How Claude Chooses Sources to Cite

Claude, Anthropic's AI assistant, can search the web in some products and configurations and cite the pages it used to build an answer. For SaaS companies, that raises a practical question about citations in Claude answers: what makes a page more likely to be found, read and cited when a buyer asks a question about your category?

It is important to be honest about the limits of what anyone outside Anthropic knows. Anthropic has not published a ranking algorithm for which sources Claude cites, and behavior can vary by product, settings and model version. What follows is a set of reasonable inferences from how retrieval-based AI answers generally work, combined with practices that are good for any search channel. It is not a formula, and you should be skeptical of anyone who presents one.

What is generally understood

Citations come from search results

When web search is enabled, Claude retrieves pages through a search step and then reads relevant content from those pages to answer the question. A page that does not come back from that search step cannot be cited, however good it is. This means general discoverability still matters a great deal. If your page does not rank or is not indexed for the relevant queries, it is unlikely to be part of the material Claude considers.

This has a useful implication for small teams. Much of the work that helps you appear in AI answers is the same work that helps you appear in traditional search: pages that are indexed, that target the questions buyers actually ask and that are linked from elsewhere on your site and the web. AI visibility builds on SEO rather than replacing it.

Content must be accessible

If a page blocks crawlers, hides its content behind a login, or renders key text only through heavy client-side scripts, it is harder for any AI system to read. A page that looks fine in a browser may present an almost empty document to a crawler that does not execute JavaScript, or may be blocked entirely by a firewall rule the marketing team does not know exists.

Anthropic publishes information about its crawlers and how site owners can control their access through robots.txt. The user agent names and their purposes can change, so check the current documentation rather than relying on a list copied from an old blog post. Some sites block AI crawlers by default through their CDN or security settings, which can prevent pages from being read even when the robots.txt file allows it.

Relevance to the question matters

Answers draw on pages that address the specific question being asked. A focused page that directly explains how to calculate CAC payback is far more useful to an AI system answering that question than a general marketing page that mentions CAC in passing. This mirrors how good search results work: the page that best matches the intent behind the question tends to be the one that gets used.

For SaaS companies, this argues for pages built around specific buyer questions rather than broad pages trying to cover everything. A product page that tries to rank for every feature is less likely to be the best answer to any individual question than a set of focused pages, each covering one topic properly.

Clear, specific information is easier to use

Pages that state facts plainly give an AI system concrete material to summarize and attribute. Clear definitions, explicit steps, stated prices and direct statements about which features or integrations are supported are all easy to use. Pages written in vague marketing language, full of claims like "powerful" and "seamless" without saying what the product actually does, give the system little to work with. If an AI system cannot extract a clear fact from your page, it has little reason to cite it.

What is not publicly known

Several important things remain unknown from the outside. Exactly how sources are weighted or ranked once they have been retrieved has not been published. Whether specific technical signals, such as structured data, directly influence whether a page is cited is not documented, even if those signals help with search visibility more broadly. How often behavior changes between versions is also not something you can rely on, and an approach that seems to work today may behave differently after an update.

Given that uncertainty, be wary of anyone promising a guaranteed method to get cited by any specific AI assistant. Claims of that kind usually rest on a handful of observations that may not generalize, and tactics designed to exploit a supposed ranking quirk tend to stop working, or backfire, when the system changes.

Practical steps for SaaS sites

Make sure you can be crawled

Start by reviewing your robots.txt file and any firewall, CDN or bot-management settings that might block AI crawlers. Decide deliberately which crawlers you want to allow, rather than inheriting a default from your hosting provider. Some companies have good reasons to block certain crawlers, and that is a legitimate choice, but it should be a conscious one made with an understanding of the trade-off.

Then check that your important content is served in the initial HTML of the page. A quick test is to view the page source, or fetch the page with a simple command-line tool, and confirm that the main text appears. If your product pages, pricing or documentation load their content only after scripts run, consider server-side rendering or static generation for those pages.

