How to measure agentic commerce: a merchant KPI framework
Measure agentic commerce across three separate outcomes: whether agents can discover your offers, whether customer-authorized purchase attempts complete correctly, and whether those orders produce value after returns and service costs. Crawler requests, mentions in AI answers and paid orders are different events. Combining them into one “AI conversion rate” makes the result difficult to interpret.
A merchant needs this distinction before deciding where to invest. More mentions may justify better content coverage. More checkout failures may justify engineering work. More orders with poor contribution may require a different offer or customer segment. Each problem has a different owner.
Start with a measurement contract
For an initial pilot, define the merchant, product range, destination markets, participating agent channels and permitted purchase flow. Record the start and end dates, exclusions and observation window for cancellations and returns. Keep browser referrals from assistants separate from transactions submitted through an agent integration.
Then agree what counts as one opportunity. A single customer request could generate several searches, product lookups and basket revisions. Counting every call as a new shopper would make a busy integration appear to have weak conversion. Counting every retry as another checkout would distort reliability in the same way.
Use identifiers that connect the relevant events without treating a reusable customer identifier as necessary for every report. The operating question is whether this checkout produced this order and outcome. Access to more personal information does not automatically improve the answer.
Use a funnel with explicit evidence
The following is an editorial measurement framework, not an industry benchmark. Each stage has a different denominator and source of evidence.
| Stage | What to count | What it establishes |
|---|---|---|
| Technical discovery | Relevant crawler or product-endpoint requests | The content or offer was accessed |
| Answer visibility | Product or merchant appearances in a defined prompt sample | Presence in that observed sample |
| Offer engagement | Distinct qualified offer or checkout sessions | A prospective transaction reached the merchant |
| Authorized submission | Distinct purchase submissions passing required authority checks | A permitted attempt was made |
| Accepted order | Unique merchant-confirmed orders | The merchant accepted a purchase |
| Retained outcome | Cohort orders and net value after the chosen adjustment window | Commercial performance after observed reversals |
Do not calculate a conversion rate from crawler requests to orders unless the underlying data supports that connection. A crawler may be indexing a page for later use. An answer impression may happen without a merchant-side event. These are useful signals, but their identities and populations are not automatically joinable.
Our article on how AI shopping agents rank products separates product discovery, comparison and seller selection. Keep those questions separate in the prompt sample too: a merchant can be visible for a product name while absent from broader buying requests.
Separate visibility reporting from revenue attribution
For a repeatable answer sample, record the prompt, model or search surface, market, language, date and available session conditions. Distinguish a brand mention from a supporting citation and a purchasable merchant offer. Repeat comparable tests and report the sample size. Results describe the observations; they do not establish an overall market share of AI answers.
Google’s AI features guidance, updated December 10, 2025, says traffic from AI Overviews and AI Mode is included in Search Console’s overall Web search reporting. A rise in that aggregate therefore cannot, by itself, be labeled growth caused by AI Mode. Combine the available search data with landing-page, referral and transaction evidence, while preserving what remains unknown.
Keep a separate unattributed category. A customer might learn about a product in an assistant, then return directly on another device. Conversely, an assistant referral can arrive after the customer has already decided to purchase. Referral attribution describes the observed path, not necessarily the cause of the sale.
Use a worked funnel to diagnose the constraint
Suppose an illustrative pilot records 1,000 distinct eligible checkout sessions. Of these, 600 reach an authorized purchase submission, 540 produce accepted orders, and 45 are fully cancelled or refunded during the chosen observation window. Assume, for this simplified example, that the remaining orders have no partial refunds.
The session-to-submission rate is 60%. The submission-to-order rate is 90%. The retained-order count is 495, equivalent to 49.5% of the original sessions. Those are three different statements. Reporting only a 90% success rate would hide the 400 sessions that never reached submission and the later adjustments.
Investigate the reasons before prescribing a change. The 400 sessions might include customers declining an expensive delivery option, incomplete addresses or requests outside the pilot’s permitted scope. The 60 unsuccessful submissions might include payment declines, expired offers or execution failures. Only some of those are evidence that checkout engineering needs repair.
Compare complete cohorts. Last week’s orders have had more time to return than yesterday’s. Label recent cohorts provisional and show the observation window beside the result. If partial returns are common, report retained revenue and item-level outcomes as well as order counts.
Calculate contribution with the costs attached
Use the same illustrative cohort to build a financial view. Start with revenue after discounts and refunds, excluding taxes collected on behalf of authorities. Deduct the relevant cost of goods, payment charges, fulfillment, return processing, channel charges and attributable service costs. Align definitions with the finance team so costs are not counted twice.
For illustration only, $40,000 of net revenue less $24,000 in goods cost, $4,000 in fulfillment, $1,200 in payment and channel charges, and $2,800 in return and service costs leaves $8,000 before other overheads and pilot implementation costs. A larger order count could still produce lower contribution if the mix shifts toward expensive fulfillment or high returns.
To assess incrementality, compare a suitable control or a staged rollout where feasible. Account for promotions, inventory availability, seasonality and customer mix. A before-and-after increase is useful descriptive evidence, but it cannot isolate the integration’s effect when other factors changed too.
Include identity decisions in the operating view
Where an offer depends on membership or another customer attribute, measure whether eligible customers can present the required evidence and receive the intended offer. Separate a failed evidence check from a valid customer who does not qualify. Otherwise a team might misdiagnose a targeting decision as technical friction.
Our merchant Know Your Agent checklist sets out the distinct trust decisions behind a purchase. AIR Identity supports credential-based recognition; any measured effect on conversion or acquisition economics must come from the merchant’s actual implementation and comparison data.
Frequently asked questions
What is the main KPI for agentic commerce?
There is no single sufficient metric. Pair correctly completed purchases with retained contribution, and use discovery, offer engagement and failure reasons to explain changes.
Do AI crawler visits count as shoppers?
No. They show technical access to content or endpoints. They do not establish a customer’s buying intent, a recommendation impression or a purchase.
Can Search Console prove revenue came from AI Mode?
Not by itself. Google includes AI feature traffic in overall Web reporting. Revenue attribution requires additional transaction and journey evidence, with gaps acknowledged.
How should merchants compare agent checkout conversion?
Use a consistent denominator, such as distinct eligible checkout sessions, and report submission and accepted-order rates separately. Compare similar products, customers and observation periods.
When should returns be included in pilot results?
Track them from the start and compare cohorts over a defined window. Label newer cohorts provisional, and report partial refunds and retained revenue where order counts conceal the effect.
Partner with AIR
Evaluating how customer credentials affect eligibility and conversion in agent-led commerce? Talk to AIR about a measurable use case for your business.
Moca Network, the identity network of Animoca Brands.
Partner with AIR
AIR is one integration suite for identity, money and loyalty. AIR Identity lets trusted issuers issue verifiable credentials and businesses verify them, so acquisition starts with users who already qualify, with minimal data custody. AIR Money embeds compliant global financial infrastructure, opening new markets without building new rails. AIR Loyalty gives brands a programmable, stable-value points engine that keeps users coming back.
Partner with us to grow your business.
AIR is built by Moca Network, the identity network of Animoca Brands.




.png)