How we measure whether AI chooses you

The unit of measurement: a buying conversation

We do not score single prompts. Each measurement is a probe: one complete, multi-turn buying conversation in which an AI assistant or agent works through a customer request, from the first question to a final recommendation. Customers compare, challenge and refine before they decide, and recommendations change as the conversation develops, so the decision is measured at the end.

What is tested

• Assistants and agents: ChatGPT, Gemini, Claude and Perplexity, run as shopping agents. For enterprise work, AIVO Meridian also uses agents built to replicate the journeys and actions of personal agents such as Meta's Muse, OpenAI's Dots, Instinct and Duet.
• Journeys: realistic customer requests for a defined category and market, agreed with the client for enterprise work, or entered by the buyer for self-serve tests.
• Language: any language is available. Each market and language is scoped separately; the Sprint and the 90-day program each cover one market in one language, with more on request.
• Competitors: the brands the assistant actually recommends instead, recorded on every run.

Two metrics: PSOS at the first prompt, AI Win-Rate at the decision

AIVO measures two points in the buying journey, and many clients use both.

• Prompt-Space Occupancy Score (PSOS): how often a brand appears in AI answers to a large, standardized set of first prompts in its category, with repeated runs under fixed conditions. It is a robust, reproducible measure of presence at the start of the journey, for companies that want a governance-grade first-prompt metric. PSOS is defined under the AIVO Standard.
• AI Win-Rate: how often AI systems select a brand head to head at the decision stage. It is measured as the share of buying conversations, run through to a decision, in which an AI assistant or agent recommends the brand outright. Full definition.
• Presence vs decision: a brand can score well on PSOS and still lose at the decision, which is why AIVO measures the whole path. The cause AIVO identifies is the Linkage Gap: the facts an AI model holds about a brand but does not carry to its recommendation.

From runs to a AI Win-Rate

The full method is published as AI Win-Rate: A Board Metric for AI-Mediated Buying (Methodology v2.0) on Zenodo.

• Repeated runs: AI answers vary between runs, so every journey is run several times on each assistant. AI Win-Rate is the share of runs in which your brand is the final recommendation.
• Confidence ranges: results are reported with 95% confidence ranges, not single scores, so you can see how precise each figure is. More runs give narrower ranges; small self-serve tests are labeled as directional.
• Where you drop out: each run records whether you were found, shortlisted, compared and chosen, and at which turn you were displaced.
• Why you lost: the sources and claims the assistant relied on are recorded, so each loss can be traced to a cause and a fix.

Stability and model change

The model version and web-search state are logged on every turn. When results are compared over time, the same journeys are re-run under the same conditions. In improvement programs, part of the journey set is held back untouched and an untouched market journey set nets out category-wide shifts, including model updates. Lift counts only above the upper bound of the baseline's confidence range.

Keeping measurement separate from improvement

AIVO both measures recommendation and helps brands improve it, so the measurement is protected from the improvement work:

• Fixed before work starts: in improvement programs, the journeys, competitors, baseline, held-back control and success threshold are agreed in writing and frozen before any remediation begins. They cannot be changed afterwards to flatter the result.
• A named method version: scoring follows a named, published version of the AIVO Standard, recorded in every report.
• The same conditions: re-measurement repeats the baseline journeys under the same conditions, alongside the held-back and market controls.
• Results reported either way: findings are reported whatever they show, including no lift. Any performance fee is payable only on qualifying lift: the day-90 AI Win-Rate minus the upper bound of the baseline's 95% confidence interval.

About the figures we publish

• First-turn displacement (87.3%): from a defined study of 1,427 brand probes across ten industries. 87.3% of brands an AI assistant named at the first turn of a buying conversation had been displaced by a competitor by the fourth turn (95% CI 85.5–88.9%). The unit is the probe: one multi-turn conversation about one brand. Sheals and de Rosen, "Beyond Visibility", working paper v1.1.
• The research base (27,252 conversations, 380+ brands, 27 October 2026): the running total of AIVO's measurement across research and public rankings. It is a count of work done, not the sample behind any single figure.
• Banking study (2,160 conversations, 30 US banks): a separate study of small-business banking recommendations, reported by American Banker in September 2026.

