Your buyers ask Claude which vendors to shortlist, and it names other companies. How to get cited by Claude in 2026 depends on three things. Claude searches the web only for some prompts, mostly comparisons and current facts, and answers the rest from memory without sources. When it searches, it cites from Brave Search's top results for short queries it writes itself. From each page it keeps the one passage that matches the exact words of those queries. Being cited, which means your URL appears in the sources list, is a different outcome from being named or recommended in the answer. A brand is often recommended through a third-party ranking list in Brave's top results while its own site is never cited.
Statcounter counted Claude at 5.7% of US AI chatbot use in October 2026. At BrightonSEO that month, Josh Blyskal of Profound argued that the web share understates Claude's use in business, citing Ramp's company spending data.
For a UK or EU B2B seller, measure ChatGPT and Google's AI answers first and add Claude third. To tell within a week whether your buyers use Claude, ask each inbound lead which assistant they asked. Also look for claude.ai among your referral sources.
Most guides to this question cover robots.txt, structured data and "write clear answers". In my view they miss an earlier step, and that is where B2B companies fail. The page is missing from Brave's index, the list of pages Brave Search knows and can show. Or it ranks far from the top for the queries Claude writes.
At Agenzy we have analysed 1,000,000+ AI chats across 15,000+ tracked prompts. Three studies of Claude's citations, linked below, show the pattern. Brave's index decides which pages Claude can cite, and the passage decides which of them Claude quotes.
This article explains how Claude chooses its sources and how a page enters Brave's index and ranks there. It then covers how to write the passage Claude quotes and how to measure it.
Two limits apply. We do not track Claude's answers to our own prompts with a tool. The evidence here is Brave's index and rankings, Anthropic's documentation, Brave's browser code and the three studies. Our indexing test had no pages left out for comparison.
How Claude chooses its sources
Claude cites a page in 2026 when that page sits in Brave Search's top results for the short queries Claude writes. The page must also hold one passage that matches those queries word for word.
Anthropic's documentation says Claude decides per prompt and searches when a question depends on information that may have changed, such as facts about specific organizations, products, prices or statistics. A single fact takes one to three searches, and a comparison across several companies can take ten or more. Without a search, Claude shows no sources, so nothing is cited.
When Claude does search, it rewrites the prompt into several short queries, called fan-out queries. It adds words the user never typed, such as the year and the audience. Those queries go to Brave Search, the provider Anthropic names for web search on its subprocessor list. Most of what Claude cites comes from Brave's top results.
A Profound and MERJ analysis of about 35,000 Claude citations over 400 queries was published on MERJ's blog in July 2026. It found 79.2% of the cited URLs in Brave's top 10 for the same query, compared with 34% in Google's top 10. Peec AI's October 2026 study of 8,500 URLs found 89.1% in Brave's top 20 for Claude's own queries.
Claude then reads about an eighth of each page, roughly 1,900 of 14,600 characters. Which 1,900 characters Claude keeps is predicted best by BM25, a keyword formula that rewards rare exact words. In Peec's test, BM25 pointed at the passage Claude actually quoted 41.6% of the time, compared with 34.1% for always taking the top of the page.
The chain from prompt to citation, from Anthropic's documentation and the Profound, MERJ and Peec AI studies.
This chain of search, Brave ranking and passage choice decides which page gets cited. Which company gets named depends on whether those pages name it, which I come to under ranking in Brave.
Overlap of Claude's citations with each engine's results, Profound and MERJ (July 2026) and Peec AI (October 2026).
Sidenote: Anthropic's list also names TurboPuffer and Parallel Web Systems for web search. What they do inside Claude's search is not public.
Claude compared with ChatGPT
ChatGPT reads three indexes, Bing, Google and an index of its own, as our explainer on whether ChatGPT uses Bing or Google sets out. Claude reads one, Brave Search. So a page missing from Brave is invisible to Claude in a way that no single miss is for ChatGPT.
Two more things differ in practice. First, Claude searches on fewer prompts. In three Peec AI projects we tracked in October and November 2026, ChatGPT searched the web on 83% to 91.5% of chats. So Claude can only cite a company on prompts where the buyer asks for a shortlist: which vendors to consider, product comparisons, prices in 2026.
The second difference is how often Claude names a company at all. For a B2B hardware manufacturer we track, Claude named the company in 15% of its answers in September 2026. ChatGPT named it in 45% of its answers to the same prompts (the report example we published).
Share of answers naming the company, Agenzy tracking data in Peec AI, September 2026.
How a page enters Brave's index
Brave has no tool like Google Search Console for submitting pages and does not accept sitemaps. Its crawler, the program that collects pages for the index, learns about pages partly from the Web Discovery Project. That is an opt-in feature of the Brave browser that reports what its users search and visit.
We read the Web Discovery Project's code in Brave's public repository, and it has two routes into the index.
The search route needs one user. When a Brave user with Web Discovery on searches Google or Bing, the browser downloads the results page a second time, anonymously. It reports the query and the result URLs to Brave. One user is enough, so Brave learns about a page that ranks on Google or Bing as soon as one such user runs that query.
