If you run SEO for a B2B SaaS client, the monthly review probably has a slide now about whether AI assistants recommend them. The usual way to fill it is to type the category into ChatGPT and screenshot whatever comes back.
I ran two categories properly this week to show what that screenshot hides. Same method both times. Six to eight neutral buying questions per category, no product names in the prompts, one prompt one engine one call, production default settings, five assistants, samples spread across four weeks, and every answer scored for whether a tool was recommended, listed in a comparison, or mentioned in passing.
63 sampled answers, four weeks to 9 September. ChatGPT recommended Clay most, in 11 of its 14 answers. Claude recommended Apollo AI most, in 10 of 14. Perplexity recommended ZoomInfo most, in 13 of 14. Gemini had ZoomInfo and Apollo AI level. Grok, on 7 answers, had three tools level.
Same question, and the first answer depends on which assistant the buyer happens to open. A screenshot from one of them is one draw from that distribution, presented as the distribution.
The mention trap is in there too. Cognism was named in 84% of the answers and was the first recommendation in none of them. Listed in 21, recommended in 21, never first. A screenshot shows it present. The full sample shows it losing.
63 sampled answers, same window. Zendesk is the tool the assistants recommend most, in 45 of the 63 answers. Of the three tools they recommend most, it is also the one they put first least often. First name in 11 answers. Intercom and its Fin agent, counted together, take the first slot in 25.
So the category has two leaders depending on which column the slide reads. Read the recommendation column and Zendesk leads. Read the first-name column and Intercom does. The assistants split as well. Claude recommends Intercom most, ChatGPT and Perplexity recommend Zendesk most, Gemini has five tools level on 14 answers, and Grok, on 7, has Zendesk and Intercom level.
Four things, each with a sample size and a date next to it.
Recommendation share, not mentions. The share of answers in which the client was recommended, with the number of answers it came from. Named in 84% and recommended first in none describe the same product.
First-name rate. How often the client is the first tool the assistant reaches for. It moves independently of recommendation share, and it is the number closest to the click.
The split by assistant. One number across five assistants hides that ChatGPT and Claude can have different winners, and the client's buyers are not spread evenly across them either.
The limit. 14 answers per assistant is enough to show a split and not enough to rank the middle of the field. Say so on the slide. A client who hears the limit from you trusts the rest.
If you fill that slide for a client, tell me the category. I will run it the same way and publish what it says, whatever it says. The full rankings with confidence intervals are on the index pages for B2B lead generation tools and AI customer support agents, and the methodology is public, including the parts that do not flatter it.
Submit your SaaS to curated directories, track proof, and manage launch visibility from one dashboard.