How to get cited by Claude: a guide for B2B SaaS teams
Claude cites a page when it searches the web to answer, and Anthropic says it searches when a question depends on information that changes, such as current prices or the details of a particular product. Most of what a software buyer asks is of that kind, and in our view four jobs follow. Let in the two Anthropic agents that serve answers, Claude-SearchBot (indexing pages for search) and Claude-User (fetching a page when someone's question calls for it); ClaudeBot, the third, only gathers training material. Keep the pages buyers ask about indexable and outside any login, because Anthropic says a noindex tag keeps a page out of Claude's search answers. Put the facts in the HTML your server sends, since Anthropic's page-fetching tool does not handle sites that need JavaScript to render. And state each fact in a short sentence that stands on its own, because Claude's citations point at passages, not whole pages.
The Rank Master does this work for B2B SaaS teams and measures Claude alongside the six other engines buyers ask first, with published pricing and no annual lock-in. Book a 30-minute call to see what Claude says about your category today.
When Claude searches, and when it answers from training
Claude does not search for every message. Anthropic's documentation for its web search tool lists what sets a search off: recent events and announcements, current prices and statistics, facts about specific organisations, people or products that may have changed, and explicit requests to look something up. Requests that rest on stable knowledge, such as settled facts, coding concepts or material already in the conversation, are answered without one.
For a software vendor, that puts "How much does [product] cost?", "Does [product] integrate with [CRM]?" and "[product] vs [rival]" squarely in the searched group, while "what does a customer data platform do?" is more likely to be answered from training. The same documentation says a simple factual question typically takes one to three searches and comparative research across several entities ten or more, which is the kind of research a buyer comparing five vendors asks for.
Training sets the baseline. As of 17 October 2026, Anthropic's help centre gives June 2026 as the training cutoff for Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, and warns that a model may not know what happened after its cutoff. A plan you repriced since then reaches Claude only when it searches.
Three settings decide whether that search happens:
- The workspace. On Team and Enterprise plans, an Owner or Primary Owner has to switch web search on for the organisation, under Organization settings and then Capabilities, before any member can use it (web search help). A buyer whose company never did gets answers with no live sources behind them.
- The prompt. Anthropic suggests writing "Search the web" into a prompt to make sure Claude searches. In what Anthropic calls the new Claude experience there is no toggle at all: Claude searches the web when it judges that a search will help.
- Research. On the paid plans (Pro, Max, Team and Enterprise), Research runs several searches that build on one another and can read the user's connected tools, such as Gmail, Google Calendar and Google Docs, for the same report. Your public pages may be read next to a buyer's own notes about you.
Anthropic's three agents, and what a robots.txt rule changes for each
Anthropic describes three agents in its help article for site owners, last updated on 7 April 2026, and each answers to its own robots.txt token. Our guide to AI crawlers and robots.txt sets them beside the agents of OpenAI, Perplexity and Google in one table.
| Agent | What Anthropic uses it for | What blocking it does |
|---|---|---|
| ClaudeBot | Gathering web content that may be used to train Anthropic's models | Tells Anthropic to leave the site's future material out of its training datasets |
| Claude-SearchBot | Crawling and analysing pages to improve the quality, relevance and accuracy of Claude's search results | Stops Anthropic indexing the pages for search, which Anthropic says may make the site less visible, and less accurately described, in search answers |
| Claude-User | Fetching a page when a person's question to Claude calls for it | Stops Claude retrieving the pages for users' questions, which Anthropic says can make the site less visible when users have Claude search the web |
Anthropic's page spells out the training role: ClaudeBot works by "collecting web content that could potentially contribute to their training", meaning its models' training. In our view, the sensible set-up for most SaaS marketing sites allows the two agents that serve answers and decides about the third on purpose:
User-agent: Claude-SearchBot
Allow: /
User-agent: Claude-User
Allow: /
User-agent: ClaudeBot
Disallow: /
Drop the last group if you are happy for your pages to be used in training, and think twice before keeping it. Claude answers stable questions from training, and Anthropic frames a ClaudeBot block as a signal about future material. Our reading is that a block does not remove what current models learned; it means later models learn about your category from what other sites say about you.
