The Cost of a Video Ad Fell 90%. Here's What That Broke
When a usable 15-second ad costs the price of a coffee, production stops being the bottleneck. The constraint moved to judgement, rights, and platform trust — and most teams haven't adjusted.
Key takeaways
- ✓Production capacity is no longer a moat; taste and testing discipline are.
- ✓Volume without a kill rule just produces expensive noise at scale.
- ✓Rights, likeness, and disclosure are now the real risk surface for AI-made ads.
- ✓Best results come from hybrid assets: real product or customer footage plus generated environments and variants.
What changed in the economics
A short-form ad that required a shoot, an editor, and a week of turnaround can now be assembled in an afternoon from text prompts, generated footage, synthetic voice, and automated captioning. For a founder-led brand, that is the difference between four assets a month and forty.
The consequence is not simply cheaper marketing. When everyone can produce forty, the auction fills with competent, forgettable video, and average performance regresses. The premium moves to the small number of concepts that say something true and specific.
The new bottleneck: selection
- ●Define the performance floor before launch, not after you get attached to an asset.
- ●Test at the concept level first, then permutate only proven winners.
- ●Cap tests to 3–5 days with a fixed minimum spend per variant.
- ●Keep a written library of winning hooks — that library is the real asset, not the video files.
Where AI video still looks like AI video
Hands, text on physical objects, complex physics, and continuity across cuts remain the recognisable tells. So does the uncanny middle ground of a synthetic human speaking directly to camera about a real product.
The practical fix is hybrid production: shoot the product, the founder, or the customer for real, and use generation for backgrounds, b-roll, localisation, and aspect-ratio variants. Audiences forgive stylisation far more readily than they forgive a fake person making a claim.
Compliance is the underrated risk
- ●Ad platforms increasingly require disclosure for realistic synthetic media, especially in political and health categories.
- ●Provenance metadata is becoming a default expectation for published assets.
- ●Never generate a recognisable person's likeness or a competitor's trademark, however easy the tool makes it.
- ●Keep a record of prompts, models, and source footage for every published ad.
Frequently asked questions
Do AI-generated video ads perform worse than filmed ones?+
Not inherently. Performance tracks the strength of the hook and offer. What consistently underperforms is generic AI output with no specific claim, proof, or point of view.
Do I have to disclose that an ad was made with AI?+
It depends on platform and category, but disclosure requirements for realistic synthetic media have been expanding. Assume disclosure is required for anything photoreal involving people, and check current platform policy before launch.
How many video concepts should a small brand test per month?+
Five to ten genuinely distinct concepts, each with a handful of generated variants, suits most budgets under €20K per month. More than that usually outruns the team's ability to judge results properly.
Sources
- [1]Sora product and safety documentation — OpenAI
- [2]Runway research and product releases — Runway
- [3]Content provenance and authenticity specification — Coalition for Content Provenance and Authenticity (C2PA)