How we measure.
Neutral shopping queries, replayed every week on 8 AI assistants, judged against your catalog and shown with their confidence interval.
Each AI is queried in the market's language and currency.
Four principles.
We sample, we don't measure once.
Each query is replayed several times on each AI: the number of answers and the cadence are settings of the brand. Figures are shown with their 95% confidence interval.
Panel queries are neutral and do not name your brand.
We ask what a shopper would ask: a category, a budget, a use.
Web search is on.
To read today's catalog, not the model's memory.
Every answer is judged against your catalog and listings.
Price, stock, features: accurate, missing, distorted or wrong.
The method, in detail.
Updated October 20, 2026 · by the skoup team
What we count
An answer is the text an AI assistant returns to a shopping query, at a given moment. skoup extracts the products and brands it recommends, in the order they appear, and the sources it cites.
A product's shelf share is the number of answers citing it, divided by the number of answers analyzed for the same query, market and period. A product cited twice in one answer counts once. Its position is the first rank at which it appears; the average position only covers the answers that cite it.
When, and how many times
Each brand sets its cadence and the number of answers asked per query and per assistant. By default, on the weekly cadence, each query is replayed 4 times on each assistant, every Monday at 02:00 UTC. On the daily cadence (Growth and Agency plans), it is replayed every day, and the answers of one week are grouped into a single cycle: both cadences read week by week. The more answers, the narrower the confidence interval.
The same assistant can answer the same question differently. That is why a skoup figure is always a proportion over several answers, never a screenshot.
Which assistants, under which conditions
skoup queries 8 assistants: ChatGPT, Claude, Gemini, Perplexity, Grok, Mistral, DeepSeek and Qwen. Each brand picks the ones it measures, within its plan. For each, skoup queries through the API the model the consumer app serves, with web search on: the assistant reads today's web. The answer can differ from the app's, which adds its own context.
Each market is queried in its language and currency. Cited products are extracted by a separate model, which is never one of the assistants measured.
Neutral shopping queries
The query panel skoup proposes asks what a shopper would ask: a category, a budget, a use. These queries do not name your brand, to measure whether the assistant thinks of you unprompted.
You can add your own queries, including comparisons that name a product. They are measured the same way, and stay identifiable in the reports.
Where our queries come from
The suggested panel is written by a language model from your store: the catalog's most populated categories, with three titles as examples, then what your site sells that the catalog does not cover yet. The model also receives your audiences, your buying criteria and the competitors your site names, and writes in the market's language. Any query that contains your brand's name, a proper noun read on your site or, for a store, a title from your catalog is dropped.
These queries are probes, not queries with a known volume: skoup does not know how many people typed exactly that sentence, and never shows such a number. You see each of them, you edit or disable it without losing its history, and each keeps its origin: suggested by skoup or added by you.
Two checks link the panel to real people. In Results, the effect of a fix is read against control queries, those of the same market that no published fix touched over the same weeks. And next to the visibility measured, skoup shows what people did: visits coming from assistants, orders and impressions in Google's AI answers, taken from Search Console's "Generative AI" report, never added to visits.
The confidence interval
Each share is shown with its 95% confidence interval, computed with the Wald method: ± 1.96 × √(p × (1 − p) / n), where p is the observed share and n the number of answers. With 256 answers and a 34% share, the interval is about ± 6 points.
A change smaller than the interval is not a movement: it is noise. skoup applies a floor to the variance: the interval therefore exists even when the share is 0% or 100%, because a small sample never proves a certainty.
Judging against your catalog
The prices and availabilities assistants state are compared with your catalog. The reference price is the price range of the product's variants; a former struck-through price is not a truth, and a price stated in another currency is never compared. An availability is only judged when the stock is known.
Features are compared with the product's identity sheet. Each element gets a verdict: accurate, missing, distorted or wrong. A verdict always quotes the words of the answer. A feature is only called invented when the answer states it with a value, and it appears neither in the sheet nor in the product's title and description.
Attributing orders
An order is attributed to an assistant when the visit that converted came from it: skoup reads the referring site of the session, then the utm_source parameter of the landing URL. It is a last-touch attribution, as the store knows it.
No personal data about your customers is kept: no name, no e-mail, no address. About your customers, only the country of the order is kept. What a fix changed is read by comparing its target queries with the queries it did not touch, over the same weeks: skoup says “since” or “after”, never “thanks to”.
What the measure does not say
The assistants' mobile apps do not always send a referring site: attributed revenue is therefore a minimum. Attribution does not follow a journey across several visits, and amounts are not converted from one currency to another.
The answers measured are those of a user with no history or personalization. The raw text of the answers is kept 16 weeks; the measures and the evidence drawn from them are kept beyond.