Original measurement

How many companies can an AI assistant name at all? Between 8 and 53, depending on which question you ask.

Repeat the same question and count the distinct companies that accumulate. The pool behind “build me a new site” is roughly 8 companies. The pool behind “who helps businesses like mine” is roughly 53.

8–53

estimated companies an assistant can name, depending on the question asked

Key findings

Repeating the same question and counting distinct companies, the pool an AI assistant draws from ranges from about 8 for “build me a new site” to about 53 for “who helps businesses like mine”.
For the broadest question we tested, the number of distinct companies named rose from 5 to 10 to 18 across three readings and was still climbing.
The implied chance that any one company appears in a given answer ranged from 11% on the broadest question to 53% on the narrowest.
A vague question puts a company in a room of roughly 53 competitors; a specific one puts it in a room of roughly 8.

How this was measured

When: Readings taken 30 and 31 August 2026.
Instrument: Capture–recapture on repeated readings of one identical question: the companies named on reading one are the marked population, and the overlap with later readings estimates how many were never caught.
Surface: Logged-in ChatGPT, United States, no metro specified.
Sample: 19 readings across 5 questions, all in a single trade. Company names are folded to a single case before counting, so one company spelled two ways is counted once.
Pre-registration: Pool size was one of three reads fixed in writing before the run.

The count, question by question

The question asked	Readings	Distinct companies after each reading	Estimated pool	Implied chance of appearing
A one-off report	4	5 → 9 → 14 → 18	~28	19%
Monthly monitoring	4	5 → 9 → 10 → 15	~12	42%
Fix my site	4	4 → 4 → 6 → 8	~14	25%
Build me a new site	4	5 → 7 → 9 → 9	~8	53%
Who helps businesses like mine	3	5 → 10 → 18	~53	11%

The middle column is the running union: how many different companies had been seen at all after each successive reading. A union still climbing at the last reading means the pool is larger than the estimate.

Four of the five were still climbing

Only 1 question - “build me a new site” - stopped producing new names by the last reading. Its 9th company arrived on reading 3 and reading 4 added nobody.

The other 4 were all still finding companies they had not named before. That is the signature of a pool bigger than the readings have exhausted, so every estimate in the table above should be read as a floor rather than a measurement of the ceiling.

The broadest question is the clearest case. 3 readings, 5 companies each time, and 18 distinct names - barely any repetition at all.

Why the number differs so much by question

The narrower the job, the smaller the room. “Build me a new site for a massage studio” has a describable answer and roughly 8 companies who plausibly own it. “Who helps businesses like mine get found” has no natural boundary, and the assistant has 50-odd candidates it can reach for.

This cuts against the instinct to target the broadest phrasing. The broad question has more people asking it and worse odds inside it. The specific question has fewer askers, a shorter list, and a far higher chance that any given company is on it.

What this means if you are buying

“Get into the AI answer” is not one objective. It is at least 5, and the odds differ between them by a factor of 5.
If you are choosing where to compete, the specific job beats the general one: fewer competitors in the pool and a materially higher chance of appearing in any single answer.
It also sets what an honest report can promise. Where the pool is around 50 and the answer holds 5 names, a single appearance is a low-probability event, and reporting one reading as a result would be misleading.

Honest limits of this measurement

Capture–recapture assumes every company is equally likely to be caught on any reading. That is certainly false here - some companies are named far more often than others - and that bias pushes pool estimates downward. Read them as lower bounds.
3 or 4 readings per question is thin. These are the right order of magnitude, not precise counts.
One trade and one assistant on this surface.
The estimator needs repeat readings on one model tier; questions with only 1 or 2 such pairs carry correspondingly more uncertainty.
Two things separate these pools from the ones on our category pages. These are service-intent questions, answered in prose with no business list; a local question gets a business list, and there the pool runs to five or six. And the figures here estimate how many companies exist behind the readings, including ones we never saw, while the category pages count only the names that actually appeared - across 20 such questions the observed pools ran from 2 to 10.

Questions about this research

What does “estimated pool” actually mean?

It is an estimate of how many distinct companies the assistant could name for that question at all - including ones that did not turn up in our readings. It is inferred from how much two readings of the same question overlap: heavy overlap implies a small pool, little overlap implies a large one.

Why is the implied chance of appearing so high for some questions?

Because the pool is small. If an assistant names 4 companies out of a pool of about 8, any one of them is roughly a coin flip to appear. That is arithmetic, not a promise - real companies are not equally likely, and a company nobody has ever heard of is not in the pool at all.

Does a bigger pool mean the category is more competitive?

It means the answer is less concentrated. More companies can plausibly be named, so any one of them appears less often. Whether that is good or bad depends on whether you are already in the pool.

Can you measure the pool for my category?

Yes - it is the same procedure and it runs on the free rig: repeat one question several times, count what accumulates. It takes readings spread over time rather than a single check.

More measurements from the same instrument

We run this measurement for a living. If you want it run on your own category and your own city, the free scan is the same instrument on one question, and the report is the full set. Nothing here is behind a form.

Every category we measure has its own page, with the questions customers put to an assistant in that category and what a reading costs. See all industries we measure →

Published 31 August 2026. Every figure on this page is recomputed from raw observation files by a single script, so any of them can be traced back to the readings behind it. If you are named here and believe a reading is wrong, tell us: we will run the measurement again and publish what it returns, whichever way it goes.