Methodology / Version 0.2 / August 2026 / Public and versioned
How we measure AI responses.
We sample answers to a defined set of buying questions, keep the evidence and explain how the findings are calculated. We do not inspect a model’s internal memory or monitor private conversations. Every number we publish is produced as described here; when the method changes, the version changes, and results across versions are never compared silently.
On this page
AI availability, defined
AI availability is how easily a brand gets chosen when an AI assistant answers a buyer’s question, and how easily an AI agent can read, verify and buy from that brand. It has two halves, reported separately and never averaged: answer availability, the brand’s share of the answers across the buying situations of its category, from memory and with live search; and agent availability, whether an agent can find, read, trust and act on the brand’s pages. It extends the mental and physical availability of the Ehrenberg-Bass research tradition to the AI layer. The full definition, with the metrics and what it is not.
The machine is the respondent
Single AI answers are unstable: identical questions rarely return the same brand list, and different models often disagree on a category leader. We treat that as the starting point, not as a nuisance. Instead of surveying people, we put the buying questions to the assistant itself, around a thousand times per category, and report the pattern with its uncertainty.
Two settings, reported separately: memory, with browsing off, shows what the model already knows. Search, with browsing on, shows what it finds right now. A brand can be strong in one and absent in the other; that gap is a finding in itself. The tests use the official developer interfaces of ChatGPT, Gemini and Claude; they do not reproduce every feature of the consumer apps, and every report says so.
Buying situations, not random prompts
For each category we derive 8 to 16 buying situations, the moments in which the need arises, phrased the way real buyers ask assistants. This applies the category entry point framework from marketing science to machines. Each situation is asked in several phrasings, because phrasing drives more answer variance than repetition, and each is repeated. For published reports the complete design is frozen before the first question is asked and its hash is kept with the results and shared on request, so anyone can check that nothing was selected after seeing the data. Each study states its brands, language, market, settings, dates and method version.
What we report
- Share of answers
- How often a brand appears in the recorded answers of a buying situation, with a 95% confidence interval. Never “rank”: rank order in AI answers is noise, membership in the answer is the stable construct.
- Share of brand mentions
- The brand’s share of all brand mentions across situations, weighted by situation importance; shares sum to 100% among the tracked brands within a setting. The machine analog of mental market share.
- Cross model agreement
- Reported, not averaged away. Platform disagreement is information.
- Entity integrity
- Wrong or confused brand claims in answers, each with the recorded evidence.
- Citation sources
- Which domains actually feed grounded answers in the category. Observed from citations, not guessed.
- Reliability of our numbers
- Design size, split-half stability and minimum detectable change, stated with every report. When a difference is within noise, the report says so.
Two shares, two denominators
Share of answers counts answers: the percentage of recorded answers in a situation that include the brand, with a confidence interval. Share of brand mentions counts mentions: the brand’s share of all brand mentions, weighted by situation importance, so the tracked brands sum to 100% within a setting. Citation counts are different again. Every exhibit states which one it shows, and the two are never mixed in one chart. Repeated answers and reliability checks help assess variation; they do not establish accuracy, causation or commercial impact on their own.
The machine-facing surface
This is the agent half of AI availability: ten checks of the brand’s public surface, fetched live, covering crawler access, entity records, structured data and its freshness, readability without JavaScript or login, and a plain path to purchase. Content behind registration or social walls is largely invisible to assistants. Walls are findings, not measurement obstacles: what is walled off is, to a machine, barely stocked, and we say so with evidence. Website checks describe the selected pages at the time of inspection.
Evidence and what stays internal
Raw responses and the sources used for factual checks are archived per run, and every report is reproducible from its archive and pinned to a methodology version. The operational recipe, meaning the situation weights, the prompt lists, the matching rules and the check procedures, is internal documentation. Clients receive the parts that concern their own measurement inside their deliverable. What is public is everything needed to judge the results: the definitions, the design summary of each published report, and the reliability of its numbers.
Vocabulary
- AI availability
- How easily a brand gets chosen when an AI assistant answers a buyer’s question, and how easily an AI agent can read, verify and buy from it. Two halves: answer availability and agent availability. Full definition.
- Buying situation
- A moment in which the need for the category arises, phrased the way a buyer asks an assistant. 8 to 16 per category.
- Memory and search
- Two settings of the same question. Memory: browsing off, what the model already knows. Search: browsing on, what it finds now. Reported separately; the gap is a finding.
- Share of answers
- The share of recorded answers that include the brand. Denominator: answers.
- Share of brand mentions
- The brand’s share of all brand mentions across situations, weighted by importance. Denominator: mentions of all tracked brands. The two shares are never mixed in one exhibit.
- Entity integrity
- Whether what assistants say about the brand is true. Each wrong or confused claim is recorded with the answer as evidence.
- Citation sources
- The domains grounded answers actually cite in a category, counted, not guessed.
- Minimum detectable change
- The smallest movement a design of this size can tell apart from noise. Stated with every report so nobody reads a wobble as a trend.
Version history: 0.2 (August 2026) added pre-registered designs, minimum detectable change and conservative matching for brand names that are ordinary words; 0.1 (July 2026) was the first public version. Metric definitions are unchanged between them.