How do you get Claude to recommend your brand?

The first source is training. Anthropic's Transparency Hub says its recent models were trained on “a proprietary mix of publicly available information” from the internet, alongside other sources such as public and private datasets and synthetic data. Its crawler for training is ClaudeBot. Anthropic publishes two cutoff dates for each current model: a training data cutoff, and a reliable knowledge cutoff, which it describes as the date through which the model's knowledge is most extensive and reliable. A model's training covers only what was published before its cutoff, on your site and on everyone else's.

The second is search. Anthropic says that when a question benefits from current information, Claude uses a search tool to ground its answer in content from the live web, with citations you can open. In what Anthropic calls the new Claude experience there is no switch for it: Claude searches the web when it helps. Elsewhere web search is a setting a person can turn on or off, and on a Team or Enterprise account an owner has to allow it for the whole organisation first. Anthropic's developer documentation for its web search tool is more specific: Claude searches when a request depends on information that is current, changing or outside its training data, such as current prices or an organisation or product that might have changed, and answers directly when a request draws on stable knowledge. So whether a buyer's answer came from a search or from what the model learned depends on the question and on how their Claude is set up.

Claude-SearchBot is the crawler Anthropic says improves the relevance and accuracy of Claude's search results, and Claude-User is the agent it says fetches a page when a person asks Claude something. Anthropic's list of subprocessors, the outside companies that handle data on its behalf, also names Brave Search for web search across all of its products, so what Claude's search can find is not limited to what Anthropic's own crawler collected.

For a brand, that makes the crawler setting and its own pages the quickest levers: search can pick up a change once it reads the page again, with no wait for a new model.

Access is the one Anthropic publishes instructions for, and they run through robots.txt: a rule naming the agent, in the file at the top of each subdomain it should cover. Each agent takes its own line, so refusing ClaudeBot leaves Claude-SearchBot and Claude-User as they were: Anthropic says a ClaudeBot refusal signals that a site's future material should be left out of its training datasets. It also says its agents will not try to get past a CAPTCHA, so a bot check at your network edge can keep them out of pages your robots.txt allows, and it publishes the IP addresses they come from, so a request claiming to be one of them can be checked. Anthropic's agents for search, for requests a person makes, and for training, and what Anthropic says each is for:

• Claude-SearchBot. Anthropic publishes it as the crawler that improves the quality of Claude's search results.
• Claude-User. Anthropic publishes it as the fetch made when a person asks Claude something that needs a page, and says all of its agents honour robots.txt.
• ClaudeBot. Anthropic publishes it as the crawler that collects web content to improve its models.

Which crawlers to allow for the other assistants, and why a rule at your network edge can refuse one your file allows, is its own answer.

Then the pages themselves, and here the advice is ours rather than Anthropic's. Buyers usually describe a situation rather than name a product: a small team, a budget, a named use. Answer that situation directly on the page, in a passage that still makes sense read on its own.

And the facts on them. Your own pages are among what Claude's search can read, and the requests Anthropic says send Claude searching include current prices and products that might have changed, so a wrong price, date or specification there can reach an answer. Where an answer and your page disagree, the fix can sit on either side. Which of the three to start with depends on where your brand is weakest, though a search crawler you refuse comes first whatever else is true, and that order has its own answer.

You can see which of those pages Claude reads in your own category. Anthropic says putting “Search the web” in a prompt makes sure Claude searches, and that a searched answer carries direct citations and source links. Ask a buying question that way and note the domains it points to. Where the same domains keep coming back on questions a competitor wins and you do not, those domains are where the work goes. Why a competitor comes up first has its own answer.

Those pages sit at different distances from you, and the work changes with the distance:

• A listing you supply. A retailer page, a directory profile, or a listing on an app people connect to Claude, such as Tripadvisor, Booking.com or Thumbtack, which Anthropic says Claude can suggest for what a person is doing. You can influence it directly by correcting what you supply, though the page is not yours.
• An independent review or comparison. You can give the writer accurate facts, access to people who can answer questions and, where it fits, the product to try. What they conclude is theirs.
• A forum or community thread. You can answer the question openly, under the brand's own name. The opinion belongs to the people in the thread.