Answer questions directly

Put a clear, direct answer near the top of each page, then provide depth below it. If the page is about how to calculate CAC payback, the first paragraph should state the formula and what it means before going into nuance and examples. Use headings that match how people actually phrase their questions, because those headings help both readers and AI systems find the relevant section quickly.

This does not mean writing for machines at the expense of people. A page that answers the question clearly in its opening is also a better page for the human buyer who arrived from search and wants the answer without scrolling.

Be specific and accurate

State plainly what your product does, who it is for, what it integrates with and how pricing works. Many SaaS sites are surprisingly vague about these basics, sometimes deliberately, which leaves AI systems to fill the gaps from third-party sources that may be wrong or outdated. If you do not publish pricing, explain how pricing works in general terms, such as whether it is per seat or usage-based.

Keep this information current. When you add an integration, change a plan or retire a feature, update every page that mentions it, including help articles and comparison pages that are easy to forget.

Publish genuinely useful content

Guides with clear steps, worked examples, templates and honest comparisons make better sources than thin pages. If you want to be cited when buyers ask how to solve a problem, publish the most useful explanation of that problem you can, written by someone who understands it. Content that simply restates what is already on many other pages gives an AI system no reason to choose yours.

Build credibility beyond your site

AI answers often draw on information from many sources, and your own site is only one of them. Reviews on reputable platforms, coverage in respected publications, detailed public documentation and genuine discussion in communities all help establish what your product is and does. When independent sources describe your product consistently and accurately, AI systems have a clearer picture to draw on, and your own pages are more likely to be treated as credible when they appear in results.

Keep content fresh

Outdated pricing, discontinued features or old screenshots can lead to inaccurate answers, and an inaccurate answer about your product can be worse than no answer at all. Review important pages on a regular schedule, update facts that have changed and note the date of meaningful updates where it helps readers.

How to check

The most direct check is to ask Claude, with web search enabled where it is available, the questions your buyers actually ask. Note which sources appear in the citations, whether your pages are among them and how your product is described. Repeat the same set of questions periodically so you can compare results over time.

You can also track referral traffic from claude.ai in your analytics, which shows some of the visits that come from cited links. It will not capture everyone who saw your brand in an answer without clicking. Adding AI assistants as an option in a "How did you hear about us?" field on your signup or demo form helps fill that gap by capturing buyers who discovered you through an AI answer and then came to your site directly.

Results vary between runs, even for the same question, so avoid drawing conclusions from a single test. Look for patterns across many questions and several weeks.

A worked example (illustrative)

Consider a small company selling subscription billing software. Its team asks Claude a set of buyer questions, such as how to handle proration on plan upgrades, and notices that its own pages never appear among the citations. On inspection, its proration guide loads its main text through a client-side script, its pricing page says only "contact us" and its firewall blocks unfamiliar crawlers by default. The team serves the guide as static HTML, adds a plain explanation of how pricing is structured, reviews the bot settings deliberately and rewrites the guide's opening to state the proration formula directly. Weeks later, the same questions begin to return its pages more often in tests. Nothing about that change relied on a secret signal; it removed obstacles and made the content easier to use.

What to avoid

Avoid creating thin pages aimed only at AI crawlers, such as near-duplicate pages targeting slight variations of a question. They offer nothing to readers and are unlikely to be treated as strong sources. Avoid stuffing pages with your brand name or keywords in an attempt to strengthen associations, which makes content worse for people and does not reliably help with machines. Avoid misleading comparisons that misstate competitors' features or pricing, which can damage trust and create legal risk. And avoid manipulating reviews or communities, which violates platform rules and tends to be discovered.

The takeaway

There is no published formula for how Claude chooses sources, so focus on fundamentals that hold up regardless of the details. Make sure your pages can be crawled and read, answer specific questions clearly, keep product facts accurate and current, and earn credibility across the web. Those steps help with AI assistants generally and with traditional search at the same time.