Mostly, Claude cites what Brave Search ranks — not what Google ranks, not what ChatGPT sees. That single fact changes how you should think about optimizing for it. Anthropic’s own support documentation confirms Claude’s search and Research features run on live results “powered by Brave Search,” and independent analysis from Profound found an 86.7% overlap between the URLs Claude actually cites and Brave’s top organic results. Compare that to ChatGPT, where the same analysis found only 26.7% overlap with Bing’s top results, and the picture becomes clear: Claude’s citation behavior is unusually transparent and unusually dependent on one search index most businesses have never optimized for.
We started digging into this after a client asked a blunt question in a strategy call: “we rank fine in Google, why does ChatGPT know about us and Claude doesn’t?” The honest answer is that these are three separate visibility problems wearing one “GEO” label, and Claude’s is the least understood of the three.
How does Claude actually decide what to cite?
Claude runs a three-step evaluation before it even searches: it weighs freshness (how current the information needs to be), specificity (how precise the query is), and intent (what the user is actually trying to accomplish). If it decides a search is warranted, it queries Brave, then filters what comes back on relevance, clarity, and extractability — and it can repeat that search-and-filter cycle up to ten times in a single conversation if the first pass doesn’t produce a confident answer.
The extractability part is where most small business sites lose out before Claude even gets to judging the content itself. Claude’s more recent models apply dynamic filtering that strips navigation, footers, ads, and boilerplate before processing a page — which means a site built on heavy client-side rendering, where the actual content only appears after JavaScript executes, is close to invisible. Claude’s crawler doesn’t execute client-side JavaScript, full stop. If your page’s real content is rendered client-side, Claude is reading an empty shell.
Why Brave, and why does that matter more than optimizing for Google?
Brave’s own index has a quirk worth knowing: a page generally needs organic visits from at least 20 unique Brave browser users with data-sharing enabled before it’s eligible for indexing at all. That’s a low bar in absolute terms, but it means brand-new pages and very low-traffic niche sites can sit outside Brave’s index entirely — and therefore outside Claude’s citation pool — regardless of how well-written they are. Getting a page some initial traffic through other channels (search, social, a newsletter link) isn’t just good practice generally; for Claude visibility specifically, it can be the difference between eligible and invisible.
What does Claude actually cite once it finds your page?
A separate analysis of over 2,000 URLs Claude cited turned up patterns worth building a content plan around. 56% of citations came from /blog/ paths rather than homepages or product pages — only 3% of citations were homepages. Nearly half (47%) used listicle-style URLs — “best,” “top-10,” “vs,” “alternatives” patterns. And two categories were almost entirely absent: major news outlets like Forbes, TechCrunch, and Bloomberg, and social/community platforms like Reddit, LinkedIn, and Medium got essentially zero citations. If your GEO strategy leans on getting mentioned in press coverage or building a presence on Reddit, that effort may pay off for ChatGPT or Google’s AI Overviews — the data suggests it does very little for Claude specifically.
Content structure that measurably helps
- Lead with the direct answer. Put the actual answer to the implied question in the first sentence or two of a section, before the caveats and context. This mirrors the guidance in our post on the answer-engine optimisation checklist, and it matters even more here because Claude’s extraction step is explicitly looking for clean, quotable statements.
- Keep paragraphs tight — roughly 40–60 words. Dense, unbroken paragraphs are harder for the extraction step to pull a clean citation from.
- Use real semantic HTML. <article>, <main>, and proper H2/H3 structure aren’t just an SEO nicety here — they’re literally what Claude’s dynamic filtering uses to separate your content from your navigation and footer.
- Show a visible freshness signal. A genuine “last updated” date and a year token in the URL or title correlate with a documented 2–3x higher citation rate for content published or updated within the last 12 months.
- Link out to 3–5 authoritative sources per page. Pages that cite their own sources appear to read as more trustworthy to the extraction and filtering step — the same “show your work” logic that helps human readers trust a page.
Should you block or allow Claude’s crawlers?
Allow them, in almost every case. Claude’s user-triggered browsing bot (Claude-User) still honors robots.txt, which is worth noting because ChatGPT’s equivalent bot stopped doing so as of December 2025 — a meaningful divergence in how these two companies are choosing to play the access question. There’s also no announced publisher licensing deal on Anthropic’s side, which cuts both ways: no revenue-share opportunity yet, but also no structural disadvantage against sites that haven’t signed one. Blocking Claude’s crawlers doesn’t protect anything commercially valuable for a typical small business — it just removes you from a citation pool you were trying to get into in the first place. If you’ve already done the work in our crawlability audit, Claude access should already be handled; if you haven’t run that audit, it’s the first thing to check, not the content structure tips above.
Does an llms.txt file help specifically for Claude?
Not in any way the current evidence supports. We covered this in depth in our llms.txt piece, and the mechanism described above reinforces the same conclusion: Claude’s citation pipeline runs through Brave’s index and its own extraction filtering, not through a manifest file your site publishes voluntarily. The budget is better spent on the semantic HTML and freshness signals above than on maintaining a file with no confirmed effect on Claude’s behavior.
How this compares to optimizing for other AI engines
If you’ve read our piece on AI browser agents researching your business, you already know Claude, Comet, and Atlas each move through the web differently. Claude’s dependency on one search index is genuinely unusual — Google’s AI Overviews draw on Google’s own index, and Microsoft Copilot Search leans on Bing, a comparison we walked through in our Copilot Search piece. That single-index dependency is actually good news for optimization: there’s one visibility gate to clear instead of several, and once a page is extractable, well-structured, and has enough real traffic to clear Brave’s indexing threshold, it’s genuinely competing on content quality rather than on domain authority or publisher deals.
What we’d actually do first
If Claude visibility specifically is the goal, we’d check three things in this order: whether your key pages render their content server-side (not behind client-side JavaScript), whether those pages are getting enough real traffic to clear Brave’s indexing bar, and whether your existing blog content leads with direct, quotable answers before the context and caveats. All three are free to audit and none of them require new tooling — they require rewriting the first two sentences of your best pages and checking your rendering setup, which is a Tuesday afternoon, not a redesign.