We measure whether AI answer engines cite your company when a buyer asks them what to buy — how often, how much it varies, and who gets cited instead. Windsor, Ontario.
Three routes, and every number we report is labelled with the one that produced it.
The hand-collected route is slower and it is the reason this holds up when a technical buyer asks how we know.
A citation is a domain the engine actually pointed at — either in its own grounding metadata, or as a URL in the answer itself. A brand name in a sentence is a mention. We record mentions separately and they never count toward a citation figure. Merging the two is how a visibility number stops meaning anything.
This is the part the industry skips, and it is the one that decides whether any of the rest is worth reading.
AI answers are not deterministic. Ask the same question twice and you can get two different lists of companies. So a single query proves nothing, and a handful of queries returning nothing proves much less than it appears to.
We measured one real buying-intent prompt against Gemini grounded search, 50 runs, one company. It was cited in 22% of runs — 95% confidence interval 12.8% to 35.2%. Every one of those citations carried the domain in the engine's own grounding sources, so this is a citation rate and not a mention rate.
At a true rate of 22%, here is what asking a few times actually tells you:
| Runs | Chance of a zero anyway | A zero bounds the true rate at |
|---|---|---|
| 3 | 47.5% | 63.2% |
| 6 | 22.5% | 39.3% |
| 12 | 5.1% | 22.1% |
| 20 | 0.7% | 13.9% |
Read the top row again. A three-run zero is close to a coin flip, and it is consistent with a company being cited in nearly two-thirds of sessions. Anyone who runs a prompt three times, sees nothing, and tells you that you are invisible in AI search has told you almost nothing — and you can disprove it from your phone in ten seconds, which is the worst possible way for a measurement to be wrong.
So we do not say it. Three runs is our floor for showing that a company is cited. Twelve is the floor for claiming it is not, and below that the report reads "not seen in N runs, which only rules out a rate above X%".
No report, email or deck from us carries a number saying remediation will improve visibility by some amount. Not "+40% citations", not "2× more likely to be cited", not a hedged "up to 30%".
The reason is not caution. A 2024 Princeton paper reported visibility lifts of 30–40% from nine content tactics, and much of this industry still sells against those figures. In 2025 C-SEO Bench retested nine such methods across six domains and 1,921 queries under realistic retrieval: three of fifty-four method-domain cells reached significance, adding statistics — the strongest lever in the original paper — lowered ranking in nineteen of twenty-four settings, and a plain SEO baseline was roughly 7.6× more effective than the best conversational method tested. Whatever advantage exists decays as more people adopt it.
C-SEO Bench, NeurIPS 2025 Datasets & Benchmarks — arXiv:2506.11097
A lift percentage in a client document is therefore a number a technical buyer can refute with one citation. Remediation here is scoped as hygiene — the crawler can reach the page, the entity resolves, the company appears on the third-party pages the engines actually cited — and as something whose effect we measure again at 30 and 90 days, never as a gain we forecast.
A document can shape an answer without being cited in it. Everything we measure is a citation — the part that sends a reader back to you. A company whose content is being absorbed and uncredited looks, in this data, identical to one being ignored. That is a real limit and it belongs in the answer whenever someone asks.
EntityGround is an independent measurement practice based in Windsor, Ontario, working across Windsor-Essex and southwestern Ontario.
Contact: contact@entityground.ca