None of it moves on your schedule. A page about you changes when the people who publish it decide to change it.

Anthropic set this out in February 2026: Claude will remain ad-free, its users will not see sponsored links beside their conversations, and its responses will not be influenced by advertisers or include third-party product placements users did not ask for. It repeated it in April 2026, when it added apps such as Booking.com and Instacart to Claude: there are no paid placements or sponsored answers, and where two connected apps could help, both are shown, ranked by what is most useful to the person asking.

Anthropic says it makes its money from enterprise contracts and paid subscriptions instead. So, by Anthropic's account, a competitor Claude recommends did not pay to be there, and the reason it comes up first is in what Claude reads and finds about it.

That is Anthropic's account of its own systems. It has said that if it ever revisits the approach it will be open about its reasons, so its announcement is the page to check.

Anthropic's developer glossary says identical inputs may produce different outputs across calls to its API. The variation is large: in Delphi's study of answers collected up to 24 August 2026, an assistant asked the same question more than once inside one measurement named the identical set of companies 21.5% of the time. A before and after built from one answer each compares two draws, and it can show a gain or a loss that nothing caused. How much evidence a change needs has its own answer.

The free check by hand is written for ChatGPT and works the same way in Claude: it sets out which questions to ask and what to write down for each answer, and comparing a before with an after adds to what you record each time. Four things are specific to doing it in Claude:

• Keep your own history out. Anthropic says Claude's memory is on by default on its Free, Pro and Max plans, and carries context such as your role and projects into new conversations. Start an incognito chat, outside any project: Anthropic says it does not use Claude's existing memory, though Claude can still read your profile preferences there, so check that those say nothing about your company.
• Choose whether it searches. Anthropic says putting “Search the web” in a prompt makes sure Claude searches, and that you can also tell it not to. Asking each question both ways separates what its search finds from what its model already holds, and the two change on different clocks.
• Ask from the same place. Anthropic says Claude may use your location, inferred from your IP address, for localised results. Ask the before and the after from the same place and network.
• Allow for the clock. A robots.txt change can only show up once Anthropic's agents read the file again, a rewritten page once Claude's search reads it again, a third-party page once its publisher changes it, and a change in what the model learned only once a model with a later cutoff replaces it.

Traffic shows only part of it. An answer that recommends you and sends nobody to your site leaves no visit at all.

The study publishes that figure with the counts behind it and how each was made. A brand's standing in Claude is one part of its AI visibility. Some of the work itself carries across and some does not. Accurate pages and third-party coverage are there for any assistant that reads the web. Crawler access is set operator by operator: most publish their own crawler, so allowing Claude-SearchBot says nothing about the rest.

The same question has its own answer for ChatGPT, Gemini, Perplexity and Grok.

Delphi measures them side by side. Each time it measures a brand, it puts the same questions to ChatGPT, Gemini, Claude, Perplexity and Grok more than once and keeps every answer word for word. Its score, the Delphi Index, carries a margin set by how far the assistants disagreed with each other on your questions, and the dashboard and the history page do not report a change in the Index smaller than that margin as a change. Each assistant's share of answers naming you is shown beside the others' from one measurement to the next, so a change in Claude alone is visible, though that figure carries no margin of its own.

Delphi asks each assistant through its API, the route developers use, rather than through the consumer app. Anthropic's developer documentation describes web search there as a tool a request adds, and most of Delphi's questions reach Claude without it. So for most questions a Delphi reading of Claude is its model answering from what it learned, and a page or crawler change shows in the answers Delphi collects first where an assistant searches as it answers, and elsewhere only once a newer model has learned it. Each paid level includes one or two measurements per brand a month. It is not a daily tracker, and an extra measurement costs $49 when a before and after needs one sooner.