Agent-originated conversions
Direct answer
An agent-originated conversion is a purchase whose journey ran partly or wholly through an AI answer engine or shopping agent — a recommendation, comparison, or agent-placed order — rather than a classic search results page. MIZ OKI is building additive event fields, a flag-off classifier, a shadow table and read-side stratification to measure them. Capability label: [IN BUILD].
Why this is a new measurement problem
The referral path is changing. Traditional search engine volume is forecast to fall 25% by 2026 as AI answer engines absorb queries (Gartner, Feb 2024). Traffic to US retail sites from generative AI sources rose 1,200% between July 2024 and February 2025 (Adobe Analytics, Mar 2025). ChatGPT reached 800 million weekly users (OpenAI DevDay, Oct 2025), and enterprise adoption is following — 78% of organisations already use AI in at least one business function (McKinsey, Mar 2025) and 23% are scaling an agentic AI system somewhere in the enterprise (McKinsey, Nov 2025).
Attribution built for click-through from a results page mislabels these journeys. A visitor who arrives from an answer engine looks "direct" or "referral"; an order placed by an agent on the buyer's behalf may carry no session at all. Counting them wrong distorts two things at once: the incrementality of the media that reached the buyer earlier, and the measured value of being cited by the engine.
What "agent-originated" means at MIZ OKI
The definition is deliberately narrow and evidence-led:
- Origin is classified from the canonical event, never guessed from the destination page. The classifier reads additive fields on the platform's journey-event schema — referrer class, agent-protocol markers where a protocol exposes them, and order metadata — and emits a class with a confidence, never a bare label.
- Classification is additive. New fields are appended to the schema; old events validate unchanged; nothing existing is rewritten.
- Shadow first. Classified events land in a shadow table that feeds read-side stratification (agent-originated vs not) in reports; they change no budget decision until a registered holdout shows the stratum behaves differently.
- Consent and erasure travel with the event. The consent gate runs fail-closed on every event regardless of origin, and the erasure path covers the shadow table like any other store.
The citation KPI, and what it is not
The site's own referral class is measured the same way. The identity-free site-events schema classifies the Referer host into direct | search | ai_answer_engine | social | other and stores only the class — the host itself is dropped before the row is built, and the schema test asserts no host, IP, user-agent or cookie can be stored. referrer_class = ai_answer_engine on a page view is therefore the citation KPI: the share of visits that arrived from an answer engine. It says nothing about which engine, which query, or who the visitor was, and it is a page-view count, never a conversion count.
What is shipped today — honestly
- [IN BUILD] Additive schema fields, the flag-off origin classifier, the shadow table and read-side stratification are being built in the current wave; nothing is customer-facing and no customer figure exists.
- [PARTIAL] The identity-free site-events schema with the
ai_answer_enginereferrer class is built and flag-gated OFF on the site. - [PROPOSED] A governed decision feed that would let an agent read a tenant's eligible decisions through the Decision Control Plane is specified only.
- No conversion figure on this page is a MIZ OKI result. Third-party statistics carry their source and month inline.
FAQ
Does MIZ OKI track visitors coming from ChatGPT or Perplexity?
It classifies the referring host into a coarse class and stores only the class, aggregated to the hour. The host, the query, the IP, the user agent and any cookie are never stored — the schema forbids the columns.
Will agent-originated conversions change my budget decisions?
Not yet, and not automatically. They are measured in a shadow table and shown as a stratum in reports. Only a registered holdout showing the stratum responds differently to media can change a decision, and that follows the normal promotion path.
Is this the same as "AI attribution"?
No. Attribution assigns credit by rule. Agent-originated classification labels the journey's origin with a confidence, and incrementality is still measured against a holdout. The label says where the buyer came through, not what caused the purchase.