We scanned ten companies nobody has heard of

Everyone writing about AI search is measuring the same companies. Stripe, Notion, Figma, whoever is large enough that the answer is already known. That tells you what visibility looks like at the top, and almost nothing about what it looks like where most software companies actually sit.

So we bought the other end of the distribution. Ten companies from recent Y Combinator batches, teams of three to fifty, all with live products and real customers. Same buyer questions, same engines, same pipeline we run for anyone. It cost thirty dollars.

What we ran

For each company the engine generates the questions a real buyer in that category would ask, puts them to the leading AI search engines, and records every product named and every source cited. Nothing was hand-tuned per company and nothing was rerun to get a better number.

We had already measured ten household names the same way. That gave us both ends of the distribution on one instrument, which is the part nobody else has, because measuring the bottom end costs money and produces numbers that are unflattering to whoever paid for them.

Half of them were never named once

Five of the ten were named in none of their answers. Not ranked low, not mentioned in passing, not described inaccurately. Absent, across roughly 58 questions each, in the category they sell into.

The number that reframed this for us is what those answers actually contained. Across a single company's scan the engines cited around 740 sources, about a dozen per answer. The engines were not short of things to say. They were citing steadily throughout, and for five of these companies not one of those citations was them.

The exception, and it is one company

One company in the cohort was named in 91% of its answers. That is higher than nine of the ten household names. It is the largest and longest-established in the cohort, and it sells into a narrow technical niche with a lot of documentation written about it.

That is a single data point and we are not going to build a theory on it. It is worth recording because it points somewhere interesting: the thing that predicts being named may be less about how famous a company is and more about how crowded and well documented its category is. A giant competing in a category with a dozen credible alternatives can lose to a small company that owns a narrow one. Testing that needs a different cohort, designed for it, and we have not run it.

The engines agree about obscurity

We have published before that the answer engines do not share a playbook, and they do not. Measured on the household names, the gap between the harshest and the most generous engine is 15.6 points.

On the unknown companies, all four engines landed within 1.2 points of each other. They disagree about which known companies deserve to be named. They barely disagree about whether to name an unknown one.

That refines our own earlier finding rather than contradicting it. Engine divergence is real, and it is a property of the contested end of the distribution. Choosing which engine to optimise for is a question you get to ask once you exist in the source material at all. Before that, all four give you the same answer.

We predicted the opposite

Before running this we wrote down what we expected, which is the only way a result means anything afterwards. Two predictions. The first was that unknown companies would be named far less often, and that held.

The second was wrong. We expected that when an unknown company did get named, the engine would link to its own website more often than it does for a household name, on the reasoning that with no third-party coverage there is nothing else available to cite. The opposite happened. Own-domain citations were 0.82% for the unknown cohort against 2.63% for the household names, about a third rather than more.

The reasoning was wrong in a way that is worth understanding. The engine is not looking for something to say about you and settling for your own site when nothing better exists. It answers the question from whatever has been written, and if you are not in that literature you are missing from the whole answer, its citations included. There is no fallback to your own domain, because you were never the subject.

What we would do with this

• Find out whether you are at zero before doing anything else. The difference between 3% and 0% is not a matter of degree. At zero, tuning your own pages changes nothing, because the engines are not reading a page about you and deciding against you. They have nothing to read.
• Get written about before you get optimised. The work that moves a company off zero happens on other people's pages: comparison posts, forum threads, videos, documentation that mentions you alongside the alternatives. Your own site confirms what the engines already believe. It does not create the belief.
• Ignore per-engine tactics until you are on the board. All four engines agreed to within about a point on these companies. Advice about optimising for one engine over another is advice for companies that already exist in the source material.
• Watch the category, not just yourself. When we asked about these companies, the engines answered with someone. Knowing who gets named instead of you is a more useful competitive list than the one on your pitch deck, because it is who your buyer actually hears about.

How we measured this, and what it cannot show

Ten companies from recent Y Combinator batches, teams of three to fifty, every one with a live product. Ten household names measured the same way earlier. Around 58 generated buyer questions per company, put to the leading AI search engines, with every answer and citation stored. The scans cost thirty dollars.

The limits, and they matter. Ten companies is a small cohort. The mention-rate gap is large enough to survive that; the smaller numbers in this piece are thinner and should be read as indicative. The two cohorts are not matched on category, because recent Y Combinator batches contain no equivalent to the deploy and work-tracking companies in the household set, so this compares unknown against household name rather than young against old within one market.

We are reporting the unknown cohort in aggregate and will not name which company scored what. Naming a large public company inside a measurement is fair, and we have done it. Naming a small company as invisible is a different act, and it would cost them something real to make a point we can make just as well without it.