GEO
Claude searches the web through Brave. Gemini leans hard on YouTube. Perplexity is the most citation dense of the four. Where each engine gets its sources changes what you should do, and it is why we score every engine separately instead of blending an average.
People talk about "AI search" as though it were one thing. It is not. ChatGPT, Perplexity, Gemini and Claude are four products with four different supply chains, and the supply chain is where your visibility is actually decided.
An AI answer is assembled in two stages. First the engine goes and fetches some pages. Then it writes an answer using what it fetched. Most advice concentrates on the second stage. The first stage is the one you can influence, and it works differently in each engine.
Here is a plain-English map of the back doors.
Claude's web search is served through Brave Search, an independent index built by the Brave browser company rather than a rebadge of Google or Bing.
The alignment is unusually tight. Profound compared Claude's citations against Brave's results and found 86.7% overlap between what Claude cited and Brave's top non-sponsored organic results. In the same comparison, ChatGPT aligned with Bing's top results only 26.7% of the time.
Read that twice, because the practical consequence is odd and specific. Your ranking in Brave Search, an index most small businesses have never looked at, quietly decides a large part of your Claude visibility. Meanwhile ChatGPT's relationship with Bing is loose enough that a Bing ranking tells you much less than you would assume.
What to do with that: search your own money question in Brave and see where you land. It is free, it takes a minute, and for one of the four engines it is close to a direct readout. Brave's index is smaller than Google's, which cuts both ways. Less competition, but also a real chance you are simply not in it.
Ahrefs studied 75,000 brands and measured which signals move together with visibility in AI answers. The strongest single correlate was YouTube mentions at roughly 0.737, ahead of branded web mentions at 0.66 to 0.71 and far ahead of backlinks at 0.218.
A correlation is a measure of how tightly two things rise together, on a scale where 1 is lockstep and 0 is no relationship. A YouTube mention here means your brand name appearing in a video title, description or transcript. It does not have to be your video.
The mechanism is not mysterious. Transcripts are clean, dense, plainly worded text about who does what, and Google owns the archive. If you were building a system that had to describe businesses, you would use them too.
What to do with that: you do not need a channel. You need to be named in videos other people already make. A trade roundup, a local tour, a supplier's walkthrough, an interview. One transcript that says your name next to your category and your town is worth more than a video you make and nobody watches.
Of the four, Perplexity is the most citation dense in the answers we read. It names more sources per answer and puts them in front of the user rather than behind a fold. Community threads, Reddit especially, turn up in its citation lists more often than in the others'.
We are describing what we see, not a measured study, so treat it as an observation. But the direction is consistent enough to act on: if your buyers use Perplexity, the forum where they already ask questions is part of your visibility surface, and answering honestly in that forum is a legitimate move. Posting fake praise under names that are not yours is not, and moderators remove it faster than it earns anything.
ChatGPT uses Bing as a backend, and yet the 26.7% alignment figure above says the backend ranking is a weak predictor of what gets cited. Something sits between the fetch and the answer, and none of us outside OpenAI can see it.
That is worth stating plainly rather than papering over. ChatGPT is the engine where "do this specific thing and you will appear" advice is least trustworthy, and where the general work matters most: consistent facts about you, on pages that are widely read, in plain text.
Now put the four together. Claude tracks a Brave ranking. Gemini tracks a YouTube footprint. Perplexity tracks a citation-dense pool that includes forums. ChatGPT tracks something we cannot cleanly describe.
Average those four and you get a number that hides every actionable thing in it. A business scoring 90 in Perplexity and 10 in Claude has the same blended average as one scoring 50 everywhere, and the two need completely different afternoons of work.
This is why we score per engine and publish the four scores separately, even though a single number would look tidier in a dashboard. The tidier number would be less true. It is the same reason we insist that being named is not the same as being recommended, which we set out in three ways a brand appears in an AI answer, and only one counts.
Per-engine scoring also tells you something a blend cannot: which back door is shut. A low Claude score with a healthy Gemini score is a Brave indexing problem, not a content problem. You would never see that in an average.
Two things are shared, and they come first.
All four need to be able to read your site. If your server blocks their crawlers, a crawler being the automated fetcher that collects pages, the supply chain never starts. That is a five-minute check and it is the cheapest fix in this whole field. See is your site blocking ChatGPT, and note that the Cloudflare change on 15 September puts a lot of sites into this bucket without their owners doing anything.
All four also need something clear to read once they arrive. Same plain sentence, same consistent facts, on pages that render as text. We have covered what AI search engines actually read on your site, and the honest maths on llms.txt for anyone tempted by the exotic fixes first.
Everything above is an observed pattern, not a permanent law.
Search backends get swapped. Perplexity has run on more than one index. Anthropic, OpenAI and Google are all free to change suppliers, weighting or retrieval logic without announcing it, and correlation studies describe a moment in time rather than a mechanism anybody has confirmed from the inside.
So use this map the way you would use a map of roadworks. Useful today, worth re-checking in a quarter, and never a substitute for looking at the actual answers your buyers are getting.
Which is the part you can do right now. Run a free scan at whocanfindme.com. No signup. Takes about ten seconds.
When a buyer asks an AI one question, the engine quietly turns it into about ten. Seer Interactive measured 10.7 hidden sub-queries per prompt across 501 prompts on Gemini 3. Most sites answer the headline question and none of the ten. Here is how to find your gaps.
Read moreThe same question, same engine, same day can produce different answers. That is the nature of AI search, not a bug. The 3 rules we use to avoid showing you fake statistics, and how to read a visibility score without fooling yourself.
Read more