Two sources, both ours: anonymised analysis of customer dialogs held by AI assistants on business websites, and the behavioural telemetry of those websites. Both are collected under the privacy policy of the service and processed only in aggregate.
A company is a root domain: branch subdomains and sections of one website count once. Our own websites are excluded from every cohort.
A niche is published when at least 10 companies and 300 analysed dialogs stand behind it, and the largest company supplies less than half of the dialogs. Each block has its own floor: at least 100 dialogs from at least 5 companies, again with no company above half. A block below its floor is not shown rather than shown on thin data.
Company names, domains, personal names, phone numbers and addresses. Customer phrases are grouped into topics by a language model instructed to return only generic labels; the grouped phrases themselves are not published.
Dialog figures are per calendar quarter, the first full quarter being Q2 2026. Behavioural telemetry is available for longer and is reported for the same quarter.
Every figure stands next to its sample size. Behavioural figures are medians across companies, so one large website cannot set the number. Shares of phrases and stages are shares of the dialogs where that field was filled, not of all dialogs.
