How this is measured, and where it stops
Last updated
One measurement is one prompt, on one engine, at one moment
Everything in the product is built from that unit. A window’s visibility is a count of those units; a competitor’s lead is a count of those units; an action’s follow-up compares two sets of them. Nothing is modelled, estimated or extrapolated on top.
Engines that can be asked today
ChatGPT
ChatGPT (API)
Microsoft Copilot
Google AI Mode
Claude
Gemini
Perplexity
Grok
Meta Muse
DeepSeek
Three collection methods, and they are not the same evidence
Each answer records which of these produced it, because the difference matters when you are deciding how much weight to give it.
Model API
The prompt is sent through an API — the provider's own, or a gateway such as OpenRouter that routes to it — with web search enabled where it is offered.
The requested model, the reported model and token usage when supplied, and any returned citations. Unreported fields stay unknown.
An API answer is what a model gives an application. It is not guaranteed to be word for word what a person sees in the consumer app, which has its own system prompt, memory and interface.
Consumer UI
A vendor drives the real, logged-out interface — chatgpt.com, copilot.microsoft.com, Google AI Mode — and returns what a person would have been shown.
The answer a visitor actually sees, with the links attached to it.
The model that answered is usually not disclosed, the surface changes without notice, and one of these engines returns no served-country field at all, so its answers are labelled Global rather than claiming a country we cannot verify.
Search provider
A third-party search API runs the query and returns the AI overview the results page carried.
A view of the answer shown beside conventional search results.
The provider decides the locale and can serve a country other than the one requested, so country is asserted only where it can be read back from the response.
One engine brand, more than one method
ChatGPT can be asked through an API and through the interface a person uses, and they are listed as two engines because they are two measurements. Google appears both as AI Mode and as the overview on a results page. Each answer carries the method that produced it, so you can compare like with like — and see where a brand is strong on one surface and absent on another.
Reported queries are the engine’s, not your buyers’
Some providers report the searches their engine ran on the way to an answer. Those are useful — they show the words an engine reaches for on your topic, and the optimizer writes against them. They are the engine’s own queries, not a count of what people typed, and no figure here is a search volume.
A sample of your questions, not a census of your market
The prompts are the ones you wrote, or ones suggested to you and kept. They are asked on a schedule from a country you set. That makes them a deliberate, repeatable sample — which is what allows a comparison over time — and it also means they are not a record of what real people typed, how many typed it, or what each of them was shown.
Consumer assistants also personalise: memory, account history, location and the version rolling out to one user and not another all change an answer. What a repeatable sample gives you instead is a trend you can act on and check again.
The schedule
Scheduled collection runs every day. Every enabled prompt is asked on every enabled engine, within the project’s daily attempt allowance, and each attempt is deduplicated per UTC day so a retry cannot inflate a count. You can also run a prompt yourself at any time; a manual run is recorded exactly like a scheduled one.
An assistant does not give the same answer twice, so one day is a single draw. A daily schedule means a fortnight is a few hundred answers to read a movement from, and a month is a trend rather than a handful of samples joined by a line.
What counts, and what is counted separately
Most disagreements about a visibility number are really disagreements about its denominator, so here is ours.
- Successful answer
- An attempt that came back with text to measure. This is the denominator: visibility is answers naming you over successful answers, and nothing else.
- No answer
- The engine replied and had nothing — an empty overview or a provider-reported no-answer result. These are counted and shown separately, and they are excluded from the denominator. Treating them as a zero would turn a quiet engine into a visibility collapse.
- Failed attempt
- A provider error, a timeout, an expired lease. Counted separately again, never retried silently, and never counted as a measurement. A week where a third of attempts failed is a week whose numbers you should distrust, so the report says so.
- A day with no collection
- Stays blank on a chart rather than being drawn as a drop to zero.
Inspecting a page tells you about the page, today
Page Audit, the cited-page checks in Gap Analysis and the optimizer’s page reading all work the same way: a bounded HTTP fetch of the document, with a size limit, and parsing of the HTML that came back. No browser is run, so content that only exists after JavaScript executes is not seen. “Not found” always means not found in the inspected text.
This cannot prove what an engine read. A page you fetch today may have changed since the answer was given; an engine may have used a cache, an index, a snapshot, or nothing from the open web at all. A page check is evidence about your page, and a hypothesis about the answer.
What a robots.txt result means
Allowed means the site’s robots.txt permits that crawler on that URL after redirects. It does not mean a crawler visited, that an engine used the page, or that a blocked crawler is why you are absent — some engines are grounded by an index rather than by their own crawler. An unreachable page, or a robots file that returns HTML, stays unknown rather than being read as permission.
How much this moves on its own
- The same prompt asked twice can answer differently: different brands named, a different number of sources, sometimes no sources at all. That is the engines, not the collection.
- Providers change the models behind a product without announcing it, so a fortnight-long comparison can straddle two different models. Requested and reported model fields are kept separately; an undisclosed model stays unknown.
- Three collection days a week give a week's worth of samples instead of one lucky reply, and comparisons are always shown against the previous window of the same length rather than as a lone number.
- Small movements may be sampling variation. Compare matched questions and repeated observations. Read the sentences and the sources before you read the delta.
Reading a before-and-after
When you record a change, the follow-up compares the same prompt wording, the same engine, the same geography, the same model and the same collection method on both sides of the date, over a window you choose. If either side is missing that coverage, it reports insufficient comparable data and shows the counts rather than printing a difference. What it prints when it can is an observation about two samples — the world changed too, and this measurement cannot separate your work from the rest of it.
What this product claims, and what it refuses to
It does
- Ask the prompts you wrote, on the engines you enabled, from the country you set.
- Store every answer with its sources, its model, its collection method and its timestamp.
- Count answers, empty replies and failures separately, and show all three.
- Open the underlying answer from every number in the product.
- Refuse a before-and-after comparison when the two sides were not collected the same way.
It will never claim
- That an assistant will recommend you, or that any amount of work will make it.
- Why an engine chose one source over another. No engine publishes that.
- That a score is the probability of being cited.
- That a prompt we generated is evidence of customer demand, or that any figure here is a search volume.
- That a change after your edit was caused by your edit.
- That this is everything your buyers ask, or everything any of them sees.
Data is held in the European Union. Your prompts go to the engines you enable; your account details do not. The privacy policy lists every provider involved.