An AI assistant can recommend your product while the customer buys from another seller. I measure both stages: the suggested model and the merchant included in the buying decision.
For manufacturers and online retailers, brand mentions cover only part of the buying journey. I study product recommendations, seller-selection criteria and merchant recommendations across a stable set of real buyer questions.
This page develops my author-defined approach, originally published in Russian on Habr in September 2026. The indicators are explained below with operational counting rules. Exact definitions are agreed before the first audit.
The five indicators reveal where an AI assistant stops supporting the selection of a specific store. Inquiries and orders are measured separately when business data are available.
This is an operational measurement protocol. Each rate is computed from the received answers in its eligible query set. Branded questions and missing responses are reported separately.
| Metric | Numerator | Denominator | Practical question |
|---|---|---|---|
| PRR | Answers recommending the target model or product line | All eligible answers to product-selection questions | Was the relevant product recommended? |
| SCR | Answers spontaneously adding a merchant-selection criterion | All answers to buyer questions that did not mention merchant criteria | Did warranty, service or delivery enter the answer? |
| RAR | Answers correctly attributing a real benefit to the target retailer | Answers in which the relevant merchant criterion appeared | Was the actual merchant identified as providing that benefit? |
| RRR | Answers explicitly suggesting the target retailer as a purchase option | All eligible answers where selecting a retailer makes sense | Did the retailer enter the shortlist? |
| PPSR | Answers maintaining the retailer recommendation despite a verified cheaper identical SKU | All eligible paired tests with a cheaper offer for the same SKU | How does the answer respond to a growing price difference? |
Each rate is numerator / denominator × 100%. RAR uses a conditional denominator: only answers where the relevant criterion appeared. Absolute counts remain visible, and PPSR is reported by each price-gap tier.
Start with 20–30 product queries that omit store names, warranty or servicing. Keep the original answers, interfaces, dates and sources.
The published article uses a hypothetical example: 3 of 30 AI answers mentioned after-sales service, giving a baseline SCR of 10%. In a later illustration, 10 of the same 30 answers mention it, giving 33.3%. The hypothetical increase is 23.3 percentage points.
These figures demonstrate counting, not observed campaign performance. The next step is RAR: does AI associate the new criterion with the actual retailer?
Compare two offers for an identical SKU while varying the competing price. Any benefits of the specialist retailer must be documented and currently available.
| Price tier | Specialist store | Competing offer | Scenario |
|---|---|---|---|
| Test 1 | RUB 100,000 | RUB 100,000 | Equal price; compare verifiable service |
| Test 2 | RUB 100,000 | RUB 95,000 | Modest price difference |
| Test 3 | RUB 100,000 | RUB 93,000 | Larger difference |
| Test 4 | RUB 100,000 | RUB 90,000 | Significant price pressure |
These are hypothetical prices. PPSR describes the AI answer under controlled conditions. It does not measure what a person is willing to pay; that requires behavioral and sales data.
Record exact queries, language, region, date, AI interfaces and counting rules. New questions form a separate series.
Verify model code, variant, seller, current price, stock and terms. A category page does not verify the price of an individual item.
Separate a brand mention, product recommendation, merchant recommendation and cited URL. Multiple products listed in one answer still form one answer.
Store full responses and source links. Report expected versus received counts and refusals.
Compare referrals, inquiries, orders and available revenue data separately. Changes in PRR or RRR alone cannot prove sales growth.
An audit should lead to prioritized catalog and buyer-scenario work. Product/Offer data, current stock, retailer terms and independent evidence are reviewed according to where recommendations fail.
Technical implementation: buyer-intent pages, Product/Offer data and catalog feeds. Commercial application: GEO for online stores and brands. General methodology: BMR and Share of Voice.
The full methodology was published in Russian in September 2026. The Habr edition examines product selection, seller advantages and price-gap tests. Oborot.ru offers a practical edition of the same research.
Metric concepts are drawn from my published article. The numerators, denominators and sampling rules on this page are an operational GAEO project protocol, not an established industry standard.
BMR tracks whether a brand name appears. Retailers also need to know which models are suggested and whether an AI answer recommends their store.
For the conditional RAR measurement, report the indicator as not calculable for that series. Show SCR and the count of answers in which the criterion appeared.
The controlled PPSR test needs offers for an identical SKU, including the correct variant. Different models belong to a separate research question.
PPSR evaluates AI answers in a controlled test. Actual willingness to pay and purchasing behavior must be studied using business analytics.
Yes. First track brand and product recommendations, then the choice of a seller on a marketplace or an official store. Define the matrix for the business model.
I will review buyer queries, actual AI answers and merchant-selection criteria.
Request a mini-audit