Each published figure states its own sample. When a total is updated, the date beside it changes.

Figures last updated: 30 October 2026.

Changes: 30 October 2026: the 90-day program's performance fee, the AIVO Meridian and AI RQ starting prices, and the one-market, one-language scope are now published; the Linkage Gap is now defined as in the published working paper (the facts an AI model holds about a brand but does not carry to its recommendation), and the 87.3% figure is labelled as displacement. 21 October 2026: the 87.3% figure is now stated exactly as in the published working paper (1,427 brand probes, ten industries), replacing earlier wording that attributed it to the wider research base; the brand count is updated to the full research corpus of 380+.

What AI Win-Rate does not tell you

AI Win-Rate measures what AI assistants recommend. It is not, on its own, a measure of customers won or revenue. Whether AI recommendation predicts commercial performance is the question the AI Win-Rate Study is testing with an independent academic, and the findings will be published whatever they show.

How to check our work

• Published methodology: the AI Win-Rate method is published on Zenodo as Methodology v2.0. Definitions, measurement rules and versions are published by the AIVO Standard, with research released openly through AIVO Journal.
• Sample report: a public sample report shows the full output of a real test, including a step-by-step run replay.
• Full logs under NDA: every run is recorded. Full conversation logs and raw outputs are available to clients, and to organizations evaluating AIVO, under a mutual NDA.
• No client influence: clients, including those in improvement programs, have no input into how their scores are calculated or reported.

Frequently asked questions

What is AI Win-Rate?

AI Win-Rate is how often AI systems select a brand head to head at the decision stage. It is measured as the share of buying conversations, run through to a decision, in which an AI assistant or agent recommends the brand outright. It is reported with 95% confidence ranges. Full definition.

How is it different from AI visibility or share of voice?

Visibility tools count how often AI mentions or cites a brand. AI Win-Rate measures whether the brand is chosen when the conversation reaches a decision. AIVO also measures first-prompt presence with PSOS. A brand can score well on PSOS and still lose at the decision.

Which AI assistants and agents do you test?

ChatGPT, Gemini, Claude and Perplexity, run as shopping agents. For enterprise work, AIVO Meridian also uses agents built to replicate personal agents such as Meta's Muse, OpenAI's Dots, Instinct and Duet.

Which languages?

Any language is available. Each market and language is scoped separately; the Sprint and the 90-day program each cover one market in one language, with more on request.

How much does it cost?

The AI Win-Rate Sprint for mid-market brands is $3,500 fixed over 30 days. The AI Decision Risk Review for consumer brands and banks is fixed-price, from $7,500. The 90-day Win Rate program is $9,500 plus $750 per percentage point of qualifying AI Win-Rate lift, capped at $15,000 (maximum total $24,500). AIVO Meridian engagements start from $3,500 with the Sprint, and AI RQ reputation studies from $10,000. Self-serve tests for small businesses and agencies start at $4 a run.

Do you need access to our systems or customer data?

No. AIVO measures what public AI assistants say. Nothing is installed and no customer data is needed.

Can we see the raw conversation logs?

Yes. Every run is recorded, and full logs and raw outputs are shared with clients, and with organizations evaluating AIVO, under a mutual NDA.

Where is client data held?

In AIVO's proprietary database, securely backed up in the United Kingdom and mirrored in the United States. A security and procurement brief is available on request.

Does a higher AI Win-Rate mean more sales?

Not on its own. AI Win-Rate measures recommendation. Whether it predicts commercial performance is what the AI Win-Rate Study is testing with an independent academic, and the findings will be published whatever they show.

Questions about the method: tim@aivocorp.com.