Learn how Claude selects product recommendations, how Claude crawler access works, and which pages can improve your visibility in Claude answers.
The premise behind most of the worry about Claude visibility is that you can win a fixed spot in Claude's answer, the way you'd win a search ranking. There is no fixed spot to win: Claude decides whether your product fits the question, which evidence to use, and which products to name.
In CiteGraph's sample, Claude named the scanned product in 6% of 1,467 answers and cited 4.8 sources per answer on average, according to our methodology.
Start with a plain statement that gives Claude something usable. Say what the product is, who it is for, what it costs, and what it replaces or competes with. Put that statement on a page Claude can read without having to infer your business from slogans, navigation labels, or a product video. Founders sometimes call this Claude SEO, but it is really just writing a page that says what you sell.
Our readiness checks show the gap. Of the 121 sites we audited, only 20 passed the check for a plain statement of what the product is, for whom, at what price, and against what. By comparison, 119 passed the check that the homepage answers a plain request. These are two different checks: most sites answer a plain request, but few state plainly what they sell, for whom, and at what price.
A clear description is the useful starting point.
Use the page title and description to say what the product does. Keep the important facts in ordinary text rather than putting them only in JavaScript or an image. Add Organization schema where it reflects the business accurately, and Product or SoftwareApplication schema where it reflects the product. These signals support the sentence. They do not replace it.
The same audit found that 72 of 121 sites passed the Organization schema check, while 60 passed the Product or SoftwareApplication check.
Claude cannot use a page it cannot retrieve. Check robots.txt, server responses, access controls, and the rendered page. The page should contain its core answer without depending on a browser action or a script that fails during retrieval. Our readiness check found that 116 of 121 sites allowed AI crawlers in robots.txt, but only 92 had content present without JavaScript.
Robots.txt is one check; retrieval is another.
Anthropic explains the difference between its crawlers on its public crawler support documentation:
“When individuals ask questions to Claude, it may access websites using a Claude-User agent.”
Anthropic, in its crawler support documentation
That quote separates training access from user-directed retrieval. ClaudeBot helps collect public web content that may contribute to model training, while Claude-User can retrieve content in response to a person's question. Anthropic also says Claude-SearchBot navigates the web to improve search result quality. Blocking one route does not make your site invisible in every context, but blocking user-directed retrieval can reduce visibility for a query.
Anthropic says disabling Claude-User can reduce your site's visibility for user-directed web search, and disabling Claude-SearchBot can reduce its visibility and accuracy in user search results. For a founder, the practical job is to inspect the rules that govern the page you want Claude to use, then confirm that the response contains the text a crawler needs.
A sitemap and canonical tags help machines locate the right version of a page. In the 121-site audit, 100 sites passed the sitemap check, while 92 passed the canonical and Open Graph check.
Build pages that answer the buying question Claude is trying to answer. Product pages establish facts. Comparison pages explain the trade-offs. Alternatives pages help when the buyer has already named a competitor. A clear pricing page gives Claude a concrete fact to use. The page needs a point of view that can survive being quoted in a short answer.
The citation mix tells you where supporting evidence often lives. Listicles and roundups account for 67.5% of the citations recorded in our research, while competitor product pages account for 24.4%. Together they take 91.9% of citations. The scanned product's own site accounts for 1.5%.
Your own page is necessary. It is rarely sufficient.
The useful target is a set of pages around the product, rather than one heroic homepage. A roundup can frame the category. A comparison can make the differences explicit. Your own page can supply the exact price, limits, use case, and product facts that those pages need. Community threads account for 3.4% of citations in the sample, so outside discussion exists, but it is not the main path shown here.
An alternatives page can matter even when it does not mention you first. It gives Claude a structured way to explain when your product fits and when another product may fit better. Honest boundaries are useful evidence. They also make the recommendation less likely to read like an unsupported sales claim.
Write for the decision inside the prompt. State the buyer, the job, the constraints, the price, and the meaningful trade-offs. Use headings that match the questions a buyer asks, then answer each one directly. If your product is for a small team, say so. If a feature belongs to a higher plan, say so. If a competitor is stronger for a certain use case, say that too.
Pricing is a good test of whether the page is usable. In the readiness audit, 90 of 121 sites passed because a price was visible or a pricing page was linked. The remaining sites did not necessarily have bad pricing, but the route to it was less clear. Claude has more usable evidence when the page gives it a concrete fact instead of a vague promise.
A buyer should not need a treasure hunt.
Your title and description should say what the page is about. A single clear H1 helps the reader and the retrieval system orient themselves. Questions answered on the page also matter: 56 of 121 audited sites passed that check, while 63 received a warning.
Our position has an obvious weakness. A page can pass every technical check and still be absent from Claude's answer because the assistant has better evidence elsewhere, the prompt does not fit the product, or the product is not relevant to the request. Readiness improves access and clarity, but selection still depends on the question, competing pages, retrieved evidence, and the model's judgement. Anyone offering a guaranteed mention is selling certainty that the data cannot provide.
Guarantees have no place in this measurement.
Test prompts that represent real buying decisions, rather than only your brand name. Include category questions, comparison questions, alternatives questions, and use cases where your product should or should not fit. Record whether Claude names the product, which pages it cites, whether the wording is confident or qualified, and whether the cited page actually supports the claim.
You can use the free AI visibility checker to inspect whether your pages are appearing in tracked answers. The free AI readiness check covers basic access and clarity problems, while our weekly research on who owns AI answers shows the broader page-type pattern. These tools measure different parts of the problem, so keep them separate in your reporting.
Compare Claude with other engines carefully. Across CiteGraph's per-engine samples, the scanned product appeared in 8.2% of 1,787 Perplexity answers, 6.8% of 1,784 Gemini answers, 4.4% of 1,781 ChatGPT-family answers, and 6% of 1,467 Claude answers. The Claude sample is smaller than the samples for the other three engines, so treat the comparison as directional. Average source counts were 5.3 for Perplexity, 4 for Gemini, 3.4 for ChatGPT-family, and 4.8 for Claude.
Claude names the product less often than Perplexity or Gemini, and slightly more often than ChatGPT-family. The sample does not tell us why, but a mention among fewer named products may carry more weight. Measure the mention and the supporting citation separately, because a brand can appear without its own page doing the explanatory work.
The caveat that matters most is simple: CiteGraph measures samples, we print the margins, and we do not know what we have not measured. A 6% naming rate is evidence about the sampled Claude answers, not a promise that Claude will recommend your product next time.
It can retrieve website content when the relevant access path is available and the page fits the question. Anthropic says Claude-User may access websites when people ask Claude questions, but retrieval does not guarantee a citation or recommendation.
Blocking ClaudeBot may exclude future materials from Anthropic's model training datasets, according to Anthropic's crawler guidance. ClaudeBot is different from Claude-User and Claude-SearchBot, so check which access you intend to allow before changing robots.txt.
A page with a plain product statement, clear pricing, use cases, and specific comparisons gives Claude material it can use. Our sample also shows that listicles, roundups, and competitor product pages make up most recorded citations.
No. Clear pages and available crawlers improve readiness, but Claude still chooses based on the question, competing evidence, relevance, and its own judgement.
Cite this: CiteGraph, “How to get recommended by Claude”, 11 Sept 2026, https://www.citegraph.app/blog/how-to-get-recommended-by-claude