Claude SEO: be a source Claude can reach and trust

Anthropic's support article on web crawling describes three bots (the article shows no date; we read it on 15 October 2026). As of that reading, ClaudeBot collects web content that could contribute to training its generative AI models. Restricting it signals that your future materials should be excluded from training datasets.

Claude-User supports Claude users. It may access websites when a person asks Claude a question. Anthropic says disabling it blocks retrieval of your content for user queries and may reduce your visibility in user directed web search.

Claude-SearchBot navigates the web to improve search result quality. Anthropic says disabling it stops indexing of your content for search and may reduce your visibility and accuracy in user search results. In short, these two are the ones that bear on being cited. ClaudeBot is a training question and a separate decision.

• Claude-SearchBot is allowed if you want to be indexed for search
• Claude-User is allowed if you want live retrieval for user questions
• ClaudeBot is a deliberate training choice, not a visibility lever
• Each subdomain has its own robots.txt that you have checked

Claude is often used to summarize, compare and explain. That kind of task works best on pages with clear statements, labeled sections and visible sources. A page that says what you do, for whom, where and at what level of evidence gives a summarizer something solid to work with.

We cannot claim this is how Claude scores sources, because Anthropic has not published that. It is our working judgment from testing, and we label it that way. What we do rely on is the documented part: if the right crawler is blocked, Claude cannot retrieve your page for a search.

We also treat other places where your brand is described as part of the job. A model that answers from memory draws on what the wider web says about you. Consistent facts on your site and in third party descriptions lower the risk of wrong answers. If something false is already out there, our guide to fixing wrong information about your brand in AI shows the approach.

Claude is often asked to compare options, explain trade offs and summarize long material. Pages that suit that use include comparison tables with stated criteria, step by step procedures, definitions that state scope and limits, and policy pages that say plainly what is and is not covered. Marketing copy with no specifics is the weakest input.

Where claims need support, give the source and the date. If you did the work, say which work and for whom, in a way that can be checked. If you cannot share names, describe the situation honestly rather than inventing detail. See original data and statistics for AI citations for how to publish your own data in a way that holds up.

Keep documents navigable. Labeled sections, a clear title and a short summary at the top help any reader, human or machine. Link to your fuller pages from the summary, so the detailed version is easy to find. This is the base of GEO content strategy.

A model can answer from what it learned in training or from content fetched at the time of the question. Anthropic's documentation separates the bots by those jobs: ClaudeBot for training data, Claude-User for user requests, Claude-SearchBot for search quality. That is the documented part.

What follows is our reasoning, not a published rule. For recent changes, such as a new product or price, memory is likely out of date, so only a live fetch can be right. For long standing facts, memory may be enough, and then what the wider web said about you earlier matters. This is why we work on both: crawler access for live fetches, and consistent descriptions across the web for memory.

It also explains why we log whether web access was used in each test. A wrong answer without sources points to memory and to third party descriptions. A wrong answer with sources points to a page you can fix. Our post on how AI search engines choose sources describes the general pattern.

Example, hypothetical: a UK accounting practice asks Claude who the best advisers are for small business tax in London, and Claude describes the firm as a payroll bureau. The audit finds that a wildcard rule blocks Claude-SearchBot, the service pages call everything "financial solutions", and an old directory lists payroll as the main service.

We would split the robots rules so that training stays blocked if the firm wants that, while Claude-SearchBot and Claude-User are allowed. We would rewrite the services page with plain names for each service, who it is for and what is not offered. We would correct the directory entry and align Organization data. Then we would retest.

If the description improves in some runs and not others, we report both. Read GEO for professional services and GEO agency in the UK for the context.

Our Claude report lists each prompt, the run dates, whether web access appeared to be used, whether your brand was named, whether a page of yours was linked and whether the description was accurate. We keep it by prompt so you can read the evidence yourself.

We do not turn this into a score or a rank, because Claude has no stable ranking to report. We report presence and accuracy across a fixed set, before and after. If you want a deeper method, see how to get cited by Claude and our GEO reporting template.

Claude traffic is often small in analytics, because many answers do not send a click. Treat visits as one signal and named mentions as another. For professional and technical buyers, a correct mention in a research answer can matter long before a visit shows up.

Claude is one of five engines in the Pilot baseline. The fixes overlap strongly with the other four engines. The free AI visibility checker is a quick first look.

The GEO Pilot is a fixed $999 (USD), one time price, delivered inside 30 days, done for you. It covers one domain and one language. A site build, ongoing digital PR, link building and paid media are not included, and we make no promise of mentions. See the GEO Pilot page for the full scope.