Platform · Measure
Five marks on every number
The standard every AEOSearch report is held to, in the words a buyer can check in a minute and repeat in a sentence.
A visibility score with no sample size, no model and no range is an opinion with a decimal point. So every number a report prints carries five marks, and a number that cannot carry all five is not printed.
- 01
Sample Size Stated
Every rate says how many answers it counts.
A mention rate is a count over answers, and the count is printed with it: 20 questions, three passes each, per engine. One pass is sold as one pass and never described as three.
- 02
Model Named
The model that answered is stored on every row.
GPT-4o mini and GPT-5.6 Sol answer the same question differently, so the report names which one answered, per engine, read back from the provider's own response rather than from what we asked for. Since 16 September 2026 the row also keeps the model the provider named, beside the one we requested, so a silent alias change shows up as a mismatch instead of passing as the same measurement.
- 03
Grounding Flagged
Web search off measures memory, not the assistant.
An ungrounded answer tells you what a model remembers from training. A grounded one tells you what ChatGPT, Gemini, Claude, Grok or Perplexity says today with search on. The flag is stored per answer and printed per engine, and an ungrounded result is never presented as the live assistant's answer.
- 04
Range Shown
Every mention rate carries its 95% interval.
Forty percent on 20 questions is a wide band, and the page says how wide, beside the number rather than in a footnote. The three passes of one question count as one question: more questions narrow the range, more passes of the same question do not.
- 05
Raw Answer Kept
Any number traces to a row you can read.
The full answer text, its citations and the provider's response stay stored. A number that cannot be traced to a stored answer is not printed, and an engine whose calls failed is disclosed as an outage, never counted as a zero.
When a page is changed
Measuring is free to read and changing a site costs three more marks. A change missing any one of them is a defect, not a feature.
Previewed
The WordPress plugin and the Shopify app show the exact change before it lands; the platform itself never writes to a site.
Receipted
Every applied change is recorded with a rollback token, bound to the site that opted in, and the report says “applied, rollback available” beside it rather than “nothing has been changed.”
Reversible
A change that cannot be undone is refused before it is recorded, and revoking consent can never block a rollback.
What we refuse to print
- Never a promised mention or a projected ranking. AI answers change between runs, and every figure is a dated snapshot.
- Never an outage sold as invisibility. A failed call is stored as a failure, excluded from every rate, disclosed on the report, and not charged for.
- Never a crawl count we did not fetch. The pages a report says it read are the pages it read.
- Never an inferred AI-traffic number. Google AI Overviews and AI Mode visits arrive as Google organic and cannot be separated from it, so they are reported as exactly that.
Against the field
On 12 September 2026 we read every AI-visibility listing in the Shopify App Store we could find, 37 in all, for the same five marks. The column on the right is what those listings state, counted, not what we think of them.
| Mark | AEOSearch | 37 App Store listings |
|---|---|---|
| Model version stated | Per engine, per answer | 0 of 37 listings |
| Grounding stated | Per engine, per answer | 0 of 37 listings |
| Sample count stated | On every rate | 2 of 37 listings |
| Interval published | Beside every mention rate | None found |
| Raw answer kept | Every answer, indefinitely | Not checked by the sweep |
Listings were read from their public store pages. A listing that keeps one of these marks inside the app rather than on its page would not be counted here, which is a limit of the sweep, and the reason the column says “stated”.
The one-minute version
Every number on an AEOSearch report says how many answers it counts, which model gave them, whether web search was on, how wide the range is, and where the stored answer lives. If a number can't say all five, it isn't printed. When a page is changed, the change was previewed, receipted and can be undone.
The long version is the methodology. The free scorecard runs on the same marks as a paid report, so the cheapest way to check us is to run one.
Start with the free scorecard.
Let's run the free scorecard and read the five marks on your own numbers.
$0 to start · no card required · every number traces to a stored answer