A funnel of answer evidence

The KI-Empfehlungstrichter asks how recorded AI answers progress from addressing a category to mentioning a relevant brand, recommending a catalogue product, and recommending the retailer. These are stages in answer content, not sessions of shoppers moving through a checkout. A drop is an evidence gap to inspect, not lost orders or measured revenue.

The four stages and brand scope

The reviewed extraction version is recommendation-funnel-v1. The brand stage can use own-brand scope or stocked-brand scope. A retailer carrying a manufacturer is not automatically that manufacturer, so the scope choice materially changes what qualifies.

StageQualifying observation
CategoryThe answer addresses the applicable product category
BrandAn applicable own or stocked brand is mentioned, according to the selected scope
ProductAn identified relevant catalogue product is positively recommended
MerchantThe own retailer is recommended using supported merchant evidence

Strict membership uses one complete cohort

An answer joins the strict funnel only when all four stage signals are assessed as true or false. Unknown is not false. To reach stage three, that same answer must have passed stages one, two, and three. Every strict-stage rate uses the same fully assessed cohort, which guarantees nested counts. This prevents a product rate assessed on a tiny subset from being drawn as a later stage of a broader unrelated population.

A numerical walkthrough

Suppose twelve answers were requested, ten have complete stage assessments, and two have unresolved catalogue or merchant evidence. The strict denominator is ten, not twelve. If eight pass category, six pass category and brand, four also pass product, and one also passes merchant, the rates are 80%, 60%, 40%, and 10%.

Example
Synthetic strict cohort: 10 fully assessed answers
Category                  8 / 10 = 80%
Category + brand          6 / 10 = 60%
Category + brand + product 4 / 10 = 40%
All four stages           1 / 10 = 10%

Product -> merchant drop: 3 answers, 30 percentage points
Unresolved outside cohort: 2 answers

Independent stage statistics answer another question

A stage may have usable evidence even when another stage is unknown. Independent counts preserve that information and show the assessed base for that stage. They must not be connected as if they were one nested conversion funnel. A product recommendation without an own-brand mention may be a valid independent observation but does not pass an own-brand strict path.

Catalogue and evidence requirements

Product identity, active catalogue mapping, brand lists, text passages, and extraction completeness determine assessability. An incomplete catalogue must not establish a confident absence. Positive evidence includes passage offsets and exact saved text so a finding can be checked. A raw retailer link, citation order, or manufacturer mention does not substitute for a retailer recommendation.

Choosing a useful action

Inspect the largest supported drop and then open the affected answers. A category-to-brand gap suggests a different investigation from a product-to-retailer gap. Check question intent, supported product mapping, retailer wording, and source evidence before rewriting content. Small cohorts warrant caution; the reviewed summary withholds a largest-drop recommendation below five fully classified answers. A visible largest drop still does not prove a causal content problem.

AEO rates, retailer wins, and unclear answers explainedCitations, competitor evidence, and content opportunitiesCatalogue matching, coverage buckets, and the CI Score