Meta · Published 2026-08-27 · 4 min read
Testing and scaling in Meta: what has changed
In December I published my campaign structure for testing.
The terminology, briefly
- Test campaign
The campaign the test ad sets run in. An ad set always has a fixed budget, the same for every ad set.
- Scaling campaign
The test ad sets that showed a predicted ROAS above 100% or hit the KPI. We monitor daily and raise or lower the budget by 15% depending on whether we are hitting the KPI.
- 5 creatives per ad set
- budget of 5 × CPA
- a duplicate of the test campaign on the winner
- the history of every test lives in the ad sets of a single test campaign
The scheme is alive. But only eight months have passed and a lot has changed
1. Creatives per test ad set: was 5, now 20–50
- Video
Production has become faster and cheaper. $100 per clip from an editor used to be a normal price. Now I haggle over $20 for five clips. The video editor's profession has not disappeared: you still have to watch for mistakes and model artefacts and keep an eye on competitors' clips. The cost is falling not only through AI generation but through AI editing — cuts, animation and end cards are all faster to do with Claude Code or Codex than by opening DaVinci or Premiere. If it is time to test and there are no new videos to hand, generating concepts is a matter of 30 minutes.
- Images
These have gone 100% to AI generation: "launch an ad set with 50 images around the quit smoking concept". It is faster to launch the campaign than to write a brief for a designer. I have written before that on no project do designers catch up with AI on performance, and AI wins because of a better grasp of how people think, and because of sheer volume of creatives.
Across all accounts we test hundreds of creatives a week, and there is no sense in launching fewer.
2. Uploading through the API
The Meta interface used to be the bottleneck. Uploading 100 creatives by hand took a long time, and testing 500 creatives in a week was not realistic for one person. Now creatives go into the account through API or MCP. The limit on volume has gone, and uploading takes no time.
3. Scout campaigns
I have grown fond of launching scout campaigns across a broad spread of user intents. It works particularly well on new projects where there is no sense yet of which sub-niche to pick. I simply upload 50 images with the prompt "think about what points of view exist, vary the style as much as possible".
4. The "Top GEO / Top Creos" campaign
Also a consequence of AI progress. Once a week I assemble a new scaling campaign: top N geos, top N creatives, last 7, 14 or 30 days.
5. The web funnel
The web funnel has become a creative. It is built in a couple of days with all the integrations, upsells, downsells and the email chain.
- The first test I run on a Meta AEM project: build a web funnel alongside it.
- And the reverse. The first thing I try when a web funnel already exists: buying directly. The creatives are all there, the accounts work — we just add AppsFlyer, which is often already in place, and buy direct.
One works in some places, the other in others. I have no preference. What matters is that on every project the performance of the funnel and of direct buying differ a great deal.
It used to feel as though you had to choose — funnels or direct buying. Now half my clients have a mixed picture: funnel tests here, app tests there.
It is good to work with vibe coders in that respect — they adapt an app to a funnel in a couple of days. The industry is often sceptical about vibe coding, but I am working with two projects right now and both are in the black. Plenty of tests, funnels alternating with apps. In short, I have nothing bad to say about vibe coding.
Conclusion
The December scheme has stayed broadly the same. What changed is the number of creatives and the functional variety of the campaigns.
The full walkthrough, with example campaigns, screenshots and Notion checklists, is covered on the UA Manager 2026 course. On the new cohort I show this launch approach. Starts 15 September 2026, ends 15 November 2026. Registration is open, or just write to me directly: @SiberianPaaanda