Retail media metrics from ad request to incremental sale, with definitions before targets

Retail media key metrics become useful when the event and eligible population are clear. An England retailer should preserve delivery counts, interaction measures and sales interpretations separately. This dictionary describes what to define; it supplies no campaign values or recommended performance levels.

The source inspection date is 6 September 2026. Scope is retailer or marketplace advertising inventory and its related reporting, excluding unrelated store traffic. Every proposed metric record below needs an owner able to explain its collection and interpretation.

What to take away

Write the decision beside the report specification

An operator investigating delivery needs different evidence from an advertiser asking how sales were credited. Someone estimating causal effects needs an evaluation design as well as reliable event records. Write the decision beside the report specification.

The HM Treasury Magenta Book analytical annex, updated on 15 May 2026, gives government methodological context. It describes experimental and quasi-experimental impact methods. It is not a mandatory measurement procedure for private retailers or evidence of campaign success.

Keep the service boundary visible. Identify the retailer property, placement type, advertiser and catalogue scope. For an England grocery retailer, a sponsored product tile on its own app home screen and a display placement on a partner site are separate properties with separate scope records.

General website traffic, unrelated channels and total retail sales enter the report only where their defined role is explained.

Complete the shared definition fields

For every metric, record event, unit, denominator, cohort, period, time zone, attribution window, exclusions, corrections, exact source and uncertainty. For standalone counts, denominator is not applicable. For non-attribution measures, record the attribution window as not applicable rather than implying a default.

Use the cohort field for the actual property, format and eligible population. Preserve source-extraction time and definition version. Late data, duplicates and excluded observations need explicit treatment; no supplier behaviour or correction delay is assumed here.

Requests and returned responses

A request count records calls asking for an advertising decision. A served or returned response count records the system response according to the implementation's specification. Keep these event counts separate, with their respective logs identified in the source field.

Do not rename either count shopper exposure. If a response-to-request ratio is needed, first establish a matched eligible request base and explain retries, empty responses and unmatched records. Those are questions to resolve in the actual data, not observed platform characteristics.

Rendering, viewability and attention

The IAB/MRC retail-media measurement guidelines, dated January 2024, provide international industry methodology for media measurement. Confirm the exact event definition adopted by the retailer.

A rendered-placement count concerns the rendering event under that definition. Viewable impressions meet specified viewability criteria. If reporting a viewability ratio, identify the eligible measured-impression denominator and disclose measurement coverage separately; unmeasured records cannot silently become failures or successes.

Attention needs a stated research or observation method. Record its unit and any denominator from that method rather than borrowing an impression count. Without suitable evidence, its value remains UNKNOWN even if a placement rendered and met visibility criteria.

Clicks and attributed sales

A click metric needs the qualifying interaction and filtering rule. A click-through ratio divides qualifying clicks by an explicitly defined eligible impression count from the matching cohort and period. Leave the ratio undefined where that denominator is missing or empty.

An attributed sale is a transaction receiving credit under a specified interaction rule and window. Define whether it counts orders, items or sales value, and document how returns and cancellations affect it. A standalone count has no denominator; a conversion ratio requires a separately stated eligible base.

Do not interpret a credited transaction as a purchase caused by the advertisement. The attribution rule explains allocation of credit, not what that shopper would otherwise have bought.

Estimates and incremental outcomes

HM Treasury's Magenta Book methods annex, updated in May 2026, addresses impact evaluation. Its government methodology is useful context for causal analysis, not a private-retailer reporting mandate.

A modelled estimate requires inputs, assumptions, model version and uncertainty alongside the shared definition fields. A causal incrementality measure also requires an explicit outcome, comparison design and analysis unit. Keep inferred effects distinct from observed sales and ordinary attribution.

Publish definitions before targets

Assemble these records for the actual dataset and mark unavailable fields openly. Carry corrections into a versioned record so earlier exports remain interpretable. No rate, monetary result or benchmark has been established by this dictionary; named editorial and qualified measurement review are still required before publication.

Before you act

Common questions

Why should a request count and a returned response count stay separate?

A request count records calls asking for an advertising decision, while a served or returned response count records the system response under the implementation's specification. The article says to keep these event counts separate and identify their respective logs in the source field. Neither count should be renamed shopper exposure.

What must be recorded for a modelled or causal incrementality estimate?

A modelled estimate requires inputs, assumptions, model version and uncertainty alongside the shared definition fields. A causal incrementality measure also needs an explicit outcome, comparison design and analysis unit. The article says to keep inferred effects distinct from observed sales and ordinary attribution.

What should happen if a metric field cannot be filled from the actual dataset?

The article says to assemble the records for the actual dataset and mark unavailable fields openly. Corrections should be carried into a versioned record so earlier exports remain interpretable. No rate, monetary result or benchmark has been established by the dictionary, and editorial and qualified measurement review are still required.