Delivery
The TikTok learning phase: what really resets it
Half the advice about the learning phase is folklore. Here is what is worth acting on, and how to tell from your own data whether an edit cost you a restart.
What the learning phase actually is
When an ad group starts — or changes significantly — the delivery system has little information about who converts for this particular combination of creative, audience, bid and offer. It explores. Exploration is expensive: results are more variable, cost per result is usually higher, and the numbers you see are a poor guide to what the ad group will do once it stabilizes.
Once enough conversion events accumulate in a short enough window, delivery moves from exploring to exploiting, and performance becomes readable. That transition is the thing everyone is trying to reach — and the thing careless edits keep pushing back.
The one rule that matters
Changes that alter who is targeted, what is shown, or how the system bids will restart learning. Changes that do not, generally will not. Everything else in this article is a consequence of that sentence.
Edits, ranked by how disruptive they are
| Change | Typical impact | Notes |
|---|---|---|
| New ad group | Full restart | Unavoidable. Plan the budget for it. |
| Changing the optimization goal or conversion event | Full restart | You are asking the system to find different people. Treat it as a new ad group. |
| Switching bid strategy | Full restart | Especially between cost-cap and lowest-cost style strategies. |
| Significant targeting edits | Restart likely | Adding or removing large interest sets, changing geo, changing age brackets. |
| Adding or replacing creatives | Partial | The ad group may keep its footing, but the new ad itself has no history. |
| Large budget change | Partial | Doubling budget changes the pacing problem the system is solving. Smaller steps are safer. |
| Small budget change | Usually safe | Moderate adjustments in the 20–30% range are normally absorbed without drama. |
| Renaming, changing schedule labels, editing tracking names | Safe | Cosmetic edits do not affect delivery. |
| Pausing and resuming within a short window | Usually safe | Long pauses are a different matter — delivery effectively starts cold again. |
Exact thresholds are not published and do change. Which is precisely why you should measure the effect in your own account rather than trusting a table — including this one.
How to tell whether an edit reset learning
You do not need platform confirmation. The data shows it:
- Cost per result variance jumps. Day-to-day CPA swings widen sharply compared with the previous week.
- CPM moves and impression volume becomes erratic. The system is exploring new inventory.
- Conversion rate drops without any change in landing page or offer. Traffic quality is temporarily less targeted.
- Delivery timing shifts. An ad group that reliably spent evenly starts pacing differently.
If three of those move together right after an edit, you restarted learning, whatever the interface says.
Practical habits
- Batch your edits. If an ad group needs a new bid strategy and new creatives, do both at once and pay for one restart instead of two.
- Do not judge during exploration. Evaluating an ad group mid-learning and killing it is how good ad groups die young.
- Give it enough budget to finish. An ad group whose daily budget is a small multiple of its target CPA may never accumulate enough events to stabilize. Either fund it properly or consolidate it into something that is funded.
- Consolidate where you can. Fewer, better-funded ad groups reach stability faster than many thin ones — the most common structural mistake in accounts that never seem to get out of learning.
- Keep an edit log. Half of all "TikTok is broken today" reports are explained by an edit someone made yesterday.
Reading the curve
The useful mental model is a cost curve that starts high and noisy, falls as the system learns, and flattens once it has enough signal. Your job during the noisy part is to keep the ad group alive long enough to reach the flat part — and to avoid resetting it back to the beginning by accident.
In Freshlytics: learning-phase entries and exits are tracked per ad group, edits are recorded in an audit log next to the performance data, and rules can be configured to skip ad groups that are still learning — so automation never kills an ad group mid-exploration. Request beta access.