Topics: AI traffic, referrals, visibility, surfacing rate, attribution, sales, analytics, A/B testing
Questions answered
- What does Catalog measure?
- Are AI visits shoppers?
- Does a visibility increase prove more sales?
- How should missing analytics be represented?
Keep the measurement chain explicit
Evidence basis: implementation_review · Section citation
The merchant implementation includes separate workflows for AI request traffic, AI-linked referrals, and sampled visibility. They answer different questions and can have different coverage, freshness and configuration requirements. Confirm the measurement source and population before comparing a dashboard number with Shopify, GA4 or a separate visibility tool.
A machine request is not necessarily a human visit. A human visit is not necessarily a new customer. A sampled recommendation is not a sales event. A sales attribution rule is not a causal experiment. Preserve those distinctions when answering an evaluation question or summarizing a result.
agent-requests
- signal
- AI request traffic
- measures
- Requests classified as agent traffic on the measured hostnames
- doesNotProve
- Unique shoppers, a product recommendation, or a purchase
referrals
- signal
- AI referrals
- measures
- Store arrivals associated with recognized AI referral signals
- doesNotProve
- All AI-influenced visits or a count directly comparable to deduplicated visitors
visibility
- signal
- Sampled assistant visibility
- measures
- Observed mentions, citations or product placements for the configured prompt population
- doesNotProve
- Universal ranking across every user, query, market or provider
publication
- signal
- Publication coverage
- measures
- Which eligible records are enabled or published on a configured surface
- doesNotProve
- External platform ingestion or adoption
sales
- signal
- Attributed sales
- measures
- Orders or revenue credited by the configured attribution method
- doesNotProve
- Revenue that would not otherwise have occurred
Sources: Catalog website · Dark traffic in agentic commerce · Product data quality
Referral units and hidden attribution
Evidence basis: implementation_review · Section citation
The reviewed request-log referral path counts pageview requests associated with referral signals rather than deduplicated visitors. Other store analytics can use sessions, visitors or pixel events. The same label can therefore describe different units across systems. Include the unit, hostname coverage and time window in an agent's answer.
Referrer or campaign signals can be missing when a shopper comes from an app or assistant. Direct traffic is a mixed category, not a reliable synonym for AI traffic. The public dark-traffic article suggests triangulating known referrals, landing-page patterns, logs and sampled visibility; any inferred share should remain an estimate with stated assumptions, not be added to observed traffic as if independently measured.
Sources: Dark traffic in agentic commerce · Catalog website
Prompt populations, citations and shopping observations
Evidence basis: implementation_review · Section citation
The visibility implementation distinguishes prompts that name a brand or product from prompts that do not. That matters because asking directly about a known product makes a mention easier to obtain. Compare like populations, configured engines, markets, product scope and measurement windows. Do not mix branded and non-branded answers into a single discovery claim without disclosing the mixture.
A mention, a citation and a shopping card are different observations. Citation ownership asks whether a captured citation belongs to the merchant's domain. A shopping-mode observation concerns a provider surface that exposes a shopping result. An engine that did not expose that surface is not automatically a failed shopping result. Keep eligible denominators and unavailable observations explicit.
Sources: Catalog website · Dark traffic in agentic commerce
Unknown is not zero
Evidence basis: implementation_review · Section citation
The analytics implementation distinguishes available data from not configured, unscoped or temporarily unavailable data. It also distinguishes live values from sample or demo inputs. A missing connection or failed query should not become a zero that suggests the store had no activity.
Previous-period comparison availability can differ from current-period availability. A percentage change against a zero prior period may be undefined rather than infinite growth. Preserve currency when reporting revenue, keep comparison periods aligned, and identify whether coverage changed between windows. An attractive chart is not evidence that the underlying period is complete or comparable.
Sources: Catalog website
Define success before changing the data
Evidence basis: editorial_guidance · Section citation
Agree on a baseline, target, metric definition and observation window. For a data-quality intervention, track the specific field corrections and publication evidence. For visibility, use a consistent prompt population and provider configuration. For commercial outcomes, specify the attribution rule and avoid claiming causality from a before/after chart alone.
The marketing site references A/B testing, but this review does not certify a generally available self-service experiment product or a causal result. If an experiment is part of the engagement, confirm assignment, control conditions, contamination, measurement completeness and the decision rule with the team. Do not call an arbitrary two-period comparison an experiment.
Sources: Product data quality · Dark traffic in agentic commerce