Do AI ads work? One brand ran 1,000 of them.

The Dilution Test

Before buying or building AI ad volume, run your current plan through five checks. The AI Vidia team uses this as the first diagnostic on any new account, and it predicts outcomes better than any tool choice does.

1. Count concepts, not assets. If next month's plan is 60 assets from 3 ideas, you are scaling dilution. The published Andy Okay ratio is roughly one concept direction per nine shipped ads; treat anything past 1:15 as a warning light.
2. Name the question each batch asks. A hook matrix is a set of questions to the auction: new audience, new angle, new objection handled. If a batch cannot state its question in one sentence, it is filler.
3. Check the kill rule. Volume without a documented kill threshold turns the account into a museum of losers. Winners scale, losers get killed weekly, and the loop must say so in writing.
4. Locate the brand lock. Something fixed has to survive every variant: the palette, the voice, the character system. If nothing is locked, volume erodes recognition instead of building it.
5. Confirm the feedback loop closes. Last week's winners must shape this week's brief. If production and media buying do not meet weekly, the volume is decoration.

Kevin's take

That constraint explains why the strongest AI accounts look more distinct over time, not less. When exploration is cheap, taste becomes the bottleneck, and taste is a compounding asset.

The Weekly Drop Loop

The tactical system behind the numbers is public on the case study page, and it is deliberately boring: five steps, one week, repeated. This is the loop that has run past 1,000 ads without the account going stale.

1. Weekly brief. One theme, one drop date. The brief is fed by the previous week's winners, so the loop starts from evidence, not from a blank page.
2. Concepts. New directions are developed first, from DOGUE covers to Break Glass. This is the human-heavy step, and it stays human.
3. Variants. Each direction is cut into statics, video and hook variants. This is where AI does the volume work: every format the feed rewards, everything in 9:16 first.
4. The drop. 50 fresh ads land in the ad account, ad-ready, named and organised so the media buyer can launch without a cleanup pass.
5. Testing. Winners scale, losers get killed, and the survivors write next week's brief. The loop closes and starts again.

Proof, with names attached

The reason AI Vidia publishes client names is that unattributed numbers are worthless in 2026. Two named engagements carry the proof here. Andy Okay: 50 new ads a week, 1,000+ shipped, 115+ concept directions, documented on the Andy Okay case study. IndianBites: 142 ads shipped in 11 weeks for a DTC food brand whose account was starving for fresh creative, with winning cohorts running at 2.4x ROAS.

Nobody asks whether photography works. They ask whether the photographer is any good. AI ads have reached the same place faster than anyone admits.

If your team is weighing this build-or-buy decision on price, the 2026 pricing breakdown walks through cost per tested asset across all four buying models, and the volume playbook covers what changes past 100 variants a week.

What to measure while the loop runs

Volume without measurement is just noise with a budget. Four numbers tell you within a month whether concept-first volume is working on your account. Watch the winner rate per drop: at a healthy baseline, a low single-digit share of new ads will out-deliver the incumbent, and zero winners across two consecutive drops means the concepts, not the variants, are the problem. Watch CTR by concept direction rather than by ad, so a strong idea in a weak cut is not killed by accident. Watch frequency on your scaled winners, because rising frequency with flat CTR is the earliest fatigue signal you will get. And watch the concept-to-asset ratio itself; if it drifts past one to fifteen, the account is quietly sliding back into variant-first production. None of this requires new tooling, only that the numbers are read weekly, on the same day the next brief is written.

When AI ads do not work

Honesty about the failure cases is cheaper than learning them on your own budget. AI ads underperform when the product needs demonstration physics that current video models still fumble, when nobody on the team can write a concept brief, and when the account spends too little to feed a weekly test loop. They also fail under variant-first volume, which is the most common case and entirely self-inflicted. If you cannot name a concept owner, fix that before buying any production capacity, human or AI.

The next step

The fastest way to test the thesis on your own account is not a subscription, it is a batch. AI Vidia runs concept-first production for DTC and consumer brands as AI image ads and AI video, delivered ad-ready into your account. Book a 20-minute call and bring your current concept-to-asset ratio; the conversation is a working session on volume without dilution, not a pitch.