I keep the books for a small SaaS company, and I also sell a tool that finds unused seats. So the question I care about is boring: when someone asks a model "what should we use to find wasted SaaS spend", which names actually come back?
Not whether the answer is good. Which names.
Method
- 40 buying questions, written the way a finance or ops lead would type them.
- Four models through one API: glm-4.5, glm-4.7, glm-5.1, glm-5.3.
- Every answer stored verbatim, then counted: which product names appear, in how many answers, and per model.
- 160 attempts (40 x 4). 151 produced usable text. The other 9 stay in the denominator and nowhere else: 7 stopped at the token ceiling while the model was still reasoning, 2 were refused by a rate limit. Neither was retried into the count.
Average answer length: 4,664 characters. Nothing was trimmed before counting.
What came back
| Name | Type | Answers naming it (of 151) |
|---|---|---|
| Slack | general tool | 74 |
| Okta | general tool | 65 |
| Vendr | SaaS spend vendor | 64 |
| Zylo | SaaS spend vendor | 63 |
| Ramp | finance tool | 63 |
| Microsoft | general tool | 60 |
| Torii | SaaS spend vendor | 57 |
| Cledara | SaaS spend vendor | 55 |
| Productiv | SaaS spend vendor | 52 |
| Brex | finance tool | 52 |
Two things I did not expect
1. General tools outrank category products. Slack, Okta and Microsoft are not spend-management products, yet they outrank every one of them. When the question is about spend, the model reaches for line items people already pay for before it reaches for a category product. That is a naming habit, not a comparison.
2. The category vendors cluster tightly. Vendr 64, Zylo 63, Ramp 63, Torii 57, Cledara 55, Productiv 52, Brex 52. Within four points of each other. That pattern usually means there is no default in the niche yet: the model has seen these names together often enough to list them, but nothing tells it that one of them is the answer.
And one number that is not flattering: the tool I sell was named in 0 of 151 answers. Not ranked low. Absent. Which is the honest output of this experiment for me: a model can only name what has already been written about, and being useful is not the same as being citable.
Limits
- One provider, one family of models, one run. Another provider has different retrieval and different defaults.
- 40 questions is a sample, not a census. A different phrasing moves the counts.
- Name matching counts a mention, not an endorsement. "Slack" can appear as an example of a bill you already have.
- I wrote the questions and the tool that counts them, so read the framing with that in mind.
Where the numbers live
I published the question set, the per-model counts and the misses on the site, no email gate: https://www.saasclerk.app/research/ai-citation-2026-11
If you run the same 40 questions against other providers, I would like the counts. That is the version of this post worth reading.
Top comments (0)