The avatar lottery: cycle your AI creators as hard as your content

Which creator posts a video matters as much as the video, and you can tell whether a creator will ever work within about 20 posts. Across the 110 AI creator accounts we run on TikTok, one account earned 16.9% of all 414 million views, the top three earned 42.3%, and the bottom half combined earned 0.2%. Two thirds of accounts never produced a single 100k video. The ones that did tended to show it almost immediately, and then kept doing it.

Everyone in short-form knows to test hooks, formats, and scripts. The same discipline almost never gets applied to the creator itself, because with human creators you cannot: hiring and firing people in 20-post auditions is not a strategy. With AI creators it is a config change, and the data says it is one of the most valuable changes available. This post is the account-level companion to What 414 million views taught us.

The account power law is steeper than the video one

Videos follow a power law: our top 1% of videos earned 47.7% of all views. Rank the accounts instead and the drop is even steeper:

The median account has 54,000 lifetime views. The best has 70.1 million. You close that spread by finding more accounts like the top ones, and the rest of this post is about how to find them.

The 20-post audition

If account quality were luck, hits would land at random points in an account's life. They do not. Among the 35 accounts that ever produced a 100k video, the median first hit came at post 7:

20 of the 35 hit within their first ten posts. 27 of 35 within twenty. The latest first hit we have ever observed came at post 56, so late bloomers exist, but they are the tail, not the plan.

Read that as an audition window: a new avatar gets about 20 posts to show a hit or a near-miss. When generation is cheap, 20 posts is a small price for that much information.

Early form predicts late form

The audition window only makes sense if early performance actually predicts future performance. It does. Take every account with at least 20 posts and split them by whether they hit inside their first ten:

An account that hit early converts 11.8% of its later posts into hits. An account that started cold converts 4.3%. Same platform, same pipeline, same content engine: a 2.8x difference that was visible by post ten. Whatever an account accumulates in its first days (niche fit, the recommendation profile its early posts establish, the audience seed) persists.

Note the cold accounts are not at zero. 4.3% is still a real hit rate, which is why the audition is a rotation policy and not an execution: a cold account gets deprioritised, not deleted.

Winners hit again, and again, and again

The strongest argument for the audition is what happens after an account proves itself. Hits per account, among the 35 that ever hit:

Only 10 of the 35 were one-hit wonders. Seventeen accounts hit six or more times; the tail runs to 39, 42, and 45 hits, and two accounts have 70 hits each. A creator that proves it can hit will usually hit again, so once an avatar clears its audition, the best move in the whole system is pouring volume into it.

The playbook this adds up to

1. Launch more avatars than feels reasonable. Two thirds will never hit, and you cannot know which in advance. One avatar is not a test of anything.
2. Audition in about 20 posts. Median first hit is post 7; 77% of eventual hitters have shown it by post 20. An account that is 20 posts cold has collapsing odds.
3. Rotate, do not execute. Cold accounts still convert at 4.3%; deprioritise them and spend the slots on new auditions.
4. Feed the winners hard. Early hitters convert later posts at 2.8x, and the best accounts hit dozens of times. Concentrating volume on proven creators is where the power law works for you instead of against you.
5. Re-run the numbers monthly. Every figure here is a portfolio average, and averages only show up at portfolio scale.

What this data can't tell you

Allocation feeds back: accounts that hit early got more volume afterwards, which flatters their totals, though not their per-post rates, which is why we report rates. And this is one company's fleet over four months, in our customers' niches: expect the shape (steep concentration, early predictability, repeat hitters) to hold, and treat the exact ratios as ours.

Neither caveat changes the conclusion. Content testing without creator testing is running half the experiment. With AI creators, the other half costs a config change.