Predictive Creative Scoring Moves the Verdict Before the Spend

The creative verdict just moved ahead of the auction

Predictive creative scoring just cleared its biggest real-world test. Kroger, Vidmob and MMA Global analyzed 1,934 assets from a year of live Meta and DV360 campaigns, and the model forecast e-commerce conversion with 81% accuracy on data it had never seen. Creative built to the model's guidelines converted four times better at up to 70% lower cost per conversion. The verdict now comes before the spend.

The setup matters as much as the score. Per the companies' announcement, the three partners took 1,934 video and image assets from Kroger's live campaigns across Meta and DV360, trained on a full year of 2025 campaign data, and validated the model against an independent Q1 2026 dataset it had never encountered — a forward-looking test of whether the scorecard could predict future performance rather than explain the past. The same announcement says the model read creative variables spanning visual treatment, messaging, narrative structure, branding and on-screen human interaction, and that each finding cleared statistical significance testing before it earned a place on the final guideline list. That validation design is what separates a scoring system from a retrospective case study.

The industry already knew this mismatch from the other direction: teams bought AI video for speed and then audited it for revenue, the mismatch documented in {{link}}, and this study attacks the same gap from the opposite end. Media plans are optimized to the decimal while creative is still judged on instinct, as the Advertising Week session page for the study puts it. Asking what an unfinished asset will do — before a single media dollar moves — is the reframe.

The industry already knew this mismatch from the other direction: teams bought AI video for speed and then audited it for revenue, the mismatch documented in AI video's proof gap, and this study attacks the same gap from the opposite end.

What Predictive Creative Scoring Changes Inside the Auction

The headline numbers describe a gate, not a dashboard. Predictive creative scoring forecast e-commerce conversion with 81% accuracy across Meta and DV360. Assets that followed Kroger's data-backed guidelines converted at four times the rate, at up to 70% lower cost per conversion. And the cheapest win required no new production at all: reallocating existing budget toward higher-scoring creative could lift conversions by up to 2.2x from the same spend, a finding both partners describe as the practical starting point — meaningful conversion gains from money that was already committed, no new assets required. Vidmob's founder and executive chairman Alex Collmer framed the shift as one of ownership: "Every brand can produce more creative than ever. The advantage now belongs to the ones who know which creative will drive the outcome they need."

Context explains why this lands harder than another benchmark. Targeting, bidding and budgeting are measured on every media dashboard in the business; creative is not, even though MAGNA Media Trials and Yahoo found back in 2023 that creative quality accounts for 56% of campaign impact against 44% for media placement — the gap Kroger set out to close, according to MMA Global's session page. When the larger share of impact sits in the part nobody measures, a model that forecasts its outcome before launch is not an incremental analytics upgrade. It moves the moment of judgment out of the post-campaign review and into the briefing process.

Meta's auction already concentrates spend on a small set of post-hoc winners, a dynamic captured by {{link}}, and predictive scoring is an attempt to identify that small set before the auction does. The difference sounds subtle and is not: one method discovers winners after the money has moved, the other decides where the money is allowed to move.

Meta's auction already concentrates spend on a small set of post-hoc winners, a dynamic captured by the creative power law, and predictive scoring is an attempt to identify that small set before the auction does.

Human moments beat product-first creative

One finding held across platforms and may matter most to video teams. Creative centered on authentic human experiences outperformed creative focused primarily on products or the act of shopping, according to the study announcement, and MMA Global's session page describes the same result in its own words: relatable human moments beat product-led and transaction-led creative. For Kroger, whose stated brand purpose is to "Feed the Human Spirit," campaigns featuring people sharing and enjoying food produced stronger e-commerce results than transactional shopping imagery or product-first shots. A grocer's purchase data is saying something general about video advertising: the render can be flawless and the asset can still lose, because what the model rewarded was the presence of people, not the polish of the product.

For AI-generated video this cuts against the grain of the tooling. Generation pipelines make product-first imagery almost free — packshots, alternate angles, lifestyle backdrops on demand — and the temptation is to fill the calendar with them. The Kroger model suggests the cheapest creative to produce is also the creative most likely to score lowest. What the scoring layer prices in is judgment, not throughput.

Retailers feel this first. {{link}} has already shortened the distance between product content and checkout, and a pre-launch score decides which of those generated assets deserves to make the trip. A fourfold conversion gap between guideline-following and off-guideline creative is exactly the kind of spread that no launch-day review can recover after the fact.

Shoppable AI video has already shortened the distance between product content and checkout, and a pre-launch score decides which of those generated assets deserves to make the trip.

Scoring before launch is a workflow change, not another report

MMA Global's session page states the operational goal plainly: score creative before launch so fewer low performers reach market, rather than waiting for platform algorithms to sort it out after spend. The session treats the study as one step in a longer build — from a point-in-time analysis to an always-on capability that carries validated guidelines into briefs and ties creative investment back to measurable outcomes. Vas Bakopoulos, who leads insights and research at MMA Global, put the pressure side of it directly: "As AI makes creative production ten times faster, judgment has to scale with it." His roadmap extends the same idea inward and outward: the next chapter goes deeper, to the segment level, and wider, into mix models and buying tools, "so creative earns a seat at the table where investment decisions are made."

For teams running AI video pipelines, the practical insertion point is between generation and delivery. {{link}} has already pushed those teams to produce far more takes than they can ship, and scoring is the gate that decides which variant earns the first media dollar. Running that gate well needs its own discipline: guidelines that cleared a statistical bar rather than opinions that photographed well, scores attached to briefs instead of dashboards, and a feedback loop in which — as Collmer puts it — every campaign makes the next one smarter. A gate nobody re-validates is just a new opinion with a number attached, and it will erode exactly as fast as the platform mix it was trained on.

Variant economics has already pushed those teams to produce far more takes than they can ship, and scoring is the gate that decides which variant earns the first media dollar.

The limits of one retailer's scoreboard

The caveats deserve the same weight as the findings. This is one retailer, one category, one purchase-based conversion objective; the guidelines that survived testing describe what predicted conversion for Kroger's assortment on Meta and DV360, not universal laws of video advertising. The model was trained on 2025 data and validated on a single later quarter, and the partners publish the full methodology — including the statistical bar each guideline had to clear — for scrutiny rather than claiming the framework ports automatically to every brand. Bakopoulos's own framing points forward, not sideways: segment-level scoring, integration with mix models and buying tools. The honest read is that predictive creative scoring works where a brand has a year of clean outcome data and the willingness to let a model pre-judge its creative before humans fall in love with it. That is a high bar most teams have not cleared yet, which is precisely why the 81% figure is worth watching: it suggests the bar is clearable, and that the next competitive gap in video advertising will open between teams who score before they spend and teams who still find out after.

Most teams cannot copy the study outright, but they can copy its sequence. Start by scoring retrospectively: tag a year of already-spent creative with the same attributes, see which ones survive significance testing against your own outcomes, and let the survivors write the guideline list. Put that list into briefs, score new assets against it before launch, and record every score next to every outcome so the guideline set keeps learning. None of this requires Vidmob's platform; it requires treating creative decisions the way media decisions have been treated for a decade - as measurable, ownable and improvable. That is the real finding of the study: not that a grocer found a model, but that creative judgment can be built as infrastructure rather than exercised as taste.