Where do a GEO tool's tracked queries come from?

The objection, and what it gets right

An AI visibility tool asks assistants questions it chose itself, then counts the answers that cite your brand. Hence a fair objection: nothing proves that real customers type those sentences. A score built on made-up questions would only be worth what the questions are worth.

The objection is right on one point: a tracked query is not a query with a known volume. skoup does not know how many people typed exactly "which e-bike to commute to work" into ChatGPT this week, and never shows such a number. Even Search Console's "Generative AI" report gives impressions per day and per page, not the questions asked.

It is wrong about what the query has to prove. This page explains what a tracked query measures, how skoup builds its panel, and what confronts the measurement with real customers.

What a tracked query measures

A tracked query is a probe, not an estimate of demand. It asks a buying question typical of your category, always the same one, and checks whether the assistant thinks of you when answering it.

Two customers with the same need do not phrase it with the same words. What counts is not the exact sentence but the need it carries: a category, a budget, a use, a criterion. That is why the panel asks several queries per category, from different angles: visibility is read over the whole panel, never over one sentence.

It is also why a tracked query never names your brand: a question that names you is bound to cite you, and says nothing about what the assistant recommends on its own.

How skoup builds the panel

The panel skoup suggests starts from your store, not from a generic list:

• Your categories first. The catalog's most populated categories, with three product titles as examples for each.
• Then what else your site sells. The categories, services or destinations read on the site that the catalog does not cover yet, up to eight topics.
• The context of your market. The audiences, buying criteria and competitors named by your site.

A language model then writes short queries, in the market's language, the way a customer would type them: a budget or a constraint, a use or a person, a buying criterion, a comparison between types of products, "which X for…".

A filter then runs over every query. Any query that contains your brand's name, a title from your catalog or a proper noun read on your site (a range, a sub-brand) is dropped.

The panel is not a black box:

• you see every query, word for word;
• you edit it, or disable it without losing its history;
• you add your own, including comparisons that name a product;
• every query keeps its origin: suggested by skoup or added by you.

From one week to the next, the same queries are asked word for word. Without that stability, no curve would be comparable.

Why several runs and a confidence interval

An assistant does not answer the same question the same way twice. A single answer therefore proves nothing, either way.

By default, every query is asked 4 times to each assistant, every Monday. A product cited twice in one answer counts once. Its visibility is the share of answers that cite it, shown with its 95% confidence interval.

An example. With 10 queries, 3 assistants and 4 runs, skoup analyzes 120 answers a week. For an observed share of 30%, the interval is about ± 8 points. On a single query, that is 12 answers, and the interval rises to ± 26 points: the number of one query is a hint, the panel's is a measurement.

When a change stays within the interval, skoup writes "stable" instead of announcing a rise. The formula and its limits are in the methodology.

Control queries: the measurement checks itself

The risk with a panel chosen by the tool is seeing progress where there is only noise, or a general trend of the assistant. So skoup compares itself with itself.

When a fix is published, the Results screen compares two groups of queries on the same market:

• the targeted queries of the fix;
• the control queries, which no published fix touched over the same period.

Both groups are read over equal windows, 2 to 4 measured weeks before and after the publication week, with the same assistants. The gap kept is the change of the targeted queries minus that of the controls. A two-proportion test at 95% decides whether it is measurable; otherwise skoup writes "stable".

If an assistant cites your brand more everywhere, the controls rise too, and the gap stays at zero. This per-fix impact is available from the Growth plan.

What real customers say

A tracked query does not name a person. What links the measurement to real people are the signals they leave. skoup shows them next to visibility, on the same timeline, without ever adding them together:

• Impressions in Google's AI answers. Search Console's "Generative AI" report counts your pages' impressions in AI Overviews and AI Mode, on real searches. It is not in Google's API: you import its export into skoup, which reads it per day and per page.
• Visits coming from assistants. People who clicked a link in ChatGPT, Perplexity or Gemini, measured by our script or by Google Analytics 4. The details are in the guide on AI traffic and sales.
• Orders. For a store, the orders whose visit came from an assistant.

These signals do not say which question was asked. They say whether the visibility measured goes along with people who see your pages, visit them and buy.

What we do not claim

• No volume per query. skoup does not estimate how often a sentence is typed into an assistant.
• No market share of all conversations. Visibility is the share of answers to your panel, on the assistants measured.
• No personalization. The answers measured are those of a user with no history and no personalization.
• Always "since", never "thanks to". A gap measured after a fix is an observation, not proof of cause.

Five questions to ask an AI visibility tool

1. Can I read every tracked query, and edit it?
2. Do the queries name my brand? If so, the score measures what the assistant knows about you, not whether it recommends you.
3. How many answers are behind each number, and is the confidence interval shown?
4. Is there a control group to tell an effect from a trend?
5. Is the measurement confronted with real signals: Search Console, visits, orders?