How we measure AI citations

The raw material is a stream of answers from five AI assistants: ChatGPT, Claude, Gemini, Grok and Perplexity. They reply to the questions used every week to monitor how client websites appear in AI answers, sent through each model's API with web search switched on.

The questions are generated from a short profile of each tracked business: what it sells, in which market, to whom. They read like a prospective buyer's requests, for example which removal company to hire for a move across France, or whether a type of online pharmacy can be trusted. Generated questions never name a brand, so the answer is not steered toward the client. Clients can add their own.

The model is asked to flag each company it recommends and to write, next to it, the addresses of the pages supporting the choice. Each address is reduced to its domain, without the www. prefix, and becomes one citation for that domain. Only answers that came back complete are counted.

The unit is a rate: the citations of a domain, divided by the answers in the same scope (period, model, language, industry), times 100.

The denominator includes every answer in that scope, even those that recommend no company or give no address. A domain is measured against all its chances of being cited.

The numerator follows one rule: a domain counts once per recommended company and per answer. If an answer recommends three removal firms and backs each with a page from the same review site, that site receives three citations. Two pages of that site for one firm still make one citation for that firm. The figure is therefore a rate of citations, not a share of answers.

Over several months, the rate is not an average of monthly rates: it divides the citations of all the months by the sum of their answers, so a busy month weighs more than a quiet one.

• Mostly French, some English. The language of an answer is that of the market the analysed website targets, read from its locale (fr-FR, fr-BE, fr-CH for French; en-US, en-GB for English) rather than guessed from the wording. Most tracked websites address French-speaking markets, so the French sample is far larger. The few answers for markets in other languages count in the overall figures only.
• Questions about businesses and markets. Choosing a supplier, comparing offers, checking prices or reliability. Topics outside commerce are absent.
• Industries. An answer belongs to the industry of the analysed company, defined by what it sells rather than by its customers: software for estate agents is filed under Software, a web agency under IT Services. This classification starts in July 2026. An industry gets its own page once it gathers at least three companies and around 500 answers, so that no sector mirrors a single client.
• Models. ChatGPT, Gemini and Perplexity receive the questions of every tracked website; Claude and Grok only part of them, so their figures rest on fewer answers and a different mix of businesses.

• Google's domains. google.com and its country versions, the vertexaisearch.cloud.google.com redirects found in Gemini answers and technical hosts such as gstatic.com are removed at every stage: they are the plumbing of the answers, not pages a reader would consult. Google products recommended by name, such as Google Cloud, can still appear in the brand rankings.
• The publisher's own website. sorank.com is tracked like a client, with questions about SEO software that would pull the English figures toward SEO tools. All its answers are removed. For a few months of early 2026 they could no longer be isolated, so those months count in French only.
• Client websites. The tracked companies' sites disappear from all rankings, whether they show up as a domain or as a brand name, even when cited only through a subdomain. Their answers stay in the sample: the sites below move up and the denominator does not change. A public platform a client gave as its address, such as a LinkedIn page, is never removed for that reason.
• Text and personal data. No question, answer, client name or user data is published.

Raw answers are kept for a few weeks only. Each month is aggregated while they still exist: totals per model and language, a complete series for every tracked platform, and top lists of domains (the first 25 of the month overall, the first 20 per model). Older months survive only as these aggregates.

A ranking over the whole period adds up the monthly top lists. A domain that missed a month's list counts zero for that month, even if it was cited. Leading domains appear in nearly every list, so the top of a ranking is reliable; further down, positions and counts are minimums, especially for domains with an irregular presence.

Tracked platforms such as YouTube, Reddit, Wikipedia or Trustpilot escape this limit, because their citations are recorded in full every month. When a platform page and a domain ranking disagree over the same period, the platform page holds the complete figure.

Each address is attached to its registrable domain, the name a company actually buys, identified with the Public Suffix List, the shared reference of domain extensions. help.example.com is counted with example.com, entreprendre.service-public.gouv.fr with service-public.gouv.fr. The detail is kept: a merged row lists its subdomains and their counts in a drop-down list.

Two kinds of domains are never merged, because their subdomains belong to different owners:

1. Site builders and hosts where each subdomain is a separate website, such as blogs on substack.com or wixsite.com.
2. Institutional umbrellas where each subdomain is a separate public body, such as nih.gov, europa.eu or admin.ch.

Zones such as gouv.fr or gov.uk act as extensions in that list, so impots.gouv.fr and service-public.gouv.fr remain two sites. An old address that redirects to a new one is merged into it (service-public.fr into service-public.gouv.fr), and tracked platforms gather their addresses: twitter.com and x.com count as Twitter/X.

One side effect: if one answer backs a company with two subdomains of the same site, the merged domain row counts both, while the platform figures count the platform once.

Questions run every week; the site is refreshed with a monthly snapshot, whose date is displayed on the site. Each snapshot adds the latest month without erasing the history: when two versions of a month exist, the one with more answers wins, so a partial month never replaces a complete one.

A change of method, such as a new grouping rule, reaches every page at the next snapshot, past months included as far as their aggregates allow.

The data is compiled by Thibault Besson-Magdelain, founder of Sorank. All aggregates behind the charts sit in one file, data.json, at the root of data.sorank.com: domain and brand rankings overall and per model, monthly platform series, answer totals by month, model and language, and industry breakdowns. Counts and rates, nothing else.

• Not a census of AI usage. The data shows which sources AI assistants cite about companies and markets, not what the public asks them in general, and sectors with more tracked websites weigh more.
• French-heavy. Overall figures lean toward sources that matter in French-speaking markets; English rests on a smaller base and other languages have none. The results do not transpose as they are to other countries.
• Models change. Providers release new versions and adjust their search tools. Figures are grouped by assistant, not by version, so a shift in a curve can come from a model update as much as from the web.
• Answers vary. The same question asked twice can return other companies and other sources. Volume smooths this out, but narrow scopes (one model, one industry, one month) move more.
• API answers, unchecked addresses. Public chat apps, with their own settings, may cite other pages. Cited addresses are not verified one by one: a page can be broken or not say what the model attributes to it.