Nobody has an accredited sales number

Retail media will take about $71bn in the US this year, and the metric the whole category is bought on has never been independently audited anywhere. Four of the six largest networks hold MRC accreditation; every one of those covers impressions, clicks and viewability. Attributed sales, ROAS and incrementality are accredited at zero networks. One network has put them into audit, and it is not one of the two taking the money.

Ipalibo Da-Wariboko · CULT+MATH · November 9, 2026 · 7 networks scored · 22 sources · public data only · no instrument run

The decision this is for

Whether a challenger network differentiates on audited outcomes rather than on reach. It is a planning-cycle decision because an audit takes quarters, not weeks, and because the first network through it writes the definition the rest have to answer.

The case against is immediate and real: audited numbers are smaller numbers, and a network that reports honestly competes against six that do not. The case for is that a reported number nobody can check is already what every network sells, so the only differentiated position left in the category is the one nobody has taken.

Incrementality is not a better number. It is a number with more discretion in it

The category’s answer to untrustworthy ROAS has been incrementality, and nearly every network now reports some version of it. Albertsons Media Collective’s own research, published in March 2026, found that iROAS can vary 6.5x and flips the result in 83% of campaigns depending on filtering, matching and revenue-calculation choices.

Those choices are made by the network. A control group, a set of matching variables and a revenue definition are three places to put a thumb, and none of them is visible to the buyer. That finding was published by a company that sells incrementality reporting, which is what makes it worth more than an agency complaint: it is an admission against interest.

So a brand comparing iROAS across two networks is not comparing performance. It is comparing two methodologies it cannot see, and the measurement comparison that sits behind this study says it plainly — ROAS is not comparable across these networks. Models, windows and sales scopes differ enough that identical campaigns produce different numbers.

The part that is not about honesty

Transparency arguments lose to performance arguments, and they should. A buyer with a number that looks good does not want a lecture about where it came from.

But a network cannot improve against a measure it authors. Optimization needs an objective the optimizer did not write. A network tuning toward a figure whose methodology it controls, and which moves 6.5x on those controls, cannot distinguish getting better from counting differently — and over enough cycles the system learns the counting.

That makes the open-loop position a capability claim rather than an ethics one. The pitch is not our numbers are better than theirs. It is theirs cannot be checked, ours can, and only one of us can therefore actually get better at this. Against a closed-loop competitor reporting higher performance, that argument does not require beating the number. The number is the thing in dispute.

Who the argument has to be made to

A measurement argument only moves money if money is allocated on measurement. In retail media it is allocated on measurement and on a negotiation, and the two sit in different parts of the building.

These do not contradict each other and the difference between them is the mechanism. Forrester measures where the dollars are sourced. Dentsu measures which function holds the plan, concurrently, which is why its percentages exceed 100. The plan has moved to brand marketing. The money has not. 62% of it still originates in trade and shopper budgets, and those dollars arrive with obligations attached to a retailer negotiation.

Joint business plans are recorded crediting $1,000 of self-serve spend as $750. A plan that applies a 25%% haircut to media spend for negotiating purposes is not allocating that spend on performance. It is counting it as trade.

And the advertiser’s own account of how it allocates points both ways at once: modeling recommendations, performance measurement and business plan adherence are the top three factors, all at 38%. Performance is exactly as important as the negotiation — and “performance” here means the number the seller produced.

The case against, which the data currently wins

The networks furthest along on measurement are the challengers. The challengers are not taking share.

Share of US retail media ad spend, 2025. eMarketer, January 2026. The three networks furthest along on measurement sit inside the bottom bar.

And the direction is worse than the level. Of roughly $10.53bn in net-new US retail media spend forecast for 2026, $9.42bn — over 89% — goes to Amazon and Walmart, with every other tracked network expected to hold flat or shrink through 2027. Neither of those two has an accredited sales metric.

The forecaster’s own reading is that the other platforms have not yet proven they deliver incremental audiences or measurement quality strong enough to pull budget away from the leaders — they are growing alongside that expansion rather than taking from it.

That is the honest state of the evidence, and it says measurement has not moved share. What it cannot say is whether measurement would, because none of these networks has shipped the thing being argued for. Each has a piece. Instacart has accredited delivery metrics and an open clean room, and no accredited outcome. Albertsons has the strongest incrementality design and no accreditation at all. Kroger has the outcome metrics in audit and is alone in that.

A null against a treatment nobody has administered is a null on the market, not on the idea.

What this cannot tell you

Nothing here is measured by us. This is a review of what is publicly disclosed, and the central claim — that no outcome metric is accredited — is a claim about a public register. It is checkable, and it is the kind of claim that goes stale the day a letter is granted.

No network discloses the revenue effect of a transparency move. Retail media revenue is not separately reported by any grocery or mass retailer in the set, so the obvious objection — that honest numbers are smaller numbers and the first mover pays for it — is neither supported nor refuted here. It stays open.

And nobody publishes what share of retail media spend is incremental. We looked for it directly and found no measurement from any source. Every figure in circulation that resembles one is a network’s self-reported lift or an agency’s estimate. The mechanism in this study has no magnitude attached, and the reason is that gap rather than caution.

Three load-bearing figures reach this page through secondary outlets rather than from the publisher: the share forecasts, the budget-ownership percentages and the trade/shopper split. Each is consistently reported across independent outlets, which is corroboration and not verification. The Dentsu figures come from 100 respondents surveyed in August 2024.

Network self-reported performance is excluded entirely rather than discounted — household-penetration lifts, campaign ROAS case studies, vendor-published uplift. It is marketing material produced by the party being measured, which is the subject of this page.

How it was built

The question, the four candidate explanations, a prior for each in ICD 203 bands, three falsifiers and the stopping rule were written down before any source was opened. Two of the four explanations were registered above the one being tested, because the pass is only worth running if the unfavorable readings might hold. Three of them do, in part.

Accreditation status comes from the MRC’s own digital accreditation directory and its services-under-review listing, which is the only independent record of what has been audited and at what scope. Network capability claims come from each network’s product documentation. Share and forecast figures are eMarketer’s. Budget-ownership figures are Forrester’s and Dentsu’s.

Four predictions were registered. Three were met. The fourth was written badly enough that no available source could resolve it — it named a comparison the published surveys do not make — and it is recorded as a design fault rather than a result. The stopping rule was saturation across three test dimensions, capped at one session, and it was met at 22 sources.

The next one

Whether Kroger’s pre-audit completes, and whether a second network follows it in. Both are registered predictions with dates on them — June 30 and December 31, 2027. The working goes out first on Sound Decisions.