The visit route needs 20 users. A visited page is reported in encrypted form. Brave can only read the URL once 20 different users on different networks have sent it. The number is fixed in the browser's code, THRESHOLD = 20 in modules/hpnv2/sources/star.es.
The two routes and the checks, read from Brave's Web Discovery Project code.
Before a visit counts, the browser downloads the page a second time with no cookies and compares it with what the user saw. The rules, read from doublefetch.es and doublefetch-handler.es:
Return HTTP 200, the normal success code, with no redirect.
Answer within 10 seconds with a page size under 2 MB.
Serve the content in HTML, because no JavaScript runs during the comparison.
Keep the word "noindex" out of the page's head section entirely, including inside a script or a meta tag aimed at another bot. The X-Robots-Tag header is not read.
A page that fails any of these checks is dropped and never reported to Brave. Brave's crawler help page adds three things. Its bot has no distinct user agent, meaning it does not identify itself by its own name when it requests a page. It will not crawl a page Googlebot cannot crawl. And robots.txt is not used to prevent indexing.
There is a submit form at search.brave.com/submit-url. According to MERJ's reading of Brave's documentation, its only documented function is to fetch again a URL Brave already knows.
Our Brave indexing test
On 8 November 2026 we ran the second-download rules as a script over every page of two of our sites, agenzy.lt and webtrafficnews.com. The second had been live for eleven days. Every page passed: status 200 with no redirect, a response under a second, a page size under 2 MB, full HTML and no "noindex" in the head.
Pages of both sites were still missing from Brave's index, because nothing had told Brave they existed. Passing the technical checks allows Brave to keep a page once it finds it, but does not make Brave find it.
We then submitted 105 URLs by hand through the form across the afternoon. The home page of agenzy.lt, which Brave already held, went first at about 15:45. That submission's second fetch replaced Brave's stored copy, which still showed a figure we had already changed on the live page. Claude quotes the copy Brave has stored.
The first 16 missing pages, 11 from agenzy.lt and 5 from webtrafficnews.com, went in at about 16:00. All 16 were in Brave's index by about 17:50. The other 88 went in at about 18:00. A recount at 19:00 showed none of them yet, and at 19:45 every one was in. Each batch was indexed within about two hours of its submission.
The afternoon of 8 November 2026, from our submission log and index checks.
We took two other actions the same afternoon, so the form on its own was never isolated. Between about 16:10 and 16:25 we submitted every agenzy.lt URL to Bing through its webmaster tool.
We also opened the Google and Bing site: results for both sites, which list a site's pages, in a Brave browser with Web Discovery on. That is the one-user search route described above. No page was left unsubmitted for comparison.
Those two actions happened hours before the second batch, and that batch appeared only after its own submission. So the form is the most likely cause. That refutes the claim that Brave has no submit button and makes "submitting does nothing" unlikely, though the form alone was never isolated.
Getting into Brave's index took hours in our test and cost nothing. Reaching Brave's top results for the fan-out queries is the slow and hard part. Brave positions also move: search specialists report rankings that hold for a few weeks and then drop (GSC Wizard).
Ranking in Brave for fan-out queries
A company gets recommended by Claude through a page in Brave's top ten results for Claude's queries. That page is either its own or a third-party ranking list that names it. So a place in Brave's top ten is what you need to win and keep, and two kinds of page can take those places.
The first is your own pages. A page in Brave's top ten for a query is in the set Claude cites from. One that is not there gains nothing from changes to its text until it gets there.
The second is the pages that already rank there: ranking lists, comparisons and buyer's guides. In the same talk, Josh Blyskal put evaluative content, listicles and comparison pages at "60-something percent" of all Claude citations. In Brave's top 20 for "GEO agency" on 8 November 2026, 15 of the 20 results were ranking lists and guides. Four were agency home pages and one was a Reddit thread.
Brave's top twenty for "GEO agency" by page type, 8 November 2026.
To get onto those pages, run Claude's queries for your buyer prompts on search.brave.com and open the top ten results. Note who publishes each list and what every entry carries: what the company does, who it serves, its price range, a measurable result. Where the publisher takes listing requests, send those facts for your company. It costs nothing but the hour.
One of our own inbound leads shows the route. A Baltic software company asked Claude for AI visibility agencies in the Baltics (the report example we published). Two of Brave's top ten results for that query are third-party list pages that include us. We do not know which pages Claude read for that buyer or on which day. Today's Brave results stand in for the ones Claude saw.
Brave's first page for the buyer's query on 8 November 2026, and one inbound lead.
Writing the passage Claude quotes
BM25, which predicts which passage Claude keeps, gives almost nothing for repeating a word. So each page you want cited needs one passage of 1,500 to 2,000 characters that does two things. It carries the words a fan-out query will use: the year, "B2B", the market, the product category. And it makes sense when cut out of the page, because Claude never sees the heading above it.
The company's name and what it does are written out, with no "it", "this tool" or "our platform".