Four more points from the same article shape how you check access:
- Claude-User follows robots.txt. Anthropic says its agents honour robots.txt and presents the Claude-User token as the control for requests users start. That is not the norm: OpenAI says robots.txt rules may not apply to ChatGPT-User, and Perplexity says Perplexity-User generally ignores them. With Claude, a disallow reaches the live fetch too; Anthropic's developer documentation lists robots.txt among the reasons its fetch tool refuses a URL.
- No CAPTCHA solving. Anthropic says its agents will not try to get past CAPTCHAs, so a bot-protection rule that challenges unfamiliar visitors stops them whatever robots.txt allows. Ask whoever runs your CDN how it treats the three agents.
- One address list, and no IP opt-out. The article links to bots.json and says a crawler whose source address is on it is Anthropic's. The file we read was created on 7 October 2026 and does not split addresses by agent, so it confirms a visit is genuine while the user-agent string names the agent. Anthropic warns that blocking its addresses may not reliably keep you out, because it stops Anthropic reading your robots.txt file.
- Every subdomain answers for itself. Anthropic asks site owners to repeat their rules on each subdomain they want covered, and it supports the non-standard Crawl-delay line. Check the docs, help and status hosts as well as www.
Where Claude's search results come from
Claude-SearchBot is not the whole picture. Anthropic's help article on blocking and removing content, updated on 3 September 2026, says a noindex robots meta tag tells Anthropic's search partners not to index a page, so they do not pass it to Claude and it stays out of answers that use web search. Password-protected content stays out too. So a noindex tag left on a pricing or comparison template costs you Claude's search answers as well as ordinary search results, and the pages that answer buying questions (plan limits, integrations, security reports) need to sit where anyone can read them, even if deeper docs stay gated.
Anthropic does not name its text search partners in the help articles we read. Its web search article names Bing only as the source of image results, and we draw no conclusion from that about text results.
When Claude opens a page itself, Anthropic's documentation of the web fetch tool sets three limits. It is written for developers; the Claude apps have a web fetch feature too, though Anthropic does not say whether its limits are identical:
- No JavaScript rendering. Anthropic says the tool cannot handle pages that JavaScript renders in the browser, so a pricing table drawn by a script may simply not be there for it.
- Only addresses already in the conversation. Claude can fetch a URL only once it has appeared in the user's message or in earlier search or fetch results. Your page has to come up in search first, or the buyer has to paste the link.
- Text, HTML and PDF, possibly cached. Other content types are refused, and the tool may return a cached copy rather than the latest version.
On the search side, each result from Anthropic's web search tool carries a page_age field, which Anthropic describes as when the site was last updated. Keep the visible update dates on fact-bearing pages accurate, and move them only when the content changes.
How Claude shows the sources it cites
In the Claude apps. While Claude searches, the reader sees an indicator that it is searching the web, and the answer arrives with direct citations to its sources, links for further reading and, where they help, quotations. Anthropic's web search help says every response that uses web search includes citations, and that image results, which come from Bing, each carry a link to their source.
In products built on Claude. Many SaaS tools and internal assistants run on Anthropic's API, where the rules are stricter. Citations are always on for web search, and each carries the page's URL and title and up to 150 characters of the text it cites. Anthropic's documentation says citations to the original source must be included when an application shows the output directly to end users. Developers can restrict which domains the tool searches, and administrators can restrict domains or switch web search off in the Claude Console. The tool is not available on Amazon Bedrock, so an application that reaches Claude through Bedrock cannot use Anthropic's search to cite you.
What a 150-character window means for your copy. A citation points at a sentence or two. Put each fact a buyer might ask about in one plain sentence that names the product and makes sense without the paragraph around it. "Single sign-on is included on every paid plan" can be cited; "It's included from there on up" cannot.
Step by step: getting a B2B SaaS product cited by Claude
Step 1: Check access on every host
Read robots.txt on each host a buyer's question might reach, from www to docs and help, for groups that name Claude-SearchBot, Claude-User or ClaudeBot. Check the catch-all group too: under the robots.txt standard, RFC 9309, an agent no group names follows the User-agent: * rules. Then check the bot-protection layer, and search your logs for the three user-agent names, matching source addresses against bots.json.
