Methodology.
How we measure what AI engines say about a brand, and how we tell a real change from noise. The examples use an invented home battery brand.
The engines
We test 6 engines: ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek. Each is called with web search switched on where the engine supports it, because that is how your buyers use them, and each on a fixed model version, so a re-test compares like with like.
The questions
We build the question set from how your buyers actually ask: hundreds of questions per brand, grouped by topic and by buying stage, and agreed with you before the first run. The same set is asked on every run, so the numbers move only when the answers do.
- Discovery
- “How big a home battery do I need to back up my whole house?”
- Shortlist
- “Which home battery is most reliable for whole-home backup?”
- Comparison
- “Voltaro or Helioma for a 15 kWh system?”
- Alternatives
- “What are good alternatives to Voltaro?”
What we record
Every answer is stored in full, with the sources it cited. For each one we record whether you are named, where, whether you are the first recommendation, which competitors are named, and how you are described. Any number in a report can be traced back to the answers behind it.
The metrics
- Mention rate
- Answers that name you, out of all valid answers. The headline number: how often you are in the conversation at all.
- Top pick
- Of the answers that name you, the share where you are the first recommendation.
- Position
- Where you appear in the answer’s list when you are named. Lower is better.
- Cited sources
- The pages and domains each answer drew on, and which of them mention you. This is where the work is aimed.
- Competitor share
- The same measures for the competitors you choose, on the same questions, side by side.
- Sentiment and accuracy
- How each engine describes you when asked directly, and whether the facts are right.
Change, not noise
AI answers vary from one run to the next, even for the same question. So we never read one answer as a result. The first week sets a baseline across the whole question set; every monthly re-test is compared with it, on the same questions and the same model versions, and we report what moved, by engine and by topic.
What we don’t claim
No one controls what an AI engine says, and models change. We don’t promise a ranking. We promise a transparent method, the full record behind every number, and work aimed at the sources the engines actually use.