Home / Method
How we measure AI visibility
Ask an AI assistant the same question twice and you can get two different lists of companies. A single screenshot proves nothing. This is how we measure instead.
1. A fixed set of buyer questions
We agree the questions with you in writing before we measure anything, and we keep them fixed so results can be compared over time. They fall into four groups:
| Group | Example | What it tells you |
|---|---|---|
| Category discovery | “Who makes benchtop FTIR spectrometers for polymer QC?” | Whether you are considered at all |
| Comparison | “Brand X vs Brand Y for teaching labs” | How you are positioned against named rivals |
| Specification | “Which encoders have 23-bit resolution and IP68?” | Whether your specs are known and correct |
| Brand | “Is [your company] a reliable supplier?” | What AI says when buyers already know your name |
Discovery and specification questions never contain your company name. Including it would make the result look better than what a real buyer sees.
2. Repeated sampling, reported as rates
Each question runs several times on each assistant, at a recorded date, location and model setting. We report how often you appear, for example “named in 4 of 15 answers”, never a single “position”. Small samples get wide margins, and we say so.
We keep assistants separate. ChatGPT, Perplexity and Google AI draw on different sources and behave differently, so blending them into one score would hide what to fix.
3. Sources, not just mentions
For every answer we record the pages it cites and sort them by type: your site, competitors, distributors, directories, trade media, forums. This “source map” shows where the assistant is getting its view of your category, and where you are missing.
4. Accuracy against your own data
We compare what assistants say about your products with your datasheets and price lists: specs, compatibility, lead times, certifications, regions served. Each error is logged with the answer that contained it and, where possible, the page it came from.
5. Baselines and control questions
Before we change anything we run a baseline, usually two to four weeks. We also leave some questions or pages untouched as a control group. If everything rises by the same amount, that was the platform changing, not our work. Control groups in this field are imperfect, because AI systems change all the time, so we treat them as a sanity check rather than proof.
6. Your own data alongside ours
Sampled answers are one signal. We also look at signals from your side, with read-only access:
- Google Search Console and Bing Webmaster Tools, including Bing's AI citation reporting
- Referral traffic from AI assistants in your analytics
- Branded search volume
- Inquiry forms that ask “How did you hear about us?”
7. What our numbers can't tell you
Answers gathered through official APIs and data providers can differ from what a logged-in user sees in the app, which may be personalised. Assistants change models and sources without notice. A rise in mentions does not prove a rise in sales. We write these limits into every report, next to the numbers.
What we won't do
- Write fake reviews, or ask only happy customers to review you
- Pay for placements that pose as independent editorial
- Publish self-serving “best suppliers” lists that name you first
- Mass-produce AI-written pages
- Hide instructions for AI systems in your pages
Besides being dishonest, these tactics breach search engine spam policies and can end up harming the client. We also never scrape AI products against their terms. Sampling uses official APIs and licensed data providers.