How many AI actors to test per product before you scale. The casting math media buyers use to find a winning face fast without burning credits on dead variations.
HighQualityUGC Team
Performance Creative
We run UGC ad tests daily and publish what holds up: real credit costs, real hook rates, no vendor fluff.
Most media buyers pick one actor, love it, and scale it. Then it fatigues in nine days and they have no bench. The problem was never the fatigue. The problem was casting a sample size of one.
AI UGC changed the economics here. When a real creator costs 100 to 500 euros per video, you cast conservatively because every face is a line item. When an AI actor costs a few dollars a render and burns credits only after you approve the storyboard, casting becomes a testing variable like a hook or a CTA. You should treat it like one.
This post gives you the casting math: how many actors to test per product, how to structure the batch, and when to stop testing and start scaling the winner.
The face on screen changes conversion the same way a hook does. Same script, same product, same edit, different actor: you will see meaningfully different CTR and cost per acquisition. Buyers who skip this are leaving the single easiest test on the table.
Creative is where the leverage is. Nielsen found creative drives 47% of an ad's sales effectiveness, ahead of reach at 22% and brand at 15%. Casting sits inside that 47%. It is not a cosmetic choice.
Source: Nielsen, cited by Supermetrics
The takeaway: stop asking "which actor do I like." Ask "which actor moves the metric." Those are different questions and only one of them has a spend attached.
Test 4 to 6 actors in your first casting batch. That is enough spread to find a face that beats your gut pick and small enough that you can fund each variant to a real read.
Here is the logic. With 3 actors you cannot tell a real winner from noise. With 10 actors you either underfund each one or blow your test budget before any of them clears a decision threshold. Four to six is the range where each variant gets enough impressions to trust and your budget still covers the batch.
Match the number to your daily test budget. A rough rule: each variant needs enough spend to reach roughly 1,500 to 2,000 impressions before you judge it. If your test budget funds four variants to that bar and not six, run four. Do not spread a fixed budget so thin that every actor dies undecided. An undecided test is a wasted test.
The takeaway: 4 to 6 is the default. Your budget, not your enthusiasm, sets the exact number inside that range.
Casting is a single-variable test. Same script, same hook, same product framing, same edit, same music, same CTA. The only thing that changes across the batch is the actor. Supermetrics is blunt about this: test one variable at a time and isolate your variants or you cannot trust the result.
If you change the actor and the hook at the same time and one wins, you have learned nothing. You cannot attribute the lift. You are back to guessing, which is the exact thing testing was supposed to end.
This is where consistency tooling earns its keep. If your actor looks slightly different in every render, you are not testing casting, you are testing render drift. Contact-sheet consistency keeps each actor on-model across every take so the only difference the algorithm sees is the person, not a wardrobe or lighting glitch you did not intend.
The takeaway: lock every variable but the face. If you cannot keep the actor and product on-model between renders, fix that before you spend on a casting test.
Here is how to run a clean casting round without wasting spend or credits.
The takeaway: settle the script, cast wide, judge at a real impression threshold, keep the top half. The whole round is one variable and one decision.
Stop casting when one actor beats the batch average on cost per result by a clear margin and holds it past the impression threshold. That is your winner. Scale it and stop second-guessing.
But do not scale only one. Keep the top 2 or 3 actors from the round on a bench. Ad fatigue is not a maybe, it is a schedule. Meta's own guidance treats format and creative diversity as a core driver of Advantage+ performance, and the algorithm optimizes better when you feed it genuinely different creative. A bench of proven faces is how you rotate before a winner craters instead of scrambling after.
The math on volume backs this up. UGC-style tools push per-video cost down toward 10 to 50 euros versus 3,000 to 12,000 euros for studio, which is why testing more concepts at the same budget is the whole point.
Source: AdLibrary 2026 UGC Guide
The takeaway: scale the winner, bench the runners-up, and start your next casting round before the current winner fatigues. Casting is not a one-time event. It is a rotation you keep feeding.
A few patterns kill casting rounds before they produce a usable answer.
Casting for looks instead of fit. The most attractive actor is not the most credible one for your product. A skincare buyer trusts a face with visible texture over a flawless one. Cast for believability against the objection, not for the prettiest render.
Judging too early. Calling a winner at 400 impressions is calling a coin flip. Wait for the threshold. Every buyer who ignores this ends up scaling the wrong actor and blaming the platform.
Testing too many things at once. Actor plus hook plus CTA in one batch is not a test, it is a mess. One variable.
Never rerunning the round. A winning actor from March is not automatically a winning actor in June against a fresh audience. Recast periodically.
The takeaway: cast for credibility, judge at the threshold, isolate the variable, and rerun. Skip any one of those and the round lies to you.
Cast 4 to 6 actors per product. Hold everything constant but the face. Fund each to a real impression threshold. Kill the bottom half, scale the winner, bench the rest, and rerun before fatigue hits. Casting is a performance variable that sits inside the 47% of effectiveness that creative controls. Treat it like the test it is, not the taste call it feels like.
