We build Bersyn, a tool that tracks which products AI models name when someone asks for a recommendation in a category. So we run a lot of scans. This week we ran one that surprised us, and it is worth showing the receipts.
The question we tested is the one a real founder types into ChatGPT: "what is the best platform to host and deploy a B2B SaaS app, and what are the strong alternatives?" We asked it five different ways across four Surfaces: ChatGPT, Claude, Gemini and Perplexity. Then we measured how often each modern platform actually got named.
What the models said
Here is the Recommendation Share for three platforms developers actually talk about, measured across the four Surfaces.
ChatGPT Claude Perplexity Gemini
Railway 0% 80% 0% 20%
Render 20% 80% 20% 20%
Fly.io 0% 60% 0% 0%
Read the first and third columns again. ChatGPT recommended Railway in zero of its answers. Perplexity recommended Railway in zero of its answers. Fly.io was named in zero answers on three of the four models. These are not low scores. They are absences.
So who got named instead? AWS, every time, leading almost every answer. And the named platform-as-a-service incumbent was Heroku. Here is ChatGPT, verbatim, on the best place to deploy a B2B SaaS app:
Choosing the best application hosting and deployment platform for a B2B SaaS team depends on several factors... Here are some popular options: 1. Amazon Web Services (AWS): AWS is highly scalable, reliable, and offers a wide range of services...
A buyer reading that answer in 2026 walks away with AWS and Heroku. Railway, Render and Fly, the tools a lot of teams genuinely ship on now, were never in the room.
Claude is the outlier, and that is the interesting part
One model broke the pattern. Claude named Railway in 80% of its answers, Render in 80%, Fly in 60%. It knows the modern category. The other three do not, or barely do.
That gap has a name in our world: Model Disagreement. When one Surface knows a company and three do not, the company is not "low visibility." It is invisible on the Surfaces most buyers use, and visible on the one they use least. If you are Railway, a buyer who happens to ask Claude sees you, and three out of four buyers never do.
Why this should bother a founder, not just amuse them
The reflex is to laugh it off. AI is behind, models have a training cut, it will catch up. Maybe. But your buyer is asking the question today, and the answer they get today does not include you.
Search had twenty years to learn that Railway exists. AI recommendation answers are being formed right now, off whatever evidence the models can find, and for a lot of newer companies that evidence is thin. The result is the table above: the incumbent gets named by reflex, the challenger gets Omitted.
The four ways a company shows up wrong in these answers, in our vocabulary:
- Omitted. The model lists competitors and skips you. This is what happened to Railway and Fly on ChatGPT and Perplexity.
- Misclassified. The model puts you in the wrong category.
- Generic. The model mentions you so vaguely no buyer could shortlist you. This is what Claude did, technically naming the platforms but describing them blandly.
- Confused. The model conflates you with a similarly named competitor.
What actually moves it
We do not claim AI is biased or that anyone is paying for placement. We just read what the models say and show you the evidence. What changes the answer over time is the same thing that changed search: published, specific, verifiable evidence that associates your company with the category and the buyer question. Comparison pages that name the alternatives honestly. Documentation that states plainly what you are and who you are for. Third party mentions in the exact words a buyer would use.
None of that is fast. But the first step is not writing more content. It is finding out what the models say about you right now, so you know whether you are Omitted, Generic or Confused, because the fix is different for each.
See your own category
We built Bersyn to show you exactly the table above, for your company, with the verbatim answers behind it. You can run a free scan on your own product at bersyn.com and see which Surfaces name you, which name a competitor in your place, and why.
If you ship on Railway, Render or Fly, tell me in the comments which model gets your stack right. I read all four daily and the disagreement between them is the most interesting thing I look at.
Top comments (0)