Measurement for Agentic Commerce

Every metric in the commerce measurement stack encodes the same unexamined assumption: the shopper is a human, on a screen, in a session. The MRC viewability floor defines an impression as an opportunity for a human eye to see — 50% of pixels for one continuous second, two for video. Click-through rate assumes a finger. Attribution assumes a browser that carries identity from ad to cart. Conversion-rate optimization assumes a mind that can be persuaded by layout. None of these is wrong. All of them are now optional.

An agentic purchase is a transaction in which the economic buyer is human but the shopper — the party that searches, compares, selects, and checks out — is software acting on delegated intent. That party does not click, does not see ads, does not have a session in any meaningful sense, and increasingly does not use a browser at all: it calls a checkout API.

This is no longer a thought experiment with a rounding-error denominator. Adobe Analytics — a retailer panel that classifies AI traffic by referral, so treat every number as vendor-sourced and a floor, since API-only journeys never appear in it — reports AI-sourced traffic to US retail sites up 393% year over year in Q1 2026 (PDF; summary), with conversion flipping from 38% worse than non-AI traffic in March 2025 to 42% better in March 2026. Shopify has reported AI-attributed orders growing roughly 11x from January 2025 to March 2026 — platform-reported, via press, not independently verified. McKinsey’s consultancy estimate puts $3–5 trillion of global retail spend in play for agentic redirection by 2030; its agentic-advertising work frames the valuable position not as the viewed ad but as inclusion in the agent’s comparison set — or being its default selection.

And the clearest evidence that this is a measurement problem rather than a traffic curiosity is that the largest retail media network went to court over it. Amazon sued Perplexity in November 2025 for letting its Comet agent shop while disguised as a human Chrome session, and won a preliminary injunction on March 10, 2026. Perplexity’s defense argued Amazon was protecting its ad revenue — agents don’t see ads. Both sides are right about the mechanics: when the shopper stops being human, the surface’s ad economics and its measurement moat break in the same motion. Amazon’s preferred resolution is agents on its own rails — Rufus, which Amazon says roughly 250 million customers have used, with Rufus users 60% more likely to purchase (both Amazon-sourced), and Buy For Me — with third-party agents admitted, per Andy Jassy, on Amazon’s terms (GeekWire on the ruling).

Four checkout rails, zero shared metrics

The transaction layer standardized first, and fast. OpenAI’s Instant Checkout runs on the Agentic Commerce Protocol (ACP), an open standard co-developed with Stripe: the merchant stays merchant of record, Etsy went first, Shopify merchants followed (Stripe announcement). The widely reported ~4% transaction fee is press-reported — it appears in no primary document. Google answered with the Universal Commerce Protocol (UCP), co-developed with Shopify, Etsy, Wayfair, Target, and Walmart and endorsed by twenty-plus payment and retail companies, wired into agentic checkout in AI Mode — which Google says has passed a billion monthly users — and the Gemini app, compatible with the Agent Payments Protocol (AP2). Perplexity took the wedge position: zero fees and commissions, PayPal/Venmo checkout, merchant stays merchant of record (CNBC). Amazon, per the section above, runs its own.

Advertising arrived on the same surfaces almost simultaneously. OpenAI began testing ads in ChatGPT in the US on February 9, 2026 — labeled units at the bottom of responses, Free and Go tiers only, expanded across eight more countries by May — with a retail-heavy first advertiser wave (Target, Williams-Sonoma, Albertsons). OpenAI’s load-bearing claim, in its help center: ads “never influence the actual answers.” Hold onto that sentence — it defines the classifier every measurement program in this category must draw. The agent’s recommendation is organic, produced by the model; the ad is adjacent, sold separately. Google’s 2026 ads and commerce outlook points the same direction, with AI-powered Shopping ads and an exclusive-offers pilot inside AI Mode (Google Marketing Live 2026 roundup).

Note what all of this standardizes: checkout, payment, identity of the order. Nothing in ACP or UCP tells a brand why the agent picked it — or that it was considered and rejected. The rails are live. The measurement is not.

The triage

The useful question is not “does measurement survive agentic commerce” but which metrics break, which survive on their existing logic, and what has to be built new. The sorting rule is one sentence: a metric breaks if it presumes a human eye or a browser session; it survives if it observes money or causality; what emerges is whatever must observe the agent itself.

What breaks

Click-through rate. ACP and UCP checkout produces zero clicks by design; the agent calls an API. The click survives in exactly one pattern — the handoff, where an agent researches and a human lands to buy. That handoff is the only reason Adobe’s referral-classified panel sees anything at all, which means the industry’s best public dataset on agentic commerce is, structurally, a dataset of the journeys where the agent didn’t finish the job.

