a solicitor in Exeter, illustrative examples might be "best conveyancing solicitor in Exeter", "how much does conveyancing cost in Exeter?", "is [your firm] any good?", "[your firm] vs [competitor]" and "solicitor near me for a house purchase".
  3. Ask each one and read the citations. Note whether you are named, whether what Claude says about you is accurate, and every site it cites.
  4. Ask the same questions with search off (tell Claude not to search). This shows what it believes from training data alone. If it does not know you, or describes an old version of you, that is the knowledge-cutoff effect; only live search corrects it until the next model.
  5. Repeat on another day. Answers vary between runs, so one screenshot proves little. Three runs of five questions is a fair picture.
  6. Fix the biggest gap first, then check again a month later.

That is manageable by hand for one assistant. It gets tedious once you add ChatGPT, Perplexity, Gemini and Google's AI features, plus competitors.

How is Claude different from ChatGPT and Perplexity in what you see?

All three search the web and show sources, but they differ in when they search and how the sources are presented.

ClaudeChatGPTPerplexity
When it searchesWhen it judges the question benefits from current information; users can ask it to search or not (Anthropic, 2026)May search automatically when a question benefits from current information; users can start a search manually from the tools menu (OpenAI, 2026)Searches the internet in real time for every question (Perplexity, 2026)
How sources appear"Every response includes citations": direct citations, source links and relevant quotesCitations you can select to open, with a Sources panel for cited sources and other relevant linksNumbered citations linking to the original sources on every answer
Search crawler to allowClaude-SearchBot and Claude-UserOAI-SearchBotPerplexityBot

The practical difference is the first row. Perplexity always looks you up; ChatGPT and Claude decide, so some answers about you come from training data and may be out of date. For the other two, see how Perplexity chooses businesses and how to get recommended by ChatGPT.

How SearchCrow helps

Claude is one of the seven assistants SearchCrow tracks on the Growth plan and above (the free check covers ChatGPT and Perplexity; see pricing). Each week it asks Claude and the others the questions your customers ask, records whether you were named, who was named instead and which sites were cited, and turns the gaps into an action plan with drafts written for you. Everything in this guide works without it; SearchCrow saves you the monthly hour of asking.

FAQ

How does Claude decide which businesses to recommend? When web search is on, Claude decides from the question whether it needs current information, runs one or more searches, reads the results and writes an answer with citations to the pages it used. Anthropic says Claude searches for information about specific organisations that might have changed, which covers most business questions. It has not published a ranking formula, so the practical route is to be crawlable, clear and well documented elsewhere.

What is the difference between ClaudeBot and Claude-SearchBot? ClaudeBot collects web content that may be used to train Anthropic's models. Claude-SearchBot analyses pages to improve Claude's search results, and Claude-User fetches a page when a user asks Claude a question. Anthropic says blocking ClaudeBot only excludes your future content from training. Blocking Claude-SearchBot or Claude-User stops Claude indexing or retrieving your site, which can reduce your visibility in answers.

Does Claude always search the web before answering? No. Claude searches when a question benefits from current information, and users can tell it to search or not to search. Without a search, it answers from training data, and Anthropic says each model has a knowledge cutoff and may not be aware of events after it. The current models' reliable knowledge cutoff is June 2026, so a business that opened or changed after that depends on search to be found.

How do I check whether Claude recommends my business? Open Claude with web search on, ask five questions a customer might ask, such as the best provider of your service in your town, and read the citations on each answer. Note whether you are named, whether the facts are right and which sites are cited. Then ask the same questions with search off to see what Claude believes from training alone. Repeat another day, because answers vary.

How is Claude different from ChatGPT and Perplexity for a business? All three search the web and cite sources, but they show it differently. Perplexity puts numbered citations on every answer, ChatGPT shows citations you can select plus a Sources panel, and Claude includes citations with source links and sometimes quotes. Claude only searches when it judges the question needs it, so some answers come from training data alone. The fixes that help one help all three.

Sources

Free check

Does ChatGPT recommend your business?

Free, in under a minute. You’ll see whether AI assistants mention you, who they recommend instead, and your top 3 fixes.

We’ll read your website for your business details, so there’s less to type.

Free, no card needed. We ask ChatGPT and Perplexity 5 questions your customers might ask, and check your website. We use your email to send you your results and a few short follow-ups about them, which you can stop with one click; see our privacy notice.