What was counted
The collection design, denominators, known gaps and limits behind every published figure.
Method
and limits
How the study was run, how denominators are defined and which conclusions the results cannot support.
Six engines, all with full coverage
The same 136 questions were scheduled daily for each of the six retained engines: 25,296 scheduled runs, 25,267 successful runs and 25,264 usable answers. Combined answer results use those 25,264 usable answers. No additional model had partial coverage within the retained panel.
The scheduled answer panel covers the full month
Scheduled runs span all 31 days in each sector. The inventory reports successful answers separately from scheduled runs so incomplete model output is not treated as evidence.
Observation, not causation
The study recorded which brands assistants named and the language used around them. It did not test what caused those results, so it cannot predict what would change after a brand edits its content.
The study covers buying questions only
All 136 prompts ask which product or service is best for a UK buyer. The resulting visibility and language patterns therefore describe buying questions, not all AI searches or every possible user intent.
Positioning labels come from fixed text rules
The report assigns a positioning label only when matching words appear in the same sentence as a brand name. A label describes the language used around a brand. It does not measure sentiment, prove a claim or show that an assistant endorsed the brand. Taxonomy version: prompt-map-v1-2026-08-17.
The report tracks a fixed list of 100 brands
The report follows 100 brands tracked in UK buying questions across six sectors. Every tracked brand appeared in at least one answer. Brands outside the list were not counted, so the figures describe this tracked group, not the whole UK market.
Data inventory
| Sector | Answer days | Successful runs | Status |
|---|---|---|---|
| Fintech | 31/31 | 4,647 | Scheduled answer panel complete |
| SME accounting | 31/31 | 3,717 | Scheduled answer panel complete |
| Legaltech | 31/31 | 3,717 | Scheduled answer panel complete |
| Insurtech | 31/31 | 3,897 | Scheduled answer panel complete |
| Energy / EV | 31/31 | 3,715 | Scheduled answer panel complete |
| Cybersecurity | 31/31 | 5,574 | Scheduled answer panel complete |
What this data
cannot support
The table separates claims this study can support from claims it cannot. Use the wording on the right when reporting the results.
| Unsupported claim | What this study can say |
|---|---|
| "AI recommends X" | X appeared in N% of answers in its sector and received an explicit endorsement in M% of those appearances |
| Recommendation rates by brand | Do not report one. Recommendation data exists for only 0.54% of responses and is excluded from this report |
| "Visibility does not predict endorsement" | Visibility explains 8.3% of the variation in endorsement rate across all 100 brands. Brands with similar visibility can still differ by tens of percentage points in endorsement |
| "Visibility causes caution" | More-visible brands received more caution language in this dataset. The study did not test whether visibility caused it |
| Primary versus competitor comparison | The report does not identify a primary brand. All 100 brands are measured in the same way |
| Consumer-app or AI Overviews performance | This study used APIs. It does not measure consumer AI products or Google AI Overviews |
| Long-term trend | 31 days can show short-term movement, not a long-term trend |
| Factuality or hallucination rates | No claims in the answers were fact-checked, so the report cannot estimate these rates |
| Untracked competitor rankings | Brands outside the tracked list were not counted. Their absence does not show how they perform |
We revise a published period when we correct its data or analysis. CHANGELOG.md records every published correction.