Agent Commerce Evidence: what merchants can measure.
A practical guide to the request-level signals that AI shopping agents leave behind, why most analytics stacks miss them, and how Cartograph is being built to close that gap.
Published · Updated
What is agent-commerce evidence?
Agent-commerce evidence is the set of request-level signals that show an AI-mediated shopper interacted with your catalog, storefront, cart, or checkout. Unlike traditional web analytics, which relies on JavaScript execution in a browser, evidence starts at the origin: configured server, CDN, and application logs can capture requests that reach those observation points even when a browser-side pixel does not fire; coverage is not complete by default and depends on source inclusion, routing, sampling, retention, and collection health.
The key observation is simple: an AI agent fetching a product page, adding an item, or completing a delegated checkout still issues HTTP requests. Those requests carry user-agent strings, referrers, tokens, and timing patterns. When captured at the log layer and labeled consistently, they become evidence.
Use the provider-documented AI crawler and agent directory to distinguish automatic crawlers, user-triggered fetchers, and publisher control tokens before labeling request evidence.
Cartograph is being built to collect, normalize, and label that evidence so merchants can answer questions their current stack cannot: How much of my traffic is agent-driven? Which requests are unclassified? Where do agents drop off between catalog and checkout?
Adobe Digital Insights reported that AI-referred traffic to U.S. retail sites grew 393% year over year in Q1 2026. In March 2026, those visits converted 42% better and generated 37% higher revenue per visit than non-AI traffic. Adobe says the underlying U.S. retail analysis covers more than one trillion visits.
Adobe-reported historical data; not Cartograph customer data. Growth rate does not indicate channel share.
Your stack sees the pieces. The planned integration connects the evidence.
Cartograph is not a bot detector, PIM, analytics platform, PSP, or commerce platform. It is being built as the agent-commerce evidence layer across them.
Readable, Consistent, Transactable.
This is a readiness framework, not an ACES taxonomy, classifier, confidence model, grade, or score. Each category names the merchant surface, the observation method, and its limitation.
Can the defined scanner or collector fetch and parse the named merchant surface?
Boundary: not external-model understanding, citation, preference, or recommendation.
Do named merchant-authorized surfaces materially agree within a defined observation window and comparison policy?
Boundary: not external-agent trust, belief, confidence, preference, rejection, or consideration-set placement.
Do merchant-observed events or a separately authorized synthetic diagnostic show progression through the exact cart or checkout milestones in scope?
Boundary: not complete-journey proof, order completion, payment authorization, provider support, or conversion.
Evidence, not guesswork.
The v0.1 specification defines deterministic traffic labels and evidence codes. It does not infer intent, score actors, or treat missing agent evidence as proof of human traffic.
The v0.1 specification defines five traffic labels: verified agent, declared bot — verified, declared bot — unverified, AI-referred, and unclassified. Those labels describe available evidence at capture; they do not assert that unclassified traffic is human.
What your stack sees — and what it misses
Most merchant stacks were designed for human shoppers. Web analytics needs a browser and a consenting user. Bot management looks for abuse, not attribution. Product information management systems know what you sell, but not who asked. Payment processors see transactions, but rarely the upstream agent that initiated the journey.
The result is a blind spot: a measurable share of eligible storefront request events remains unclassified after deterministic labeling. Closing that gap requires combining log-origin request events with a labeling model that recognizes provider-documented agents and agent-like behavior.
How Cartograph plans to connect the evidence
Cartograph's planned architecture treats merchant surfaces as event producers: feeds, storefronts, carts, checkouts, and payment flows each emit request events. The platform is being designed to ingest those events from server-side logs, apply a v0.1 evidence label set, and surface an Unidentified Traffic metric: the unlabeled share of eligible request events.
Over time, the goal is to add deterministic labels for provider-documented agents, heuristic labels for unknown but agent-like fetchers, and merchant-owned signals such as signed checkout tokens. The result is a shared evidence layer that analytics, product, security, and payments teams can all reason about.
Agent-commerce evidence is the set of request-level signals that show an AI-mediated shopper interacted with a merchant's catalog, storefront, cart, or checkout — captured at the log layer rather than in the browser.
Cartograph Intelligence, "Agent Commerce Evidence: A Field Guide for Merchants." https://cartographintelligence.com/agent-commerce-evidence