--- title: How Eventabee tracks AI attribution at the event level url: https://honeybound.co/blog/eventabee-ai-attribution-event-level date: 2026-05-25 summary: AI shopping traffic is becoming a real commerce channel. Eventabee tracks AI attribution at the event level by preserving evidence, classifying identifiable AI signals, and avoiding direct-traffic guesses. tldr: Eventabee reports AI Commerce activity from identifiable event-level evidence, not guesses. It preserves bounded signals from referrers, UTMs, source fields, and agent handoff markers. tags: eventabee, ai commerce, attribution, shopify, server-side tracking --- AI shopping attribution has a boring problem before it has an exciting one. The exciting version is the headline: shoppers ask ChatGPT, Perplexity, Gemini, Copilot, or another assistant what to buy. The assistant compares products, sends the shopper to a store, builds a cart, or eventually completes more of the checkout flow itself. The boring problem is measurement. If an AI assistant helps someone choose a product, what does your Shopify analytics stack actually see? Sometimes it sees a clean referrer. Sometimes it sees a UTM. Sometimes it sees a checkout handoff marker. Sometimes it sees an order with weak source context. Sometimes it sees direct traffic and nothing else. That last case is where attribution tools can get sloppy. It is tempting to call a new direct session "probably AI." Eventabee does not do that. Eventabee's AI Commerce attribution model is built around a stricter idea: keep the evidence on the event, classify the signal conservatively, and only report AI-attributed activity when the event contains something identifiable. For the broader market context, read [Shopify UCP is open: prepare for AI commerce, then prove what it drives](/blog/shopify-ucp-ai-commerce-readiness-attribution). ## Why AI commerce needs event-level attribution Traditional ecommerce attribution expects a fairly simple path: ```text ad click -> landing page -> session -> checkout -> order ``` AI shopping does not always look like that. A shopper might ask an assistant for "the best waterproof hiking boot under $150." The assistant may read product pages, compare structured data, check policies, and send the shopper to a product page. With protocols like Shopify's Universal Commerce Protocol, that flow can also include product discovery, cart creation, checkout handoff, and order monitoring through UCP-shaped MCP tools. [Shopify's agentic commerce docs](https://shopify.dev/docs/agents) describe that buyer journey as negotiation and authentication, product discovery, carts and checkout, and order monitoring. [Shopify's UCP profile docs](https://shopify.dev/docs/agents/profiles) also describe how agents identify themselves with hosted profiles and negotiate capabilities with the merchant or platform. That matters for measurement because the signal can appear at different points: - the original referrer - the landing URL - UTM parameters - an agent or handoff marker - Shopify order source fields - a session-to-order stitch If attribution only looks at the final order, it can miss the discovery signal. If it only looks at the first page view, it can lose the purchase connection. Eventabee treats AI attribution as an event-level problem first. ## What Eventabee stores on the event When Eventabee sees a browser or order-side signal that may carry AI attribution, it normalizes that evidence into a bounded payload object on the event. The important part is not "we saw AI somewhere." The important part is what evidence justified the classification. Examples of evidence Eventabee can preserve include: ```text referrer_domain=chatgpt.com query_param=utm_source=perplexity query_param=utm_medium=ai_agent query_param=eb_agent query_param=eb_agent_session source_name=chatgpt source_url_domain=perplexity.ai marker=ec_version marker=ucp_* marker=agent_* ``` The evidence is intentionally bounded. Eventabee should not need to copy a full arbitrary URL, customer note, or noisy source string into a reporting field just to say "this looks like ChatGPT." The goal is to keep enough proof to explain the attribution without turning the report into a data leak. That is especially important as agentic commerce matures. UCP involves profiles, capabilities, carts, checkout sessions, and handoffs. Those systems can carry useful source context, but merchants still need reporting that is understandable and safe. ## The classification model Eventabee currently separates AI commerce activity into four buckets. ### AI-referred visits This means the visitor arrived with a recognizable AI source signal. Examples: ```text https://store.com/products/widget?utm_source=chatgpt https://store.com/products/widget?utm_medium=ai_agent