Most AEO advice treats "AI search" as one thing. Get cited by ChatGPT, the thinking goes, and Perplexity and Claude will follow. A July 2026 study by Searchable, covering roughly 95,000 search records across 3,000 unique Claude prompts, shows why that assumption is expensive. Claude cites sources differently to ChatGPT at almost every stage - which index it draws from, which source types it favours, even how often it bothers searching the live web at all.
Two different questions need two different plans
Claude answers some questions by reading the web right now, and others from what it already learned during training. Which one happens depends heavily on the question, not just the topic.
Searchable tracked how often specific prompts triggered a live search. "Best CRM 2026" pulled a live source 9 times in 10. "Cheapest CRM" pulled one 7 times in 10. But "how to pick a CRM" triggered a live search only once in 34 attempts, and "Zendesk vs Intercom" only once in 7. The pattern holds across freshness and comparison prompts generally: anything tied to price, ranking, or the current year sends Claude looking for a live page. Anything closer to general advice, Claude usually answers from memory.
That's two different jobs. Winning a retrieval-led prompt means winning a live search result - the tactics further down this article. Winning a memory-led prompt means becoming part of what Claude already knows, which happens through sustained third-party coverage over time, not a single optimised page. Write down the questions your buyers actually ask, sort them into the two groups, and the content plan splits naturally along the same line.
The index behind Claude's citations is probably Brave, not Google
Searchable compared the pages Claude actually read against the same day's top ten results on Brave Search and on Google. The overlap with Brave ran to 63%. The overlap with Google was 34%. Searchable is careful about what this does and doesn't prove - Anthropic hasn't named a search provider, and the study frames this as inferred index behaviour, not a confirmed integration.
It's worth checking that inference independently rather than taking it on trust, so I did. TechCrunch reported in March 2025 that Anthropic had quietly added Brave Search to its subprocessor list the day after switching on Claude's web search, that Claude's search function carries a parameter named BraveSearchParams, and that at least one identical query returned matching citations in both Claude and Brave. None of that is Searchable's claim - it's independent reporting from a year before this study, and it points the same direction.
The practical upshot: ranking on Google doesn't tell you much about whether Claude can find you. Search Brave for your own domain. If nothing comes up, submit your homepage through Brave's URL submission process, and check that nothing in your robots.txt or firewall settings is quietly blocking crawlers from getting in. Sites that have never thought about Brave visibility are, by definition, not thinking about the index Claude is most likely reading from.
Once you're in the top ten, rank barely matters
Inside Brave's top ten, Searchable found citation rates barely moved by position: 39.4% for ranks 1 to 3, 37.2% for ranks 4 to 6, 36.8% for ranks 7 to 10. Being read at all functions like a gate - pages outside the top ten were rarely read in the first place. Once through that gate, the difference between rank 2 and rank 9 is close to nothing.
That changes where the effort should go. A page stuck on Brave's second page is worth far more attention than a page already sitting at rank 3 and being polished toward rank 1. The first move is the one with a payoff. The second mostly isn't.
Claude leans on third-party coverage more than ChatGPT does
Searchable's source-type breakdown shows a consistent pattern: Claude favours external, editorially-produced content over brand-owned pages, more than ChatGPT does. National press appeared in 52% of Claude's cited sources against 19.2% of ChatGPT's. Review sites came in at 49.1% versus 44.8%. Smaller comparison blogs sat at 42.6% versus 28.8%. A brand's own domain made up 31.6% of Claude's answers, against 60.9% for ChatGPT.
Reddit and forum content sit at the opposite end for both - 8.3% of Claude's cited sources and 12.7% of ChatGPT's. That's worth flagging against other AEO research already out there: Semrush's wider study of AI citations found Reddit as the single most-cited domain across AI search as a whole. The two measure different things: Semrush looked at domain-level citation share across a broad general dataset, and this study measures citation rate for a specific set of commercial, buyer-intent prompts. Reddit can be a major citation source overall and still barely feature in Claude's answers to "best CRM" or "cheapest applicant tracking system" specifically.
One more detail worth keeping in the plan: no single publisher supplied more than 9% of Claude's cited sources in this sample. Landing one big press hit isn't the strategy. A spread of smaller, credible, third-party mentions across press, review sites, and comparison publishers does more than concentrating on one.
Local search barely touches the open web at all
For location-based prompts, Claude behaves differently again. 97.7% of local prompts triggered a live search, but Claude read an average of just 0.1 conventional web pages per query - it's pulling from a business listings service, not crawling websites. For any business with a physical location, the entire game here is keeping opening hours, categories, descriptions, and address data accurate on that listing. A well-optimised website barely factors in.
A few common follow-up questions
Does this mean a ChatGPT-focused AEO strategy is wasted on Claude?
Not wasted, but incomplete. The fundamentals - clean structure and real evidence on the page - carry across every model. What changes is where to point the extra effort: Brave visibility and third-party press for Claude, owned-domain depth for ChatGPT.
How do we check whether Claude is actually reading our site?
Start with the Brave search test in this article - search Brave for your own domain and your target queries, and see where you land. Then run your own buyer questions through Claude directly and note which ones trigger a visible source versus an answer with no citation at all.
Should we stop investing in our own website content, given how much Claude favours third-party sources?
No. Owned content is still what most other channels, including Google and ChatGPT, draw on most heavily. The finding here is about where the next pound of effort goes for Claude specifically, toward earning press and editorial mentions, not a reason to stop publishing on your own site.