1. Start with the decision
Before choosing a single question, we define the commercial decision being studied:
| Field | Example |
|---|---|
| Industry | Health insurance |
| Market | The market being studied |
| Buyer | A family comparing cover |
| Trigger | A change in circumstances |
| Job to be done | Find suitable health cover |
| Decision | Which providers to look at |
| Brands | The providers in scope |
| AI platforms | ChatGPT, Gemini, Claude, Perplexity |
2. Build the question set
Questions come from how buyers actually talk, not from one researcher at a desk:
- Sales and service conversations
- Customers' own language
- Search and query data
- Community and forum discussions
- Review by specialists in the sector
| Study | Size | Used for |
|---|---|---|
| AI readout first read | 10 questions, 1 run | A first read, clearly labelled |
| Answer Share report | 25 questions, 3 runs | A baseline with ranges |
| Audit | 60 to 150 questions, 3 runs | Full diagnosis and action plan |
| Tracking | Agreed set, 3 runs, monthly | Change over time |
Every question is tagged by buyer, market, buying stage, job, question type, whether it names a brand, and commercial importance.
3. Collect the answers
Every question is asked on every platform three times (once for an AI readout first read) within a defined window. For each answer we record the date, platform and interface, the exact question, the full response, citations, the brands named, the order they appear in, and how each is described. Reports state exactly what was collected, for example: “960 responses collected between 5 and 9 October 2026.”
4. The measures we report
| Measure | What it tells you |
|---|---|
| Answer Share | How often you are named across eligible questions, with a 95% range. |
| Answer Chain | For each question: who was named, at which stage, and which sources the answers relied on. |
| Answer Gap | Questions where others are named and you are named in fewer than half the answers. |
| Accuracy Check | Each statement about you judged against your approved fact register and graded Critical, Major or Minor. |
| Answer Grid | Your own pages placed on AI performance and Google Search Console performance. |
| Source actions | The Fix, Claim, Correct, Compete or Watch label for every source, ranked by stage, reach and risk. |
5. Trace the sources
Counting mentions is not enough; the useful question is why a brand appears. We classify every cited source:
- Your own site
- Competitor sites
- Retailers and marketplaces
- Review and comparison sites
- News and editorial
- Blogs and creators
- Communities and forums
- Reference sites
- Government and regulators
- Social and video
Then we look at how often each type appears, how varied and how recent the sources are, and which sources mention you and your competitors.
6. Ranges and the change rule
Every share is shown with a 95% range worked out from the number of answers behind it. An AI readout first read, with one run, has wide ranges and is labelled as a first read. When we re-measure, we use the same questions, assistants and number of runs, and a movement counts as real only when the new figure sits outside the previous range. Interpretation is labelled as interpretation.
This is why we write “in 18 of the 40 answers that named Harbourline (fictional), it was described as innovative”, not “AI sees Harbourline (fictional) as innovative”.
7. What we don't do
- We don't publish a single visibility score out of 100, and we never show a share without its range.
- We don't present explanations as facts. Interpretation is labelled as interpretation.
- We don't claim that visibility alone caused a sale.
- We don't invent data. Illustrative examples on this site always use fictional brands and are labelled as such.
What are the limits of this method?
AI answers vary with time, location, account settings and the interface used, and models change without notice. We control what we can, document what we cannot, and report the collection window with every result. A measurement is a snapshot of a moving system; tracking the same questions over time is what makes it useful.
How the change rule works
Before
After
40% sits outside the previous 29% to 35% range: the change counts as real.
Frequently asked questions
Which AI platforms do you test?
ChatGPT, Gemini, Claude and Perplexity as standard, with others added where they matter in a particular market.
Why ask each question more than once?
Because the same question can produce different answers. Three runs per question let us show a range, which separates consistent patterns from one-off results.
Do you use any AI to analyse the answers?
Yes, with checks. An AI judge decides whether a brand was presented as an option and quotes the passage, and a person reviews a sample. Question design, the Accuracy Check verdicts on regulated facts, and the conclusions are made by people.