Step 2: Make the answer pages readable
On the pages that answer buying questions (pricing, plan limits, integrations, security, comparisons), remove any noindex tag or login that has no business there, and make sure prices, limits and comparison tables are in the HTML the server returns rather than drawn by scripts or held in images. Requesting the page with a plain HTTP client such as curl shows what an agent receives.
Step 3: Write sentences Claude can lift
For each buying question, find the sentence on your site that answers it and make it short, specific and self-contained, naming the product rather than leaning on "it". Where no such sentence exists, that is the gap to fill first.
Step 4: Ask Claude your buyers' questions with search on
Take ten to twenty prompts from real sales conversations, from category picks and alternatives to pricing and integrations. Ask each with "search the web" in the prompt and record whether your product is named, where it appears and which pages are cited. On a paid plan, run one category question through Research and read its source list. Repeat the set on different days before acting on it.
Step 5: Work on the third-party pages Claude keeps citing
Review sites, roundups and community threads that keep coming up for your category are where Claude's picture of you forms. Keep listings complete and current, and give roundup authors accurate facts. Our guide to getting a SaaS product into the best-software lists covers the outreach, and how to get a SaaS product recommended by AI engines shows how the same sources feed the other engines.
Measuring what Claude does with your site
Server logs show the three agents separately. Claude-User requests are the most telling, since each means a person's question to Claude led to one of your pages; Claude-SearchBot visits show indexing and ClaudeBot visits the training crawl.
Analytics can show visits referred from claude.ai where the browser passes the referrer on. None of the Anthropic pages we read mentions a campaign tag on outbound links, so build the segment on the referring domain.
Prompt checks are the only view of answers that name you without linking to you, because none of the Anthropic pages we read describes a reporting tool for site owners like Google's Search Console. For a fuller method, with scoring against named rivals, see how to measure brand visibility in ChatGPT and AI answers.
How this guide was researched
Every statement here about Claude's behaviour comes from Anthropic's own pages, read on 17 October 2026: the help-centre articles on its crawlers (updated 7 April 2026), web search (updated 7 October 2026), Research (updated 14 October 2026), training cutoffs (updated 7 October 2026) and blocking and removing content (updated 3 September 2026); the developer documentation for the web search and web fetch tools and the models overview; and Anthropic's crawler address list. The comparisons with other engines come from OpenAI's and Perplexity's crawler documentation, read the same day. Anthropic does not publish how Claude weighs one candidate page against another, so this guide offers no list of ranking factors.
Questions B2B SaaS teams ask about Claude citations
Does blocking ClaudeBot keep a site out of Claude's answers?
Not out of answers that use web search. ClaudeBot gathers training material; search indexing is Claude-SearchBot's job and live fetching is Claude-User's. A ClaudeBot block does shape what future models learn, and Claude answers questions on stable topics from training.
Can Claude read a page that robots.txt blocks for Claude-User?
Not according to Anthropic. It says its agents honour robots.txt and that blocking Claude-User stops Claude retrieving your content for users' questions.
Do Claude users need a paid plan to get web search?
No. Anthropic's web search help includes advice for free accounts, whose searches count toward usage limits that reset every five hours. Research is for paid plans only.
Why did Claude answer a question about my product without searching?
Usually because the question looked like stable knowledge, which Claude answers from training, or because web search is switched off for the buyer's Team or Enterprise workspace. Writing "search the web" into the prompt asks for a search.
Can a company ask Anthropic to take its page out of Claude's answers?
Yes, for pages it owns. Besides noindex, password protection and robots.txt, Anthropic accepts requests to block a URL already appearing in Claude's outputs, with evidence of ownership such as a WHOIS record or an email from an address on the domain. That removes a page; it does not correct a description. To change what Claude says about your product, change the pages it cites.
If you would rather have Claude and the six other engines measured for you, The Rank Master pairs that measurement with a fixed amount of work each month: on the published plans, 4, 8 or 14 revenue pages and 2, 4 or 7 earned citations, from $2,000 a month, with custom programmes from $1,500. Nothing publishes until you approve it, citations are earned and never bought, and after the 90-day initial term the programme runs month-to-month on 30 days' notice. Book a 30-minute call.