Every business we work with is asking some version of the same question. How do we show up when someone asks ChatGPT or Claude who to hire? Almost every answer floating around is a guess, because almost all AI visibility work starts from citations, and a citation only tells you what the model chose to show. It says nothing about what the model actually retrieved and passed over.
For Claude, we found a way to stop guessing. Over ten weeks this summer we measured where Claude's web search gets its sources, on 1,481 matched queries. The answer is Brave. 87.8% of the URLs Claude's search tool retrieved were sitting in Brave's top 10 for the same query. A Google control on identical queries came in at 45.6%.
Anthropic doesn't say in the product where Claude's search results come from. Brave surfaced as the likely provider in 2025 reporting after it appeared on Anthropic's subprocessor list. We wanted our own measurement, at scale, on real buyer-intent queries, and we wanted to keep measuring so we'd notice if it ever changed.
Why this is measurable at all
When Claude answers with its web search tool, the API returns more than the answer. Alongside the text it returns the exact query strings Claude wrote for itself and the exact URLs its retrieval brought back, before any selection happened. Both halves of the system are visible. There's no reverse-engineering from citations and no guessing what the model saw.
That changes the kind of claim you can make. We take Claude's own reformulated query, fire the identical string at Brave's search API and at Google organic, and compare the three URL sets. Every query is matched one to one. Nothing is inferred.
The control matters more than the headline
Any two search engines overlap. If we only told you Claude's sources appear in Brave, you should ask how often they appear in Google too. So every query also ran against Google organic results as a floor. Brave held 87.8% of Claude's sources in its top 10. Google held 45.6%. On 1,455 of the 1,481 queries, 98.2%, Claude's result set sat closer to Brave than to Google. That separation held steady across three full monthly sweeps of the account base, June through August, with no significant drift between sweeps.
The matched queries come out of a much larger corpus we run continuously for clients, 44,635 AI answer runs across six engines since late May, 5,397 of them Claude with search enabled. The 87.8% rests on the 1,481 matched queries, where all three engines returned comparable result sets. The larger corpus is context for scale, and the two numbers shouldn't be conflated.
Brave rank predicts whether Claude cites you
The retrieval finding is interesting. The citation finding is the one you can act on. For every tracked prompt, we also recorded where the client ranked in Brave for that query and whether Claude's grounded answer ended up citing them. The gradient is steep.
| Brave rank | Cited by Claude |
|---|---|
| 1–3 | 92.0% |
| 4–10 | 69.8% |
| 11–20 | 38.4% |
| Not in the top 20 | 11.9% |
Measured across 1,408 prompt observations on the same 52 accounts. Businesses ranking 1 to 3 in Brave got cited in 92% of Claude's grounded answers. Businesses absent from Brave's top 20 got cited in 12%. The question of how to get recommended by Claude has a concrete answer. Rank in Brave. And Brave isn't an exotic target. The work that earns rank there is the same work that earns it in Google.
Two different problems that look the same
Once you can see retrieval and citation separately, a Claude visibility miss splits into two different diagnoses with two different fixes. A business absent from Brave has a ranking problem, and the fix is ordinary SEO. A business that ranks in Brave but still doesn't get cited has a selection problem, where Claude sees the page and passes on it, and the fix is content and entity work.
Across our client base that split runs 218 ranking gaps to 56 selection gaps, and 38 of 55 accounts skew ranking-first. Most businesses asking how to get into Claude have a plain ranking problem wearing a scary new costume. That's genuinely good news, because plain ranking problems are the ones the industry already knows how to fix.
The measurement almost fooled us once
Between the June and July sweeps, our primary overlap score dropped sharply, which is exactly what a swapped retrieval backend would look like. It turned out Claude had started returning fewer sources per query while Brave and Google held flat, and the metric we were using divides by the size of both sets combined, so it sagged on set-size alone. A containment measure, the share of Claude's sources present in Brave's top 10, held completely steady over the same window. We now watch containment, and we publish the number that survived its own false alarm. If a vendor shows you an overlap stat, it's worth asking which kind.
This is the first in a series of studies from our AI visibility tracking work. We measure how real businesses actually appear across Claude, ChatGPT, Gemini, Perplexity and the rest, every month, against real buyer-intent prompts. If you want to know where your business stands before your competitors figure this out, talk to us.
