When someone asks an AI what to buy, are you in the answer?

Up to 10 questions on a project. Each one is put to each of the 3 models every week, and every answer is stored in full, with the sentence your brand was named in kept beside the count so you can read it rather than take our word for it.

• How often you were named at all.
• How often a model linked to one of your own pages — a different and stronger signal than being named.
• Who was named instead, among the rivals your project tracks.
• Where you came in the order of those tracked names, when more than one was named.

The rate is taken over the last 5 sweeps rather than off the newest one. A single answer is a sample: ask a model the same question twice and it will not answer identically, so one week’s reading is noise and the line across weeks is the measurement.

Every rate divides by the answers that came back. An answer a model never returned is not a measurement, so it is never counted as an answer that failed to name you — a sweep in which every call failed reads as no rate at all, and never as nought per cent. The grid says which of six things happened to each question on each model: named and cited, cited, named, not named, no answer, and not yet run. "Not named" is a finding about a market; "no answer" is an outage, and the two must not look alike.

The models are named by model — GPT-5.6, Claude Sonnet 5, Gemini 3.6 Flash — and never by the name of an app built on one. That is not a style choice: a figure described as "how often ChatGPT names you" would be a claim about a product with its own instructions, its own memory of the person asking and its own way of searching, none of which an API call carries. What this product can tell you is how a named model answers a named question. What it cannot tell you is what any of those apps will say to one particular person.

A place in the order is an ordinal over the brands you asked about, not over every brand a model mentioned. Adding a rival to the project therefore moves your own place without any model having said anything different, which is why the column says what the ordering is over.

A brand whose name is an ordinary word — Gap, Next, Boots — will match sentences that are not about it, and this product does not pretend otherwise. Every match is stored with the sentence it was found in, so a false one takes a glance to spot; a link to your own domain is the signal with no such ambiguity, which is why it is counted separately rather than folded in.

A sweep spends the workspace’s monthly AI budget — a dollar figure rather than a count of calls, and the point past which this product starts no new sweep on your behalf, never an amount you owe. Scale carries $60.00 of it; the pricing table has the row for every plan, and a plan whose figure in it reads Not included does not carry the feature at all.

A workspace over its budget has that week’s sweep skipped, and nothing is recorded for it. A sweep recorded as failed would look exactly like evidence that no model mentioned the brand, when the truth is that nobody asked — which is a difference worth a missing week on a chart.

You can also run a sweep on demand rather than waiting for the week, with an hour between presses. The limit is not only about cost: two sweeps an hour apart measure the same noise twice, at full price.