Viewability and attention. The MRC floor presumes a human opportunity-to-see. An agent parsing a product feed has 100% “viewability” of everything and attention to nothing — the metric returns a perfect score while measuring a category error. The measurement, verification, and media quality stack was built to answer “could a human have seen this?”; the agentic question is “could a model have parsed this?”, and no accredited metric asks it yet.

Reach, frequency, view-through. No exposure means no view-through. Frequency capping against an entity with perfect recall and no fatigue is meaningless in both directions — one exposure is enough, a thousand change nothing.

Persuasion-surface metrics. Session-based conversion optimization and A/B tests assume layout moves minds. The agent is not moved by hero images; it satisfies constraints. What replaces persuasion design is machine legibility: Adobe’s AI Content Visibility Checker (a vendor tool — label accordingly) scores retail homepages around 75% and product detail pages around 66% visible to LLMs, meaning a quarter to a third of commerce content is invisible to the shopper that converts best.

Human-baselined bot filtering. The quietest break, and the funniest. The entire invalid-traffic apparatus treats non-human traffic as waste to be filtered — and as of this year, the highest-converting traffic in Adobe’s retail panel is non-human. HUMAN’s MRC-accredited viewability measurement is already wrestling with AI-agent classification. Every analytics team should assume its IVT rules are currently deleting buyers.

What survives

The conversion event. Money clears either way — and agentic checkout arguably improves the record, because an ACP or UCP order is a structured object with protocol-level provenance, not a pixel fired from a browser that may or may not still hold consent. The natural home for standardized outcome reporting already exists: event and conversion APIs, where IAB Tech Lab’s ECAPI is the closest thing shipping to an agentic conversion standard.

Incrementality. The three requirements in the IAB/IAB Europe incrementality guidelines — a credible counterfactual, controlled bias, separation of signal from noise — are substrate-agnostic. They do not care whether the treated unit clicked or delegated. What must change is the unit of randomization: cookie- and session-level assignment collapses when one household’s purchases route through an agent that touches four surfaces, so holdouts move to the household or account level. The discipline survives; the plumbing gets rebuilt.

Marketing mix modeling. Top-down econometrics never cared who clicked. If anything, agentic opacity at the journey level pushes more budget-allocation weight back onto MMM — the measurement family that requires the least cooperation from the surface.

The audit habit. Everything in the iROAS buyer’s protocol — this number, by which methodology, with what diagnostics? — transfers intact. The first vendors selling “agentic iROAS” will deserve those questions more than the current ones do.

What emerges

Verified agent identity — the new impression-grade primitive. Today, “agent traffic” in most analytics is a user-agent-string guess. The cryptographic substrate to fix that already exists: Web Bot Auth, an IETF draft built on RFC 9421 HTTP Message Signatures, plus Cloudflare’s signed agents and agent registry, with Visa TAP and Mastercard Agent Pay verifying the agent at the payment processor. Note the rhyme on the media-buying side: AdCP signs agent-to-agent requests with the same RFC 9421 primitive. One cryptographic idiom is quietly becoming the identity layer for agents that buy media and agents that buy products. Until a surface verifies signatures, every ”% of traffic is agents” chart you see is an inference, not a measurement.

Inclusion and citation rate. From GEO practice: was the brand present in the answer at all, and was it cited as a source or merely mentioned — distinct metrics that commodity SEO content already conflates. Treat every “citation lifts clicks by X%” stat circulating in that content as vendor-sourced and single-study until someone replicates it.

Agent share of choice. The vocabulary has a genealogy worth knowing: Share of Model was coined by Jellyfish in 2024, formalized by INSEAD in 2025, and is now tooled by Profound, Semrush, and Adobe’s LLM Optimizer. But SoM measures mentions — how often a model surfaces the brand. The agentic successor must measure selection: of the purchase tasks agents complete in your category, the share in which your product was the chosen option, or the default. No standards body or vendor has standardized that term as of July 2026; call it agent share of choice — coined here, labeled as such. The distance between share of model and share of choice is the distance between PR and revenue.

Order-object attribution. ACP and UCP orders arrive as structured objects. The emerging attribution question is which provenance fields those objects carry — surface, agent identity, whether a paid placement was present in the comparison set — and who gets to read them. Today the protocols specify the order, not the why.

Catalog legibility. A quarter to a third of retail content invisible to LLMs (Adobe, vendor-sourced, above) makes “share of catalog an agent can actually parse” a board-level metric with a direct revenue line. Expect it to be productized within a year; expect every vendor score to disagree.

