AI answers are random: the same model gives two different answers to the same question. So we do not fix the measurement, we fix the way it is measured. This page is the whole of the method in force, and it is versioned.
Every report states in its imprint which version produced it, on which date and with which prompt set. If we change the weights, older reports keep carrying the rules they were produced under.
The total score is the weighted average of these four components
How often, and in which position, AI names the brand in the prompts real buyers use. Measured by asking the AI assistants.
Whether the business AI describes matches what the customer states in writing. Measured by asking the AI assistants.
14 automated checks, robots.txt, llms.txt, schema and server-side rendering among them. Measured without asking anything.
Wikidata record, third-party directories, imprint and expertise signals. Measured without asking anything.
The two purple components are calculated without asking the AI assistants anything; the free scan shows exactly these two, 45% of the total weight. Because they are independent of the engines, they are calculated in full in every package. The two green ones require asking the engines in your package the prompts of your own sector.
These are what make the measurement repeatable
The prompts written for your sector are locked on the first scan and the same ones are asked again in every monitoring re-measurement. If the prompts change, the before-and-after comparison is void; "prompts refreshed every time" is a marketing line, not a measurement.
Looking at a single answer is rolling dice. The score is the share of repeats that name you, and the report always shows it with its variance range: not 71, but 71 ± 4.
Signed-out session, memory off, fixed location. Otherwise what you measure is not the answer the AI gives, but the history of the person measuring it.
In the same scan we also measure 2–3 competitors who are doing nothing. If everyone rose, the model changed; if only you rose, the work worked. This is the only honest way to tell the two apart.
Your package decides how many engines you are measured on · each engine gets its own report
We do not squeeze every engine into a single average. ChatGPT's reason for not mentioning your brand may sit on the index side, while Gemini's may be a missing Google Business or Wikidata record and Claude's that your site is not rendered server-side. Separate causes are reported separately, and the roadmap is ordered accordingly.
What we do, and do not do, if you give us FTP access
Only files are uploaded to your root directory; your site content and code are not touched. If FTP access is not given, the report is still delivered in full and the setup right stays open for 90 days. Because we do not touch the visible face of your site, the setup does not extend delivery time and does not require a review round on your content.
What we do not measure and cannot promise
AI engines change their indexes and answer policies without notice; nobody has the control it would take to guarantee that outcome. What we do promise is this: the measurement before and after is made with the same method and the same prompts, and whatever comes out is reported as it is. If the score drops, we write that too.
AI answers are asked by an operator under clean-room conditions and recorded in the dashboard. The operator does not award the score: mentions, position and competitors are computed by the system, so the measurement stays consistent even when the operator changes.
We do not crawl third-party directories to ask "are you listed here"; that would both break those sites' rules and produce unreliable results. What we measure is the directory links your own site declares.
It is on its way to becoming a standard, but there is no evidence that the AI engines read it. We measure it because it costs one file; its weight inside the Structure component is deliberately kept low.
It looks at sentence length, paragraph structure and the use of lists and tables. A smarter measure would need a language model, and at that point the rule would stop being explainable.
This page is the method itself, not a summary of it. When it changes the version number goes up, and older reports keep carrying their own version.