How to read your measurements

Start with the questions asked and the answers retained. Mentions are observations in this sample, not a measure of everything AI models say. The main figure is the citation rate; the three diagnostics that complement it are shown only when they were measured.

• 01

  Questions asked

  The free test plans 5 questions based on the business read on your website. One is detailed without an account; all five become readable after email confirmation without new collection. New Business and Agency plans provide for 10 questions. The report records actual coverage. A comparison retains the reference protocol’s questions.

• 02

  Our reading crawler

  To write the questions, our crawler SellomBot reads your site’s home page, then a few public pages. It identifies itself under this name (page sellom.io/bot) and, if the site refuses that name, tries once more as a browser. It reads robots.txt and reports it: a rule that blocks SellomBot or AI crawlers does not stop the measurement, the report shows it, with the lines to correct.

• 03

  Models queried

  The planned free panel is Perplexity, Gemini flash-lite and ChatGPT mini. Business and Agency share five models: Perplexity, Gemini flash-lite, ChatGPT, Claude and Mistral. The report identifies those that actually answered. Models are queried directly, outside their consumer apps; answers may differ from those apps.

• 04

  Answers retained

  The protocol plans two repetitions per model and question on Free, three on Business and Agency. Collection limits and errors can reduce coverage. Empty or failed answers do not count as non-mentions. Available excerpts, truncation notices and actual retained counts let you check the result.

• 05

  Business mentions

  Detection looks for the brand name and domain in the answers. A mention is not necessarily a recommendation. Read the number of mentioning answers alongside the number retained. Common names, variants or brands that differ from the domain can cause errors: check the excerpts.

• 06

  Available sources

  Some configurations request web search. This describes the request sent to the model, not proof that a search occurred. The report only presents sources actually returned and usable. No available source does not mean that no website mentions you.

• 07

  The citation rate and the three diagnostics

  The report’s main figure is the citation rate: the share of answers obtained with web search that cite your business, as a percentage. Three diagnostics complement it: presence in answers (out of 40), sources that cite your name (out of 30), accuracy of the description (out of 30). They do not add up to it. A diagnostic that was not measured is shown as such; a measured zero is different from missing data.

• 08

  Reading a first result

  Take the first detailed question. Read the excerpt attributed to its model, then the count of answers mentioning your business. Check that the question matches a service you offer. If it does not, establish that before deciding to change your website.

• 09

  Corrections and publication

  A dossier can propose up to three corrections in total, linked to a page and a real question. Insufficient evidence means no invented correction. Declaring a change published is separate from a dated observation of its static HTML. That observation does not certify JavaScript rendering, indexing or an effect on AI answers.

• 10

  Comparing measurements

  Comparison requires a compatible recorded protocol: the same questions and required conditions. It covers lexical mention rate, with differences expressed in percentage points. It does not calculate a causal effect of changes. A new measurement collects new answers and uses the applicable quota.

• 11

  Limits and languages

  Models change and their answers vary. Sellom does not measure your traffic, sales or market share, and promises no gains after a correction. These repetitions do not provide a statistical margin of error. Journeys are available in French, English, Spanish and Portuguese; a dossier keeps its measurement language.

Description dated 15 October 2026. Each report retains the details of its own measurement.