How to Get Recommended by Claude: Three Crawlers, Short Citations and a Search Index Nobody Names

Claude answers from memory until it decides to search

Ask Claude about a roofer in Boerne and one of two things happens. Without web search, it answers from training, which is dated and usually blank on small local businesses. With web search, it reads pages and answers with citations. Anthropic's help center says web search is switched on from the + menu in the chat window, and that in the newer Claude experience there is no toggle at all: Claude searches when it judges a search would help.

That judgment is the whole game for a local business. A question with a town or the words 'open now' in it needs current information, which means search. So the useful question is not what Claude remembers about you but what its search finds. Free accounts get web search within limits that reset every five hours. On Team and Enterprise plans an Owner must enable it first, so a buyer on a company account may get an unsearched answer.

Whose search results Claude is reading

Anthropic does not publish the name of its web search provider, so treat this part as reporting, not documentation. In March 2025 Brave Search was added to Anthropic's list of subprocessors, and press reports at the time said testers found Claude's web search returning the same citations as Brave. Anthropic's own help center states one narrower fact outright: Claude's image search uses Bing. Such arrangements change quietly, so no plan should bet on one index.

Anthropic also crawls for search itself, through Claude-SearchBot. Brave, Bing and Anthropic's crawler want the same unexciting things: plain HTML, clearly stated facts and pages that answer the question asked. Plenty of businesses rank on Google and are still invisible to AI because they miss one of the three.

Three Anthropic bots, three separate robots.txt decisions

Anthropic splits its crawling across three user agents, and each answers a different question. Block the wrong one and a business can vanish from Claude's searched answers while believing it only opted out of model training. The table uses Anthropic's own descriptions, plus Vercel's December 2024 analysis of which AI crawlers run JavaScript.

The common mistake: an owner pastes a block-everything list from a forum and shuts out Claude-SearchBot and Claude-User along with ClaudeBot. If you want out of training but into answers, block ClaudeBot alone. Each subdomain needs its own robots.txt, so a rule on www does not cover a shop or booking subdomain. Every other company's tokens are laid out in our guide to AI crawlers.

• To opt out of training only: a group for User-agent: ClaudeBot with Disallow: / and no rules for the other two bots.
• To slow ClaudeBot rather than block it: User-agent: ClaudeBot with Crawl-delay: 1. Anthropic honors this non-standard directive, which many crawlers ignore.
• To stay in Claude's searched answers: no Disallow for Claude-SearchBot or Claude-User, and no catch-all User-agent: * block that sweeps them up by accident.
• To check a visitor that claims to be Anthropic: compare its IP address with the list Anthropic publishes at claude.com/crawling/bots.json.
• To avoid a false sense of security: do not block Anthropic by IP at the firewall. Anthropic warns that a bot which cannot reach robots.txt cannot read your opt-out either.

Rendering matters as much as permission. Vercel found in December 2024 that ClaudeBot did not execute JavaScript; that may have changed, and Anthropic has not said what its other two bots do. Server-rendered pages put the text in the first response, as our Next.js builds and any well-built WordPress site do. The one-minute test is in JavaScript and AI crawlers.

150 characters: the size of a Claude citation

In Claude's developer API, web search citations cannot be switched off, and each carries the page URL, the page title and up to 150 characters of cited text. The consumer apps show direct citations, source links and relevant quotes. A 150-character cap is about one plain sentence, so the passage Claude lifts from your site has to stand on its own.

'We have proudly served our community with excellence for years' carries no fact at all. A hypothetical 'Emergency plumbing in Boerne and northwest San Antonio, 24 hours a day, $129 trip fee' carries four in under 90 characters. The longer treatment is in content AI can quote.

Claude on an iPhone sends local searchers to Apple Maps

Claude's iOS app can ask for a user's location, suggest nearby places and services, show them on a map and open a chosen place in Apple Maps for directions, according to Anthropic's help center. The location tool is not available to Team or Enterprise members. Anthropic does not say where those suggestions come from. What is certain is where the customer lands: on your Apple Maps place card, reading whatever hours and phone number Apple has.

That makes an unclaimed Apple listing a Claude problem as well as a Siri problem. The fix is free in Apple Business, and our page on Siri and Apple Intelligence covers what Apple reads. Keep the same name, address, phone and description there as on Google, Bing Places and your own site. That consistency is the heart of entity optimization.

