Why does an AI recommend one business over another?
It is not a ranking and it cannot be bought. A model assembles the answer from whatever it can read about businesses in that area …
Proof
How do you measure whether AI is recommending you?
The short answer
You ask the models the same questions on a schedule and record whether your business was named, on which platform, and in what position. That is the only direct measurement that exists. Everything else — traffic, impressions, rankings — is a proxy, and most AI visibility reporting sells proxies as results.
Probe cadenceWeekly, three platforms
Stored per runPrompt, response, citations, date
Our own trend15.2% → 6.3%, May to Aug
01 · Why
Why it matters
Most reporting in this category cannot be checked, which is the problem with it.
Most reporting in this category is unfalsifiable
It is worth being blunt. A great deal of what is sold as AI visibility reporting cannot be checked: a score out of a hundred with no method, a graph that only goes up, or plain search metrics relabelled.
A measurement you cannot falsify is not a measurement. The test to apply to any provider, including us, is simple — can they show you the exact question asked, the exact answer that came back, the platform and the date? If not, you are being shown a proxy.
A measurement you cannot falsify is not a measurement.
The number that matters can go down
An honest measurement moves in both directions. Ours does.
On the demo site we watch most closely, unbranded discovery citations went from 15.2% in May to 6.3% in August across an identical prompt set on identical models. We could have shown you the May number and stopped. Instead it is in the report, because a metric that only rises is a metric nobody is really taking.
02 · What
What it actually is
What actually gets recorded, and why the two scores stay apart.
What actually gets recorded
The exact prompt text, stored verbatim
The platform and model version it ran against
The full response, so the claim can be checked later
Whether the business was named, and where in the answer
Which other sources the model cited
The date, so change over time is visible
Branded and unbranded are separate scores
Collapsing them into one number is how this gets misleading, so they are reported apart.
Branded prompts ask about you by name and answer the verification question. Unbranded prompts ask an open question about the category and answer the discovery question. The first is winnable now; the second is much harder. A single blended figure hides exactly the thing you would want to know.
03 · How
How it works
How the probe runs — and what it cannot see.
How the probe runs
Each business gets a panel of prompts covering both shapes. That panel runs weekly against Claude, ChatGPT and Gemini, and each response is parsed for citations and stored whole.
The report goes to you by email and lives in your portal — including the questions you are losing, which is the part that tells you whether anything is actually working.
What we cannot measure
Two honest limits.
We cannot see private conversations. We measure what the models return to our probes, which is a sample, not a census of what your customers asked.
And attribution to revenue is indirect. We can show you were named more often; connecting that to a specific booked job needs your side of the story too.
04 · Questions
Common follow-ups
Can I see the raw responses?
Yes. Storing the whole response is the point — a citation count you cannot audit is just a claim.
Why weekly rather than daily?
Model answers move slowly and probing costs real money on every platform. Weekly catches change without spending your subscription on API calls.
What if the numbers go the wrong way?
Then you will see it, as we did. That is the difference between a report and a marketing dashboard.
Start where AI starts
Try it for 14 days. If it isn't what we said, we refund your first month in full and take the site down.