Everyone talks about "getting recommended by AI" as if AI is one thing. It is not. I put the same 40 buying questions to ChatGPT, Gemini and Perplexity on the same day, read every answer, and counted which products each one named. The three do not agree, and the size of the disagreement surprised me.
Here is what came back.
Across the 39 questions where all three answered, they named 437 distinct products between them. Of those:
- Only 30% were named by all three.
- 47% were named by exactly one of the three.
Read that again. Nearly half the products that showed up in an AI buying answer were named by a single assistant and invisible in the other two.
What it means if you only check ChatGPT
A single assistant shows you between 58% and 65% of the products the full panel named, depending on which one you check. So if you look only at ChatGPT and see your rival there, that is one third of the picture. Your buyer over on Gemini or Perplexity may be getting a completely different shortlist, and you have no idea whether you are on it.
Is it the model, or what it read?
This was the part I wanted to know. When I handed ChatGPT and Gemini the identical set of search results and asked them to answer from those, they agreed 64% of the time, far more than they do in the wild. So most of the disagreement is not the models reasoning differently. It is that each assistant retrieved different pages to begin with.
That is good news, because retrieval is the part you can influence. The assistants disagree because they read different sources, and the sources are pages you can be on. Change what they retrieve and you change what they recommend.
The takeaway
"Am I recommended by AI" is not one question, it is three, and they have different answers. If you measure your AI visibility on one assistant, you are measuring a third of it. Check all three, read the source pages behind each answer, and work on being in the sources that the assistants you are missing actually read.
I build a tool that runs exactly this: three assistants on one shared set of buying questions, with the sources shown, so I run it on my own category and publish the numbers. The measurements above are real, from 40 questions on 19 July 2026. If you want your own three-way picture, the check is free at glotier.com. But the method is right here, and you can reproduce it with a browser and a spreadsheet.
Top comments (1)
This is the measurement people need more of. The interesting part is not only which model recommends which product, but which evidence each system appears to reward. If the same page is visible to one answer engine and invisible to another, the practical SEO question becomes much more like data-quality debugging.