How TikTok's cold start actually works

Every recommendation system has the same problem with new accounts: it has nothing to work with. This is the cold-start problem, and it is not specific to TikTok — it applies to any system that recommends content based on behaviour.

What the system actually knows about a new account

When an account is created, the signals available are thin: device, approximate location, language, and whatever the account does in its first session. There is no watch history, no like history, no follow graph. TikTok's own published description of its recommendation system lists user interactions — likes, shares, follows, completed watches — as primary inputs. A new account has none of them.

So the first post goes out to a small test audience assembled from very little information. If that audience is poorly matched, the post underperforms, and the system reads the underperformance as a signal about the content rather than about the audience selection.

Why “just post good content” is incomplete advice

It is true that a strong clip can break out from a cold account. It happens constantly. But the variance is enormous, because the initial audience is close to arbitrary. The same clip posted from an account with a legible interest profile and from a blank account are not running the same experiment.

The useful framing: content quality decides whether the audience you reach engages. The account's history influences which audience you reach in the first place. Those are separate problems and improving one does not fix the other.

What changes the picture

The signals that build an interest profile are the ordinary ones: watching videos to completion, re-watching, liking, following creators, and searching. Watch-through is the cheapest and most abundant signal — it happens on every video whether or not the user does anything else.

An account that has spent hours watching one niche has told the system, in the system's own vocabulary, what it is about. When that account posts, the initial audience is drawn from a much better starting guess.

The honest caveats

• None of the platforms publish the weights they assign to these signals, and they change. Anyone quoting you an exact number is guessing.
• Warming up does not make a weak clip perform. It changes who sees it first.
• Activity that looks automated — too fast, too regular, too much — is itself a negative signal. Pacing matters more than volume.

That last point is why Ember's TikTok warm-up runs over hours rather than minutes, and why it ramps rather than starting at full speed.

← All posts