The separation audit. OpenAI states ads never influence the answers. That is a falsifiable claim about the most valuable recommendation surface in commerce, and testing it — systematically, with paired prompts, over time — is a measurement job nobody is accredited to do yet. Whoever does it credibly becomes the agentic era’s verification layer.

Where the standards bodies actually are

The honest map, as of July 2026:

• IAB Tech Lab has the most concrete roadmap — AAMP’s three pillars, the Agentic Real-Time Framework, buyer and seller agent SDKs v2, ECAPI, OM SDK device attestation, and an Agent Registry live since March 3, 2026. But read what it standardizes: agents as buyers of media. The agentic standards stack has no dedicated methodology for the agent as shopper — the entity in Adobe’s conversion data. That is the gap this essay maps.
• IAB (the trade body) is circling: the 2026 Measurement Leadership Summit carries a “Building Measurement for the Agentic Era” track, and the Campaign Data Standards (Project Eidos) comment window closed July 6, 2026.
• MRC has no public agentic-shopper measurement workstream — not publicly documented as of July 2026. Its viewability floor still presumes the human eye it was written for.
• The demand side is not waiting. In eMarketer’s April 2026 survey of commerce media leaders across the US, UK, and Germany, 92% said they are likely to invest in agent-specific measurement and diagnostics within two to three years — the top-ranked agentic use case — while 55% of US advertisers already report inconsistent RMN targeting and attribution. Survey data, labeled as such; the gap between 92% intent and zero accredited methodologies is the market.

For the retail media buyer, this lands on top of a measurement stack that was already un-comparable across networks — the full picture is in the retail and commerce media measurement decode and the cross-RMN comparability matrix. Agentic traffic does not create that problem. It adds an unmeasured population to it.

The buyer instrumentation list

What to stand up now, before any standard tells you to. None of this requires a vendor contract; most of it is logging discipline.

1. Split “AI traffic” into three lanes — cryptographically verified agents, declared referral handoffs (chatgpt.com, perplexity.ai, gemini), and undeclared automation. One “AI traffic” line in a dashboard is three populations with three economics.
2. Verify identity at the edge, don’t guess from user agents. Turn on Web Bot Auth / signed-agent verification where your CDN offers it, and log the verification result into analytics as a first-class dimension.
3. Audit your IVT rules before they eat your buyers. Find where verified or declared shopping agents are being filtered as invalid traffic, and carve the exemption deliberately — with the verification from step 2, not a user-agent allowlist.
4. Capture protocol-native order metadata. If you sell through ACP or UCP, the order object and the payment-rail attestation (Visa TAP, Mastercard Agent Pay) are your new attribution fields. Store them raw; the schema questions come later.
5. Baseline agent conversion separately. Adobe’s own flip — AI-referred traffic from 38% worse to 42% better conversion in twelve months (vendor-sourced) — means any pooled conversion rate is now a blend of two populations moving in opposite directions.
6. Track inclusion and citation with dated snapshots across the four assistant surfaces, in-house or tooled. Keep the raw prompts and answers; vendor indices change methodology without notice.
7. Score catalog legibility per PDP — structured data, feed completeness, whatever an agent must parse to select you — and treat the score as a merchandising KPI, not an SEO one.
8. Re-base incrementality on humans-plus-their-agents. Household- or account-level holdouts, per the IAB’s three requirements. And add one question to the iROAS audit protocol for every network QBR: how does your measurement panel classify agent traffic — and which side of test/control does it land on? No network has a public answer today.

Speculation, labeled

Everything above is sourced. This section is not — it is where the author thinks the next three years land, stated so it can be wrong in public.

• Payment rails become measurement vendors. Visa and Mastercard see verified agent transactions end to end, across every surface, with cryptographic identity attached. Receipt-rail attribution — the panel nobody has to recruit — is too valuable not to productize. Speculation.
• Protocols grow a “why” field. Some ACP/UCP successor or extension standardizes agent-declared selection disclosure — what was in the comparison set, what ranked the winner — the agentic analogue of a bid landscape. The pressure will come from advertisers who notice the order object is the only trustworthy log in the chain. Speculation.
• RMNs ship “agentic iROAS” before any neutral standard exists, measured on panels whose agent-classification rules they will not disclose. The iROAS negotiation replays with a new noun. Speculation, though barely.
• MRC accreditation for agent-traffic classification becomes the new viewability accreditation — the commodity trust layer everyone buys and nobody loves. HUMAN’s current accreditation work is the visible on-ramp. Speculation.

The pattern under all four: measurement authority is migrating from the parties that watch the shopper to the parties that authenticate the agent. In the human web, whoever owned the panel owned the truth. In the agentic web, whoever verifies the signature does.

Instrument now. The standards, when they arrive, will want your logs.