How we measure, and what we never claim.
Assistants rarely give the same answer twice. So we never report a single answer: we ask each question several times, count how often your client is named, and show how sure that count is.
Why there is no “rank in ChatGPT”
We do not report a “rank in ChatGPT”. Ask the same question twice and the list of brands changes, so a position read off one answer says very little.
SparkToro asked AI tools the same recommendation questions many times and found the list of brands almost never repeated: fewer than 1 in 100 runs gave the same list.
- What we report instead. What we report instead is visibility: the share of sampled answers that name the brand, per assistant. Assistants are shown side by side, never blended into one score.
- How the client is named. Inside the workspace we also show how the client is named when it is named: in how many answers it comes first, or ahead of every tracked competitor. That is a count across samples too, not a position read off one answer.
How we sample
Visibility is only as good as the questions and the number of answers behind it.
- Questions. Questions are the buyer prompts you track for a client. You can generate a first draft from templates, import your own list, and edit either until it reads the way buyers actually ask.
- Samples. By default each question is asked three times on each assistant you switch on, every run, and you can ask more. A cell with fewer than three answers in its window shows a dash instead of a number.
- Cadence. Every two weeks by default; weekly or daily if you switch it on.
- Through the API. Each assistant is asked through its developer API with web search switched on, not through the consumer app. What a person sees in the app can differ.
- Sources. Every answer is stored with the pages it cited, so each figure can be traced to answers and sources.
Measured by default
Switched on per client, using more AI checks rather than costing extra
Not measured
Windows and ranges
Answers are grouped by week. Screens add up a rolling window (the last 28 days by default) so a figure rests on more answers; reports compare the start of the period with the end.
- The range. The range around a share is a 95% interval worked out from the number of answers behind it. Few answers, wide range.
- Up, down, or can’t tell. When two periods' ranges overlap, the change is labelled as within what the sample can tell apart, not as up or down.
What confidence means
Confidence describes how many answers sit behind a figure, not how good the result is. The same thresholds apply on every screen, in reports and in the API.
| Label | Answers behind the figure |
|---|---|
| confidence: low | fewer than 12 |
| confidence: medium | 12 to 59 |
| confidence: high | 60 or more |
Experiments and estimated contribution
An experiment compares the topics a piece of work touched with the client's other topics that it did not touch, over the same weeks. The baseline is the 28 days before the work was marked done.
How the estimate is made
The change in the touched topics minus the change in the untouched ones, from the baseline to the weeks after the work. Baseline: the 28 days before the work was marked done.
Where its confidence comes from
Its confidence comes from whether untouched topics exist and stayed steady, how many answers came in afterwards, how deep the baseline is, and whether a new source started being cited.
How wide the band is
An action's estimated contribution is shown as a band around the estimate. The band has a fixed width: it marks the figure as rough, it is not a statistical interval.
What it cannot do
Untouched topics of the same client are a stand-in for a comparison group, not a true one. Without them, the report says movement cannot be separated from platform-wide change.
What we keep
Recorded on every answer: the raw text, the pages it cited, the model version that produced it, and what it cost to ask.
- Answers. Every answer is stored as it came back, with its model version and cost, so any figure can be checked against the text behind it.
- Decisions. Human decisions stay put. A recalculation never overwrites what your team decided about an opportunity.
- Publishing. The product reads what assistants already answer and tells you what it found. Anything that changes a client's site stays a human decision.
What we never claim
Stated here rather than in the footnotes, because you will repeat these lines to your own client.
- A rank or position in any assistant
- A single 0–100 visibility score
- That a piece of work produced a change: we show what followed, with a comparison where one exists
- A forecast of the share a client will reach
- A revenue figure for visibility
- Anything about assistants we do not measure
A quarter, not a week
Models re-crawl and re-cite over weeks. Sixty to ninety days is the honest unit here; a short period shows early signal rather than settled results.
Movement, not attribution
With one client and no untouched topics to compare against, movement cannot be separated from platform-wide drift, and the report says so where that is the case.
Ranges, not exact numbers
A share comes with a range that depends on how many answers sit behind it. An action's estimated contribution is shown as a fixed-width band around the estimate, next to a confidence level, so it never reads as an exact number.
No revenue figure invented for you
Visibility is a share of answers. What that share is worth belongs to your client's model, not to ours.
Only assistants with a public API
Microsoft Copilot and Google AI Overviews / AI Mode offer no public API, so they are not measured, and no report estimates them.
Nothing is published for you
The product reads what assistants already answer and tells you what it found. Anything that changes a client's site stays a human decision.
See the method on one of your own clients
The free audit asks the questions, keeps every answer and shows each figure with the answers behind it. Nothing is charged to run it.