Methodology
How we measure, and what we cannot see.
AI answers vary from run to run. How we collect them, how we score them, and where our numbers stop being reliable.
What runs, in order.
The stages behind the two lanes on the price list.
The answer-engine lane.
Getting you named when a buyer asks an AI app what to use.
- Map
The buyer questions in your market, by city and persona.
- Measure
Blind, so the answer is the one a buyer gets.
- Find the gap
Per market: who is named, where you are not.
- Build
The answer to the question, not a brochure.
- Publish
To your own site, after you have seen it.
- Prove
Weekly on your priority questions, 30 and 90 days on the rest.
The search lane.
Getting you found when a buyer types the same thing into Google.
- Seed
Your category, in the words your buyers use.
- Harvest
Live volumes, read first party from Google.
- Ledger
It compounds: every keyword it has ever seen is kept.
- Cluster
By intent, so two pages never chase one answer.
- Rank
By leverage: what you can win, not what is biggest.
- Publish
One page per intent, never two.
Collection, one AI app at a time
We check the real apps your buyers use, and every one of them searches the web before it answers.
ChatGPT and Gemini are read through a third-party data provider. Perplexity and Claude are checked through their official APIs, with live web search.
| AI app | Source |
|---|---|
| ChatGPT | The real ChatGPT app, through a third-party data provider |
| Gemini | The real Gemini app, through a third-party data provider |
| Perplexity | Official Perplexity API (Agent API) |
| Claude | Official Anthropic API, with live web search |
| AI Overviews | Licensed search-results provider |
Google AI Overviews data comes from a licensed search-results provider, which records the overview shown on a Google results page for the question.
Model versions
You can tell a model update apart from a real change in your visibility.
How the model version is logged
Each AI app is called with a fixed model setting where the source allows one, and the model name reported back is logged with every answer whenever the source provides it. When a provider changes its model, the change shows in your history.
The blind question rule
A tracking question never names your brand, your domain or your products.
If we named you, the AI app would talk about you because we asked, and the result would say nothing about what a buyer hears.
- Who is askingA buyer in one country, the first country on your list of places.
Other countries are not measured. AI apps answer differently in each one.
- Search firstRun a live web search before answering, not from memory.
And say so on the first line if it could not search.
- The questionThe buyer question, word for word.
- Answer normallyName the products or companies it would recommend, in order.
One line on why for each, using only public information.
- Two closing linesEvery name given, in order. Every source domain used.
The scan reads those two lines, so they are asked for by name.
A small separate set of brand questions may name you. Those check what AI apps say, not whether they recommend you.
Forced live search and the “searched” flag
Every check records whether the AI app actually searched, and the ones that did not are marked.
How the searched flag is read
An answer from memory reflects old training data. An answer after a search reflects the web today, which is what your work can change. The flag is read from the response itself, for example whether a web search call was made or grounding queries were returned.
How often we check
Weekly on your priority questions, 30 and 90 days on the rest.
Why not daily
Buyers and AI apps do not change their minds overnight.
Why weekly and not daily
The same question can get a slightly different answer minutes apart, but that is random variation, not a trend. The things that move your visibility, a page being indexed, new reviews, a place on a best-of list or a competitor's new page, show up over weeks.
Daily single checks mostly record that day-to-day noise, and they multiply the cost. Checks combine into rates with 95% ranges, and a change is called real only when it beats the noise. You can press Run checks whenever you want, and it is worth doing right after you publish a page or fix a wrong fact.
What we record
Four things per answer, and the same four every time.
| What | What it means |
|---|---|
| Mentioned | Your brand or a known alias, named in the answer |
| Cited | A page on your own domain used as a source |
| Position | Where your name first appears, counting from 1 |
| Also recorded | Competitors named, and every source domain |
A vague description does not count as a mention. Checking each answer against your fact sheet for wrong claims is on our roadmap.
95% ranges
A rate is never published without the band around it.
The Wilson score interval. Simpler methods give ranges below zero or above a hundred on small samples.
Calling a change real
A change is called real only at the 95% level. Anything less reads “Not distinguishable from noise”.
How a change is tested
Two periods, or the questions we worked on against the ones we left alone, are compared with a two-proportion z-test. The test has to clear a z score of 1.96 either way. Below that we keep measuring and say so.
Experiments compare questions your work targeted with similar questions it did not. If both groups rise, the cause was probably something outside your work, such as a model update.
Reusing a recent answer
Same question, same AI app, same country, matched exactly.
Each window is half that priority's check frequency. Each customer is scored separately against the one answer.
The rules we work by
Five rules, and they are the same for every customer.
| Rule | What it means in practice |
|---|---|
| Where we check | The real ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and nothing simulated. |
| Honest numbers | Every rate carries a range, and an unmeasured week is shown as a gap. |
| Nothing ships silently | A day to read everything before it goes live, and to stop it. |
| Reviews | Real ones only. Never written by us, never bought. |
| Unconfirmed facts | Marked for you to confirm, never invented to fill a page. |
Known limits.
Stated as facts rather than hedges, and in the open rather than behind a disclosure.
Logged-out sessions
Read from the real apps, logged out, in your chosen country. A logged-in user with chat history can see something different.
We cannot reproduce every user
Real users see answers shaped by location, language, history and account. We set a country, but we cannot reproduce every user.
The same question can answer differently
A minute later it can come back changed. That is why we use many checks and show ranges, not single screenshots.
AI Overviews come through a provider
This data comes through a licensed search-results provider, not from Google directly. Coverage and timing depend on that provider.
Timing is typical, not guaranteed
New pages typically show on Perplexity in 4 to 8 weeks and on ChatGPT in 2 to 4 months. That is typical, not guaranteed.
The questions are a sample
Results are only as good as the questions tracked. Written blind and by buyer job, they are still a sample.
Questions about the method? Talk to us. See also what the product does and the proof.
You have read how we measure. Want it run on your business?
Same method, your site, your buyers' questions. No card needed, and the free plan stays free. Pay only when you want the work done.