Referrer: https://chatgpt.com/c/abc123 Referrer: https://www.bing.com/chat?q=widget ``` Eventabee can classify known sources like ChatGPT, Perplexity, Claude, Gemini, Microsoft Copilot, and You.com when the source appears in a supported field. This is the cleanest top-of-funnel AI signal. It says: this visit carried evidence of an AI assistant or AI-search source. ### AI-assisted purchases This means a purchase event preserved AI attribution evidence through checkout or order context. For example, if an order-side source field identifies ChatGPT, or the purchase event still carries the recognized AI source context, Eventabee can classify the purchase as AI-assisted. The wording matters. This is attribution evidence, not mind reading. Eventabee should not say an AI assistant caused the purchase. It can say the purchase event preserved identifiable AI attribution evidence. ### Possible agentic checkouts This is the cautious middle category. A checkout-like flow may contain handoff markers such as: ```text ec_version ucp_* agent_* agent_session ``` Those markers can suggest an agentic commerce flow, but they are not always enough to call the purchase AI-referred or AI-assisted. Eventabee treats those as possible agentic checkouts unless stronger source evidence exists. That gives merchants visibility without overstating the claim. ### Unknown direct This is the bucket that keeps the reporting honest. If there is no known AI referrer, no recognized AI UTM, no source field, and no useful handoff marker, Eventabee does not call it AI. Direct traffic may include some AI-assisted shoppers. It may also include typed URLs, privacy tools, blocked referrers, app browsers, copied links, email clients, or ordinary direct visits. Eventabee leaves that traffic as unknown rather than inflating an AI dashboard. ## What this looks like in a payload A simplified internal Eventabee event can carry AI attribution like this: ```json { "event_type": "product_viewed", "session_id": "sess_123", "payload": { "product_id": "7997875749161", "name": "Transmission Bearing and Seal Kit", "ai_commerce_attribution": { "classification": "ai_referred", "confidence": "confirmed", "source_family": "ai", "source_name": "ChatGPT", "evidence": [ {"type": "referrer_domain", "value": "chatgpt.com"} ] } } } ``` A purchase with source context preserved through checkout might look more like this: ```json { "event_type": "purchase", "value": 202.66, "currency": "USD", "payload": { "order_id": "gid://shopify/Order/1234567890", "ai_commerce_attribution": { "classification": "ai_assisted_order", "confidence": "confirmed", "source_family": "ai", "source_name": "Perplexity", "evidence": [ {"type": "query_param", "value": "utm_source=perplexity"} ] } } } ``` A checkout with weaker agentic markers might stay cautious: ```json { "event_type": "checkout_started", "payload": { "ai_commerce_attribution": { "classification": "possible_agentic_checkout", "confidence": "medium", "source_family": "agentic_commerce", "evidence": [ {"type": "marker", "value": "ec_version"}, {"type": "marker", "value": "agent_session"} ] } } } ``` The value is not the exact JSON shape. The value is that AI Commerce reporting is built from event-level evidence instead of broad channel guesses. ## Why conservative attribution is the better product decision AI commerce is going to create pressure for inflated dashboards. Merchants will want to know whether ChatGPT, Gemini, Perplexity, Claude, Copilot, or a custom shopping agent is driving revenue. Platforms will want to show big new-channel numbers. Agencies will want proof that AI-readiness work is paying off. That makes conservative attribution more important, not less. A useful AI attribution report should answer three questions: 1. What did we see? 2. Why did we classify it that way? 3. How confident are we? If the answer is "we saw direct traffic go up and guessed," the report is not good enough. Eventabee's model is stricter: - classify known AI referrers and sources - classify recognized AI campaign parameters - preserve bounded evidence with the event - treat weak agentic handoff markers as possible, not confirmed - keep unknown direct traffic unknown - connect events to purchases only when the signal survives or can be stitched safely That tradeoff may produce smaller numbers. Good. Smaller honest numbers are more useful than large numbers nobody can defend. ## How this connects to Shopify UCP [Shopify's UCP work](https://shopify.engineering/ucp) makes the measurement problem more urgent. UCP gives agents and commerce systems a shared way to discover capabilities, search catalogs, build carts, convert carts to checkouts, hand off payment, and monitor orders. [Google has also framed UCP](https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/) as