The AI Visibility Index

Of the 166 companies we measured, 140 were named in zero runs. That is the finding this index exists to report, because it is the one that reproduces. Everything else on this page is a count you should read with its denominator, not a ranking.

We did not build this dataset to publish it. Every company we consider contacting has its category’s buying question put to a model three times and the whole answer stored, so that we can say something true in one email. Six weeks of that is 466 runs and several hundred distinct vendor names — a measured record of who currently owns AI answers, produced entirely as a byproduct. Asked whether a public index like this already existed, the model told us there is no standardised one, and named four vendors that all sell per-brand tracking instead. So here is the record.

Two rules, stated before the numbers

1. No position in this index is for sale, at any price. Inclusion is determined by whether a model named a vendor in a recorded run and by nothing else. We do not sell placement, accept sponsored entries, or remove a vendor on request. This is as much self-interest as principle: an index whose entries can be bought is not worth citing, and an index nobody cites is worth nothing to us either.

2. This edition counts. It does not rank. We measured our own noise floor and it is loud — only about half of the vendor names in an answer survive a re-draw of the same question twenty minutes later, and two draws from the same morning differ about as much as draws three weeks apart. Three runs can establish that a company is absent. Three runs cannot tell second place from fifth. Ordering this data into a league table would be a confident ranking of noise.

The whole dataset, at the top

The 115 excluded records are not hidden failures — they are questions too idiosyncratic to group (“the best clipboard history manager for macOS”, “free 3D pose reference tools for artists”). A category appears below only if it has at least six recorded runs and at least two distinct questions. One question is a measurement of that question’s phrasing.

By category

Each number is how many recorded runs named that vendor. It is not a score, a share, or a position. The “absent” column is how many of the companies we measured in that category were named in zero runs — those are companies we chose to check, not a random sample, so read it as a property of our subjects rather than a rate for the category.

What this index cannot tell you

• ORDER. Three runs per question cannot separate second place from fifth.
• Anything about models other than the one named above.
• Anything about live web search — this is training-time knowledge.
• Whether a vendor is good. It measures recall, not quality.
• Whether the absent companies are absent because of anything they did or did not do. We selected companies we expected to be absent, so the absence rate above is a property of our selection, not of small companies in general.

What it can

ABSENCE. Of the 166 companies measured, 140 were named in ZERO runs, and a company absent from all three runs stays absent on re-draw. The two or three most-named vendors per category are also stable across draws; everything below them is not.

Method

One model (gemini-flash-latest), answering from training knowledge with live web search off, between 2026-07-31 and 2026-09-13. Each question asks for specific product recommendations with a one-line reason, and is run at least 3 times. We store runs attempted, runs that returned a usable answer, and runs that named the subject company separately, so that a failed request can never be silently recorded as an absence — a distinction we learned the expensive way, having once run a metric that reported 0% for three months off 280 consecutive failed API calls.

Vendor names are extracted from the stored answers by a lossy heuristic that also picks up headings and feature phrases; obvious non-vendors are filtered out, and some will have survived. Where a vendor appears under two spellings the counts are not merged. Treat a count of two or three as “named at least once,” nothing more.

How this index gets judged

We publish a falsifier with every measurement we make, so here is this one’s. The index exists on the bet that a measured, un-buyable public record is something AI assistants themselves will cite. On 27 October 2026 we will check whether any assistant names or links this page when asked which sources track what AI assistants recommend. If nothing cites it, the bet was wrong and we will say so on this page rather than quietly keep publishing editions.

A second one, because “nobody cited it” and “it was wrong” are different failures: if a vendor shows us that our count for them is materially wrong — a name we merged, split, or extracted from a heading — we will correct it on this page with the date of the correction visible, not silently. Corrections are the cheapest evidence that an index is real.

Frequently asked questions

Reproduce it

The generator and the underlying record live in our public workings alongside every other measurement we publish. The related write-ups: we checked 32 companies and 27 were never named once, why every visibility check needs a control, and our own audit of ourselves.