Kroger Used AI to Kill Bad Ads Before Running Them

A study published this week involving Kroger, Vidmob, and MMA Global found that predictive AI creative scoring can forecast e-commerce conversion with 81% accuracy. The ads that scored well and ran converted at 4x the rate of the ones that didn't, across Meta and DV360, at up to 70% lower cost per conversion. Those are the numbers. I'll say up front that the full source page was unavailable when I pulled this, so I'm working from the study summary, and I'm not going to dress it up with details I don't have.

Here's what caught my attention, though: Kroger didn't use AI to write copy or generate images. It used AI as a filter. Run the creative through the model, see which pieces are predicted to convert, put money behind those, and let the rest die quietly. That's a fundamentally different use of AI than what most businesses are thinking about right now.

What "creative scoring" actually means

Before I ran auction events, I spent years figuring out which items to put on the front table, which to bury in a box lot, and which to photograph for the catalog. You didn't wait for the bidding to tell you. You got good at predicting before the gavel fell, because by then the money was already on the table.

Creative scoring is the same instinct, automated. You feed the model your ad assets, and it predicts performance before you spend a dollar on media. Vidmob's platform analyzes the visual and structural elements of a creative, compares them against patterns from prior campaigns, and returns a score with enough predictive power that Kroger trusted it at scale. The MMA Global study validated that 81% accuracy figure. The outcome wasn't just better ads. It was the same budget, aimed at better targets, with far less waste.

Call this the "kill filter." Not a creative tool. A culling tool.

Why this matters more than another AI content announcement

Most of the AI-in-marketing conversation is about production. Write faster, generate more, post more often. And look, I've seen that pay off. But production without curation is just more noise. The Kroger result is interesting precisely because they weren't using AI to produce anything. They were using it to stop producing the wrong things.

If your paid social or display budget is $2,000 a month and you're running four ad variations, you don't actually know which one is going to convert until you've split-tested it live. That test costs real money and real time. A scoring model that's right 81% of the time means you could, in theory, skip the expensive learning phase and let the model's homework replace your testing budget. At 70% lower cost per conversion, that's not incremental. That's a structural change in how you allocate media spend.

The math on your business will be different from Kroger's. Obviously. But the principle doesn't care about your revenue tier.

The kill filter and how to think about applying it

The kill filter is not a piece of software you buy this afternoon. It's a decision-making posture. Here's the two-question version:

1. Do you have enough past campaign data for a model to learn from? Vidmob's approach works because Kroger has years of creative performance data across massive media spend. If you're a local business with six months of Facebook ad history and four creatives, the model has thin material to work with. The accuracy drops when the training set is thin.
2. Is your current creative testing process costing you more than a scoring tool would? If you're running $5,000 a month in paid media and you're not testing systematically, you're already paying the tuition. A scoring tool that costs a few hundred dollars a month and catches even a fraction of the waste is worth the math.

For most local businesses, the version of this available today is simpler: use a creative review step before ads go live. That might be a structured prompt in ChatGPT that asks why this image and headline would or wouldn't convert for your specific audience. It's not 81% accurate. But it's better than nothing, and it's free this afternoon.

Under the hood

For the technical reader: what Vidmob is doing is computer vision plus performance prediction. The model breaks down ad creative into component signals, things like color contrast, face presence, text density, motion in video, call-to-action placement, and maps those signals against conversion outcomes from prior campaigns. The 81% accuracy claim is for e-commerce conversion prediction specifically; accuracy on other objectives may differ, and the study summary doesn't break that out. The platforms used were Meta and DV360, both of which give Vidmob enough signal-level data to do feature attribution at the creative element level. This is meaningfully different from platform-native tools like Meta's Advantage Plus creative, which optimizes after launch rather than scoring before it. Pre-launch scoring is where the differentiation lives.

If you want to see a related application of AI doing filtering work rather than generation work, the Skyline Chili org chart piece is worth a read. Same principle: AI as a decision layer, not a production layer.

What to do this week

If you're running paid media at any scale: Pull your last 90 days of creative performance by ad unit. If you haven't done that recently, the data is probably sitting in your Ads Manager right now. Look at which creative elements the top performers share. That's a manual version of what Vidmob is automating, and it costs nothing but an hour.

If you're spending more than $3,000 a month on paid social: Ask your agency or media buyer whether they have a pre-launch creative review step, and what it's based on. If the answer is "gut feel" or "past experience," that's fine, but it's worth knowing. The kill filter conversation is worth having now, before you've committed the next quarter's budget.

If you want to experiment with an AI-assisted review: Write a prompt that describes your target customer, your conversion goal, and your brand, then paste in your ad copy and image description and ask the model to score it and explain why. It's rough. But running your creative through even a structured gut-check step before launch is the posture this study is pointing toward. Our AI content strategy work includes this kind of pre-distribution review for content, and the same logic applies to paid creative.

Skip it if

You're not running paid media, or your paid media budget is under $1,000 a month. At that scale, the testing is the learning, and no scoring model is going to save you more than running the experiments yourself. Come back to this in a year when you've got enough data to train on.

The best creative brief in the world is the one you never ran

I've watched a lot of businesses spend their way to data they could have bought more cheaply if they'd paused before launch. The Kroger study is a reminder that the question isn't always "how do I make better ads." Sometimes it's "how do I stop paying to find out which of my ads is bad."

If you were sitting across from me right now, I'd ask you this: what percentage of your current ad creative do you actually believe is working, before you see the results? If the answer is "I don't know until I run it," that's the gap this study is pointing at.

Sources

• Crowdfund Insider: AI Findings Highlight Last Week's Research Roundup (October 7, 2026)