What Perplexity actually says about ATS vendors
20 prompts. 2 vendors. 1 engine. The pattern was not what the marketing pages assume.
An ATS vendor usually has two things: a category page, and a set of comparison pages. Lever, for example, publishes /alternative/best-ats-for-mid-sized-companies, /alternative/best-ats-with-crm-functionality, /alternative/lever-vs-greenhouse and /alternative/lever-vs-ashby.
So we asked an AI assistant exactly those questions — plus six more buyers type — and watched what came back.
Method
What we ran, exactly
Engine: Perplexity, default model, logged-in free account.
Prompts: 10 per vendor, written before the run and not edited afterwards.
Runs: 1 per prompt (a pilot, not a census).
Dates: Ashby on 2026-10-21, Lever on 2026-10-22.
Judgement: mention = the brand appears in the answer text; citation = the answer links the brand's own domain.
Result 1: Lever appeared in every answer, and was cited in none
| Metric | Lever | Ashby |
|---|---|---|
| Answers naming the brand | 10 / 10 | 8 / 10 |
| Answers naming it when the prompt never did | 7 / 7 | 5 / 7 |
| Answers linking the brand's own domain | 0 / 10 | 0 / 10 |
| Answers also naming Greenhouse | 10 / 10 | 7 / 10 |
| Answers also naming Ashby | 10 / 10 | — |
Both vendors are extremely well known to the model. Neither is ever the link.
Mentions are free. Citations are the scarce thing. A mention is a brand impression. A citation is a click, a visit, and a page the model will lean on next month.
Result 2: the pages already exist, and the model passes them over
This is the part worth sitting with. Lever has a page for “best ATS for mid-sized companies”. We asked that exact question. Lever was named. The page was not cited.
The same is true for the CRM question and both head-to-head prompts. The content exists. It is not being selected.
That distinction matters for anyone deciding what to do next. “Write more content” is the wrong prescription when the content is already there. The question is which sources the model quotes — and how quotable those sources are.
What we would fix first
- Make the page quotable, not just complete. A dated, direct answer with one clear claim beats a feature list of the same length.
- Earn the third-party mention. The model cites pages it can lift a sentence from. If those pages are review sites and analyst write-ups, the brand description is written by someone else.
- Re-measure on identical prompts. Any change is only worth what the delta says. Same prompts, same engine, same day of week.
Limits, stated plainly
One run per prompt is a sample, not a census; model answers vary by run, date and context. Only Perplexity was tested here — ChatGPT, Gemini and Copilot were not. The prompt list, dates and judgement rules are published above so anyone can re-run it and check our numbers. A report you cannot reproduce is worth nothing.
Want this for your category?
We run ten prompts your buyers actually type, across the engines that matter, and send a five-page report: where you appear, who appears instead, and which sources the model quotes. Free, 48 hours.