part of a broader shift toward agentic shopping across discovery, buying, and post-purchase support. For merchants, the practical question is not just "can an AI agent buy from my store?" It is also: ```text If an AI agent influenced this visit or order, will I know? ``` Today, merchants should not assume every AI or UCP-assisted order will arrive with perfect merchant-visible metadata. The safer approach is to capture every useful signal you can see at the event layer, use explicit parameters for cooperative agents, and preserve enough context to connect sessions, carts, checkouts, and orders. That is also why Shopify's newer order/cart join keys matter for attribution; see [the Order.cartToken attribution note](/blog/shopify-order-carttoken-graphql-attribution). For partner agents or custom shopping experiences, we recommend explicit handoff parameters: ```text utm_source=chatgpt utm_medium=ai_agent utm_campaign=gift_finder eb_agent=gift_finder eb_agent_session=abc123 ``` Those parameters make the downstream measurement cleaner. They also reduce the need for guesswork later. ## What merchants should do now If you are a Shopify merchant preparing for AI commerce, start with the boring foundations. First, make your store readable to AI systems: clean product titles, useful descriptions, accurate variants, complete structured data, clear shipping and return policies, and machine-readable discovery files where appropriate. If you want the store-readiness side, start with [the AI shopping agent audit checklist](/blog/shopify-ai-shopping-agent-audit). Then make the traffic measurable. That means: - preserve referrers and landing URLs where possible - use explicit UTMs for AI campaigns and agent handoffs - keep session context connected through cart and checkout - use server-side events so purchase attribution is not dependent on one browser pixel firing correctly - report AI attribution with confidence levels, not vibes Eventabee is built for that second half. It captures Shopify event context, stores normalized event payloads, respects consent, and reports AI Commerce activity from identifiable event-level evidence. If an AI assistant sends a shopper to your store, you should be able to see the signal. If that shopper buys, you should be able to connect the purchase back to the strongest evidence available. And if there is no evidence, your dashboard should say that too. ## The short version AI commerce attribution should not be a guessing game. Eventabee tracks AI attribution at the event level by preserving bounded evidence, classifying known AI sources, separating possible agentic checkouts from confirmed AI referrals, and keeping unknown direct traffic out of the AI bucket. That is the foundation merchants need before AI shopping becomes another channel they spend money on but cannot measure. [See how Eventabee handles server-side event tracking for Shopify](/work/eventabee). ## Key takeaways - AI commerce attribution breaks when tools only inspect the final order or guess from direct traffic. - Eventabee classifies known AI referrers, UTMs, source fields, and agentic handoff markers conservatively. - Unknown direct traffic stays unknown unless there is identifiable event-level evidence. ## FAQ ### Can Eventabee track ChatGPT traffic to a Shopify store? Yes, when the visit carries identifiable evidence such as a ChatGPT referrer, a recognized source field, or campaign parameters like utm_source=chatgpt. Eventabee should not classify ordinary direct traffic as ChatGPT without evidence. ### Can Eventabee track AI-assisted purchases? Yes, when the purchase event preserves AI attribution evidence through checkout, order context, or session-to-order stitching. Eventabee reports this conservatively as AI-assisted purchase attribution, not proof that AI caused the purchase. ### Does Eventabee support agentic commerce attribution? Eventabee can classify possible agentic checkout flows when event URLs or source fields contain handoff markers such as UCP or agent parameters. Stronger source evidence is required before a flow should be treated as confirmed AI-referred or AI-assisted. ### Why not count all direct traffic as AI traffic? Because direct traffic is not a source. It can come from blocked referrers, copied links, email clients, app browsers, typed URLs, privacy tools, or AI assistants. Eventabee keeps unknown direct traffic out of AI attribution unless there is identifiable evidence. ### What parameters should AI agents or partners pass for cleaner attribution? Use explicit handoff parameters such as utm_source, utm_medium=ai_agent, utm_campaign, eb_agent, and eb_agent_session. Those values make it easier to connect AI-driven discovery to sessions, carts, checkouts, and orders.