Apps built on Claude set their own rules

Claude also runs inside other companies' software through its API. A developer using its web search tool pays $10 per 1,000 searches plus tokens, as of October 2026, and can restrict results to an allowed list of domains, exclude a blocked list, and localize results by city, region, country or timezone. A booking app built on Claude may search only the directories its developer chose. You cannot control that list, but you can make sure your trade's major directories describe you accurately.

A one-hour audit of Claude's answers about you

1. 1

   Start a clean chat

   Use a new conversation with no project files or saved notes about your business. If your account shows a web search option in the + menu, turn it on; in the newer experience Claude decides on its own.

2. 2

   Ask like a customer, not an owner

   Use a buyer's words: 'Who installs tankless water heaters in New Braunfels?' rather than your company name. Write eight to twelve such questions; prompt tracking explains how to choose them.

3. 3

   Read the citations, not only the names

   For each answer, note which businesses appear, which pages are cited, and whether your site or a directory about you is among them. A competitor cited from one trade directory tells you exactly where to go next.

4. 4

   Ask about yourself by name

   Ask what Claude knows about your business, once without search and once with it. Wrong hours or a closed location usually trace back to a cited page you can open and fix.

5. 5

   Repeat before you conclude anything

   Answers vary from run to run. Ask the same set every week or two for a month, keep a dated log, and compare like with like.

If that sounds like a lot of copying and pasting, the free AI visibility check puts a customer's question to several AI engines through their APIs, live, and shows each answer, whether you were named and who was named instead. API answers can differ a little from the consumer apps, so treat it as a fast first look. For ongoing tracking, our AICiteBeam platform monitors Claude alongside ChatGPT, Gemini, Perplexity and AI Overviews.

What moves Claude toward naming you

• Let the right bots in. Allow Claude-SearchBot and Claude-User, then decide about ClaudeBot on its own merits.
• Serve the facts in HTML. Hours, services, areas and prices belong in the page source, not only in content that appears after scripts run.
• Rank where Claude reportedly looks. Check your positions in Brave Search for your main services, then in Bing. Weak results in both say more than any 'AI score' a dashboard invents.
• Earn mentions on pages that get cited. Local news, trade associations and chamber directories are the kind of third-party pages BrightLocal found ChatGPT citing in 2024; no one has run that study on Claude. See brand mentions versus backlinks.
• Keep reviews current and answered. A review page is often what a search engine returns for 'best [trade] in [town]'. See reviews and AI recommendations.
• Skip the shortcuts. An llms.txt file is not a Claude strategy: Ahrefs found 97% of llms.txt files got no requests in May 2026, and AI retrieval bots made just 1.1% of the requests that did arrive.

Claude does not rank your business. It repeats whoever the web says you are, about 150 characters at a time.

This is the work inside DataCram's AI search optimization programs, with the same customer questions tested in Claude and the other engines each month. On sites we build, the Delphi agent in Athena asks the AI engines what they say about the business and drafts the facts and FAQs they cite. Programs start at $625 a month; GEO pricing explains how the tier is set after a free GEO audit.

Correcting Claude when there is no one to call

Anthropic offers no business dashboard and no listing to claim. It says Claude can produce incorrect or misleading responses, advises against treating it as a single source of truth, and asks users to report problems with the thumbs-down button. Use it, then fix the source the answer came from, which is what changes the next answer.

Our fix pages cover the usual versions: AI made up facts about my business, AI cites old information and AI has the wrong business information.

Claude in GA4, and why the visit count stays small

When Google added GA4's built-in AI Assistant channel on May 13, 2026, its announcement named Claude as an example source; recognized visits get the medium ai-assistant. If you built a custom channel before then, the hostname to match is claude.ai, which appears in the referrer lists analytics firms publish beside chatgpt.com and perplexity.ai. Visits from desktop and mobile apps often carry no referrer at all and land in Direct, so any channel undercounts. The full setup is in tracking AI traffic in GA4.

Keep expectations for clicks modest. Cloudflare estimated in July 2025, from crawl-to-referral ratios, that getting traffic from Anthropic was about 30,000 times harder than getting it from Google search used to be. That argues for measuring mentions, not for ignoring Claude: a buyer who reads your name may simply phone, and that never shows up as a Claude visit.

Anthropic's rules against OpenAI's and Perplexity's

These three engines have the crawler rules owners most often confuse. Here they are on the points an owner can check.

Each rival has its own guide: ChatGPT, Perplexity and, if Bing matters in your market, Microsoft Copilot. The generative engine optimization guide puts all eight engines side by side, and there is no number one in an AI answer explains why the goal is a place in the short list rather than a rank.