AI Product Recommendation Intelligence

Methodology

What Gengine measures, how it measures it, and the limits you should keep in mind when reading a report.

Definitions

Observation: One recorded answer to one prompt on one repetition. A free diagnostic collects 5 prompts x 3 repetitions = 15 observations.
Mention: The product or its registered alias appears anywhere in the answer text, without being put forward as a choice.
Recommendation: The product is named as a suggested option or pick in the answer, not only referenced in passing.
Sample denominator: The number of successfully collected observations. Failed or excluded calls are never counted in the denominator; they are reported separately as sample completeness.
Citation: A link returned by the provider inside the answer. When no link is returned we display 'No source provided'. Gengine never invents sources.

Prompt design

Prompts are written as neutral consumer buying questions, in Portuguese for the Brazilian market, with no brand names in the question. You review and can edit every prompt before a diagnostic runs.

Limits

Observations come from public AI assistant APIs. They are not measurements of the consumer ChatGPT interface, and results may differ from what an individual user sees.
Counts describe what appeared in a sample of answers. Gengine does not claim to explain why a model produced an answer, and makes no causal claims.
A free diagnostic uses a single provider and 15 observations. It is a directional signal, not a market share measurement.
Audit findings are marked as confirmed error, opportunity or unverified, and never guessed.

Report metadata

Every report records provider, model, mode, country, analysis language, repetitions, run timestamps and the failure list, so any number can be traced back to its source answer.

Four separate measures

1 · GEO readiness

geo-readiness-1.1.0

Deterministic rules over the product page (JSON-LD, OpenGraph, visible text). Weights: identity 20, specs 25, use-case 15, commercial 15, machine readability 15, trust 10. Score = earned ÷ evaluated applicable weight. Unknown checks are never counted as failed; below 70% evaluated coverage no overall score is shown. On marketplaces, platform-controlled markup is not applicable to the seller.

2 · AI recommendation rate

observed

Observations in which an assistant recommended one of your exact catalog products ÷ valid completed observations. Failed attempts are excluded from the denominator but shown as coverage. Mentions, citations and recommendations are counted separately, as are branded and unbranded prompts. Publishers, retailers and generic category phrases are never ranked as products; unmatched names stay 'unresolved'. Observations come from provider APIs, not the consumer chat interface; any Google AI Mode proxy would be labeled an estimate.

3 · Sales readiness

sales-readiness-1.1.0-heuristic

Heuristic with unvalidated weights: commercial content 25, price clarity 15, stock evidence 15, delivery & returns 15, verified trust evidence 15, purchase path 15. Same unknown/coverage rules as GEO. A page scan cannot prove conversion, price competitiveness or seller reliability, and this is not a probability.

4 · Conversion forecast

unavailable

Not produced until connected, validated historical data and a validated model exist. Observed conversion (orders ÷ product sessions) is shown on Sales Performance from your own uploaded data only.