A definition sentence as many pages write it, then the same sentence carrying the words a fan-out query uses:
Before: "GEO, short for generative engine optimization, is the work of getting a business named and cited when someone asks an AI assistant such as ChatGPT, Gemini or Google AI Overviews for a recommendation."
After: "GEO, short for generative engine optimization, is the work of getting a B2B company named and cited when a buyer asks ChatGPT, Claude, Gemini, Perplexity or Google AI Overviews which vendors to shortlist in 2026."
This article's two versions with the fan-out words marked.
Sidenote: in the studies we have, structured data and tables do not raise the chance of being cited. Ahrefs compared 1,885 pages that added JSON-LD, the usual format for structured data, with 4,000 similar pages that did not. AI Overview citations moved by -4.6% and ChatGPT citations by +2.2%, and the second change is too small to be sure it is real.
Programs that read a page as plain text find a table-bound answer 5% to 24% of the time, compared with 39% to 44% in a paragraph (PixelRAG).
How to measure Claude visibility
You can check all of this in an afternoon without a Claude tracker, with a Claude account as the only cost before step five.
Run your twenty buyer prompts in Claude with web search on and mark which ones made Claude search. An answer with no sources is a memory answer.
Write down the queries Claude ran for each searched prompt. Claude shows them above the answer, and those are your fan-out queries.
Search each query on search.brave.com and note whether your page, or a page that names you, is in the top ten.
Search
site:yourdomain.comon search.brave.com, compare the result with your sitemap, and submit every missing page at search.brave.com/submit-url. Run the second-download checks on pages that stay missing.Add a Claude channel to a prompt tracker where the budget allows. In Peec AI, Claude is a paid upgrade. Our comparison of AI visibility tools lists which trackers cover Claude, and our guide to prompt tracking explains the prompt set.
Repeat steps one to four monthly, because Brave positions move over weeks.
For clients we run the Brave checks as a script on Brave's paid search API, with the prompt set and the page work, as part of our GEO service.
FAQ
How do you get cited by Claude?
Claude cites the page that Brave Search ranks near the top for the queries it writes and that answers them in their own words. That happens only on a prompt where Claude searches at all. Our guide to ranking on ChatGPT in 2026 covers the same work for ChatGPT.
Should I block ClaudeBot?
ClaudeBot collects content for training. Blocking it does not affect Claude-SearchBot or Claude-User, the two bots Claude's web search uses (Anthropic's bot page). Our AI crawlers list has the robots.txt lines for all three.
Does robots.txt affect whether Claude cites my page?
Anthropic's bots honor robots.txt, so blocking Claude-SearchBot or Claude-User can reduce your visibility in Claude's web search. Brave's crawler will not crawl a page Googlebot cannot crawl, and robots.txt does not stop Brave indexing a page it learns about from users.
Does structured data help get cited by Claude?
Ahrefs' controlled test found no increase in AI citations from adding JSON-LD. Put the facts in prose first, because Claude chooses the passage by the words in the text.
Does Claude use Google or Bing at all?
Claude's web results come from Brave Search. Claude's help center says Bing powers its image results and names no provider for web results. A page that ranks on Google or Bing still enters Brave's index through the search route.
Does Claude read content that loads with JavaScript?
Treat content that loads through JavaScript as invisible. Vercel's log analysis, from December 2024 and so two years old, found that Anthropic's crawlers did not run JavaScript. Brave's second download runs none either. Put the passage you want quoted in the HTML the server sends.
Does Brave drop pages from its results, and what do I do?
Search specialists report Brave rankings that appear for a few weeks and then drop, and Brave has not documented why (GSC Wizard). Judge your presence over weeks, re-submit a changed page at search.brave.com/submit-url, and track the buyer prompts with our guide to prompt tracking.
Does buying Brave Search Ads change whether Claude cites me?
Profound's test excluded sponsored results, and no measurement shows Claude citing them. Spend the effort on the organic position instead.
How can I see which passage of my page Claude used?
Each citation Claude returns shows up to 150 characters of the passage it quoted, which Anthropic's web search documentation specifies. Open a citation to your page and check whether those characters come from the passage you wrote.
Does my page's language or country matter on Brave?
Anthropic's web search tool takes the user's city, region, country and timezone, so results are localized (Anthropic's documentation). Write the words that describe your market, in your buyer's language, into the passage. Then search the query with the market word on search.brave.com and check that your page is in the top ten.
How is getting cited by Claude different from Perplexity?
Perplexity runs its own index and crawler, searches on every prompt, and searches the buyer's question as typed. Claude searches on some prompts, through Brave Search, with rewritten queries. Our guide to ranking in Perplexity covers the rest.
Final thoughts
Three things decide whether Claude cites you. It has to search. Brave has to rank a page about you near the top for its queries, and that page has to answer in the buyer's words.
This week, run steps one to four on your twenty buyer prompts and write down which prompts Claude searched and where Brave put you.
If something here does not match what you see in Claude, tell me on LinkedIn.
Further reading: our guide to ranking in Perplexity, which sources AI engines cite in 2026 and what LLM SEO is.
