Take the reading on the surface your buyer uses, because a model answering a prompt is not a product that retrieves

Read the surface a buyer actually opens. An interface that takes a prompt and returns text is a model answering a question. The consumer app is a product: it decides whether to search, chooses what to fetch, reads it and prints the sources underneath the answer. Everything a published page can influence happens in that middle step, so a visibility figure taken from a layer that never fetched anything is a figure about a different system. Both of our citation runs were taken on the ChatGPT and Gemini surfaces for that reason, and every cited address was recorded from the answer rather than inferred.

This is not a fine distinction. It decides whether a reading can be acted on at all, and it is the first thing to establish about any visibility number somebody shows you, including ours. We are paid to take that reading on a schedule and to write the pages that change what it finds.

The source list is the measurement

A reading without a source list cannot be scored. On 15 October 2026 we attempted 612 readings and used 597, and a reading was set aside precisely when it came back without the sources we were there to record. Those 15 are printed as the gap between the two figures rather than quietly dropped, and they are the clearest statement of what the measurement depends on: if the layer you are reading does not return addresses, there is nothing to count and no page to point at.

The second run, on 16 October 2026, recorded 1,260 readings with all 1,260 usable, holding 20,143 citations across 4,947 distinct domains under one rule: one domain in one reading, deduped. Those are counts of what two retrieval products reached for. They are not statements about what any model knows.

The two surfaces do not retrieve alike, and that is a product fact

The strongest evidence that you are measuring a product rather than a model is that the two products behave structurally differently on the same questions, asked on the same day, in the same wording.

Surface	Readings, 15 October 2026	Its most used source appeared in
ChatGPT	306	176 of its 306 readings
Gemini	291	19 of its 291 readings

One surface concentrates hard on a short list of documentation. The other spreads almost flat across many small sites. That is a difference in how each product selects and weights what it fetches, and it changes what a publisher should do in each column: on the first you are competing against an entrenched set, and on the second there is no entrenched set, which is worked through in what a top source at 19 readings of 291 means for a category. Neither sentence would be available from a layer that answers without fetching.

A layer that accepts your request is not always the layer that answers it

This failure has a general shape, and a different part of the same programme gives us a first party measurement of it. A help centre that was a source in 269 readings of our 16 October run permits all fourteen of the agents we checked in its own policy file, and refuses five of them at the edge with an HTTP 403, along with any request that gives no agent name at all. Two separate readings agreed.

One of those two layers published a permission and the other one decided the outcome. That is the question to put to any layer you are taking a reading from: not what it accepted, but what came back, and whether anything was fetched in order to produce it. Every rule in that set of files, host by host, is counted in our reading of 26 robots files.

What to record so a reading is about the product

• The surface, named and labelled. The app by name, and whether the row came from the app or from an interface. Rows from different surfaces live in different tables and are never added.
• Whether sources came back. This is a field, not a footnote. A reading with no sources belongs outside every citation rate and counted on its own.
• Every cited address, as returned. Not a judgement about the sources, the addresses themselves, so a proxy or a mirror host can be resolved before anything is divided.
• Whether the answer named the business. Its own column, with its own denominator. Appearing in a source list and being handed to a buyer as a choice break in different ways and cost different amounts to repair.
• A date for this reading, and a date already fixed for the next one.

The first of those five is the one that gets skipped, and it is the one that makes a dataset defensible a month later. What each of the names in this space refers to, model against product, is untangled in what each of those AI chatbot names actually is and what is still available.

The question this page does not answer

We have not run one question bank through an interface and through the consumer app on the same day and published the size of the difference. That comparison is not in our records, so what this page offers is the design and the reason, not a figure for the gap. When we run it, it will arrive with its own numerator, denominator and date, and the older rows will stay in print beside it.

Both runs above also cover two surfaces. The three further surfaces our product reads were never put these questions. Our product reads five surfaces and both of these datasets read two, and that distinction belongs in every sentence rather than in a note at the bottom.

Questions about which surface an AI visibility reading comes from

Can I not just call an interface and save the cost?

There are good reasons to: it costs less and it repeats more predictably. Keep those rows in a table of their own with the layer written on it, because a request that may never have fetched anything yields a figure about what a model recalls, and a published page does its work at the fetching step.

What tells me the answer involved a fetch at all?

Look for addresses. A product that retrieved will show you what it read, and our two runs are built entirely on those addresses. Where a surface tells you it searched without naming anything, record the claim as a claim and treat the reading as unscored until something resolves.

Does a reading with no sources mean the business is invisible?

It means that reading cannot answer the question. Those readings are counted separately, which is why 597 of 612 is printed as a pair rather than as a single total. A reading that fetched nothing has no slot for anybody's page, favourable or otherwise.

Should rows from two surfaces ever be averaged?

No, and the table above is the reason. A source appearing in 176 of 306 readings on one surface and 19 of 291 on the other are two descriptions of two products. Averaging the two yields a figure true of neither, and it conceals the one distinction a publisher is reading the table for.

What is the cheapest version of this for a small team?

Ten buyer questions with your name in none of them, put to the two apps your buyers actually open, with every cited address written down beside the date. That is a product surface reading, it takes an afternoon, and it can be repeated next month against the same wording.

The step you are trying to influence is the one that fetches

Everything a business can do here happens between the question and the answer: being reachable, being indexed, being the page that settles the question in its opening lines. That step exists only on the surface a buyer opens, so a reading taken at any other layer describes a system nobody is buying from. Reading five named surfaces on a schedule and writing the pages that change what they find is our business model, stated plainly rather than implied. The Search Console half of our own record, window by window, is published at aiknowsus.com/proof.