TL;DR
- ChatGPT knows these four companies well and almost never brings them up on its own.
- Their websites are healthy, fully crawlable, and already carry the comparison pages the standard advice asks for.
- Everything the YC launch playbook produces was cited three times in 120 answers.
Ask a founder how an AI-native product beats incumbent software and you get a version of the same answer. The old tool is slow, the new one is better, buyers will work it out.
We tested whether an AI already knows that. Four YC-backed recruiting startups from the W24, F24 and W25 batches. Sixty real buyer questions, each asked twice on ChatGPT with web search on.
It named HireVue. It named Paradox, Greenhouse, Lever and iCIMS. It almost never named the startups.
The engine knows them. It just never brings them up.
These are not unknown brands. Contrario is at roughly $6M annualised revenue with more than 150 placements in under six months. Ask ChatGPT about it directly and it comes back first, and well reviewed, in both runs. Ask the question a buyer would actually type, and it is gone.
That gap is two different problems, not one. A company nobody has heard of fails both halves. These four pass the first perfectly and fail the second almost completely, which means the knowledge is there and the recommendation is not.
Three real questions from the study, with the companies ChatGPT actually named:
- “What is the best AI recruiting platform for high-volume frontline hiring?” Paradox, Workday, iCIMS Frontline AI, Workstream, VidCruiter, Humanly. Lightscreen AI absent in both runs.
- “What is the best combined applicant tracking system and CRM for recruitment agencies?” Bullhorn, JobAdder, JobDiva, Recruit CRM, Recruiterflow. Spott absent in both runs.
- “Contrario vs hireEZ: which is better for startup hiring?” Contrario named first, both runs.
So who is ChatGPT recommending?
Every company here sells against hiring software that is slow, manual and stuck in 2012. Ask for the best AI recruiting platform and that is close to the entire answer. Not one of the names below is AI-native.
HireVue13 / 55
Paradox13 / 55
Greenhouse11 / 55
Lever8 / 55
iCIMS7 / 55Counted on the 55 questions that named none of the four, so every row shares one denominator. Answers name several tools at once, so these do not add up to 55.
We spent a day blaming their websites
The obvious suspect is the site. Blocked crawlers, a page that only renders in JavaScript, thin content. So we crawled all four the way an AI crawler reads them, JavaScript off and robots.txt checked per bot.
Every one of them came back healthy. Not a single AI crawler blocked anywhere.
| Company | Batch | Content score | Pages read | Crawlers blocked | Named, blind |
|---|---|---|---|---|---|
| Spott | W25 | 92 | 80 | none | 1 / 15 |
| Alex | W24 | 82 | 79 | none | 1 / 15 |
| Lightscreen AI | F24 | 74 | 15 | none | 0 / 15 |
| Contrario | W25 | 72 | 36 | none | 0 / 10 |
Spott scored 92 out of 100 with text extractability above 85 on every page, and it was named in one blind question out of fifteen.
They already wrote the comparison pages
Then we found the part that actually stings. They had already done the thing everyone tells challengers to do. Spott has comparison pages against Bullhorn, Loxo, Crelate, JobAdder and Clockwork Recruiting. Alex has more than sixteen of them.
Across twenty blind comparison questions, the four were named zero times.
The sources are on other people’s domains
When we stopped looking at the four companies and looked at what ChatGPT was citing, the pattern was immediate. Nearly half of every answer in this category traces back to one review site. None of the four have a profile on it.
G245.0%hirevue.com15.0%paradox.ai12.5%Share of all 120 answers citing that source, counted once per answer, with g2.com and learn.g2.com together.
Three of the four also publish no pricing at all, so an engine has nothing to answer one of the first questions a buyer asks.
In search, your website was the asset. In AI answers it is barely a source. You cannot write your way onto a domain you do not own.
That is why sixteen comparison pages produced nothing. A comparison page on your own domain is a sales argument. When an engine assembles a neutral comparison, it wants a source that is not you, and for these four no such source exists yet.
The YC part
A YC company starts with distribution most startups do not get: a Launch YC page, a Hacker News thread, a directory listing, usually some press. All of it lives on exactly the kind of high-authority third-party domain that should show up here.
None of it did.
TechCrunch, Forbes, Built In1 each
ycombinator.com0
Hacker News0
Product Hunt0
Crunchbase0Answers citing each source, out of 120.
Three citations, in total, from everything the launch playbook produces. The one review site in the chart above out-cited all of it by a factor of eighteen.
Launch coverage tells people you exist. Review sites tell an answer engine what to recommend. A YC batch buys you the first and none of the second.
So the honest read is not that these four did nothing. They ran the playbook they were handed, and the playbook does not touch the surface their buyers are now using.
What we would do
Your launch got you coverage. Coverage is not what gets cited. Get a review profile up before writing another page. It is the single largest gap between these four and the incumbents, and the one thing on this list that cannot be done on your own domain.
Publish pricing. Three of the four have no pricing page at all. It is one of the first things a buyer asks and right now there is nothing for an engine to answer with.
Get someone else to write the comparison. Keep the existing pages. Then earn one review, roundup or teardown you did not write. That is the harder and more valuable version of the same move.
And the negative result matters as much: when the site is already healthy, more on-site work will not move this. That is where most AI visibility budget currently goes.
How we measured it
Sixty buyer questions, generated from each company’s own site content and split evenly across category questions, head-to-head comparisons and problem-solving questions. Each was asked twice, giving 120 answers. The two runs agreed 96.7% of the time, which is how we tell a real result from a one-off. Questions naming the tracked brand are reported separately throughout, so nothing here is flattered by a question that handed the engine its answer.
Three answers were checked by hand against live ChatGPT before scaling, and all three matched. Two companies were cut at that stage: one had pivoted out of recruiting, one had become a market-research product. That is why this covers four companies and not six.
Some limits. This is one engine, and Perplexity and Google AI Mode behave differently. The per-company numbers are small, so the gap between one company at 0 and another at 1 is a single answer and is not a ranking, which is why that column stays as counts. And the review-site link is a pattern, not a test: we have not measured what happens after a company creates a profile.
ChatGPT →
Claude →
Perplexity →