Buyers are asking AI assistants which software to buy. I wanted to know whether that changes who gets recommended, so I measured it instead of guessing.
Nine B2B SaaS companies across five categories — project management, product analytics, help desk, knowledge base, CRM. For each one I generated five decision-stage buying questions, ran them through an AI assistant with live web search enabled, and recorded every product named in the answer.
The spread was much wider than I expected.
| Visibility | Company |
|---|---|
| 100% | Missive — shared inbox (named in every answer) |
| 80% | Linear — project management |
| 60% | a Gmail-based CRM |
| 40% | an open-source analytics tool |
| 20% | a knowledge base tool |
| 0% | four companies, across four different categories |
I've named the two strongest performers and anonymised the rest. Publishing a list of companies that scored badly seemed like a hostile thing to do to businesses that never asked to be measured.
Four of nine were not named a single time across five buying questions in their own category. Not ranked low. Absent.
There's very little middle. Established companies with years of accumulated content and third-party mentions score high. Newer challengers — including ones with real products, real customers, and well-optimised comparison pages — score zero.
This is not a gentle ranking penalty. In traditional search a challenger sits on page two. In an AI answer there is no page two. You are named or you are not, and four of these nine simply were not.
One company — an open-source analytics tool — had its own website cited as a source in 5 of 5 answers, and was recommended in 2.
The assistant was reading their documentation to learn about the category, then recommending competitors.
Their content is doing the work of educating the model, and someone else is collecting the recommendation. I've started calling this the citation–recommendation gap, and it showed up repeatedly:
Being the page AI reads is not the same as being the answer AI gives. Those are different problems and probably need different fixes.
Reasonable objection, so I tested it. I re-ran one company at 25 questions instead of 5. Visibility was 0% both times — same result, five times the sample.
That's one company, not a proof. But it's evidence the small sample isn't just noise, and it's cheap enough that anyone can replicate it.
Open an AI assistant and ask the question your best customer would ask right before choosing someone in your category. Not "what is a CRM" — something like "best CRM for a 10-person sales team that lives in Gmail."
Count the companies it names. Check whether you're one of them.
That's a 30-second version of what I did, and for four of these nine companies the answer would have been genuinely unpleasant.