LinkedIn
I had a home services campaign whose bids kept wobbling for five weeks. The owner checked the Learning column every morning like it was a microwave timer, waiting for the ding. I used to tell clients to give it seven days. I was wrong.
Seven days describes how long a useful sample often takes to collect, not how long the model needs by decree. Smart Bidding needs conversions to judge which auctions are worth buying. If those conversions arrive slowly, the calendar keeps moving while the evidence barely changes. The Learning badge and the underlying learning are not the same thing, either. That distinction explains why a low-volume campaign can feel stuck after the badge clears, and why most tricks for skipping the wait only make the wait start over.
First, the model needs to know what a conversion looks like
Every auction asks Smart Bidding a version of the same question: how likely is this click to produce the conversion you defined? Not clicks in general. This query, on this device, in this location, at this hour, with this wording. Google uses those auction-time signals to set a bid.
It can draw on broader patterns, but it still needs evidence about your business. A home services campaign might record calls; a store might record purchases. Neither gets much help from a model that confidently predicts the wrong outcome. The signal that matters is the conversion you ask it to pursue, recorded often enough for the model to connect outcomes back to auctions.
That is why conversion history matters more than the campaign’s birthday. An account with useful history may have a head start over a new one, and Google can use signals across campaigns in the same account. But an older campaign does not become well calibrated merely because it has been sitting there. No conversions, no calibration. Thin conversions, slow calibration.
In the home services account, the question was not, “How many days has this run?” It was, “How many calls has it recorded, and how long does a click take to become one?” Start there before you touch a setting.
Then separate the badge from the bids
Google says the visible Learning badge lasts no more than seven days. Ads keep serving while it appears, and Smart Bidding continues learning from historical and new conversion data after it disappears. A cleared badge is not a certificate saying the bids have settled.
For performance, Google points to one to two conversion cycles as the usual window for convergence. A conversion cycle is the time between a click and the conversion it produces. If your outcome takes time to show up, this week’s bids cannot be judged fairly against this week’s incomplete results.
The same distinction sits behind the rough benchmark of up to about 50 conversion events or three conversion cycles. Fifty is not a pass mark, and three cycles are not a universal deadline. They describe the evidence and elapsed feedback a model may need, not a clock you can watch count down.
Think of seven days as an estimate for filling a sample when conversions arrive quickly. The badge tells you about a visible status; the conversion data tells you how much the model has had to work with. When those disagree, judge the data.
Three calls a week changes the calendar
The home services campaign was closing about three calls a week. At that pace, collecting something like 30 to 50 events takes ten-plus weeks. Its Learning badge cleared within seven days; its bids did not magically settle on day eight. The five-week wobble stopped looking mysterious once I counted the calls.
The rough practitioner pattern tracks that arithmetic: 30-plus conversions a week may settle in about a week; 10 to 30 can take one to two weeks; fewer than 10 can take three-plus weeks and may sit in Limited. Those are useful planning ranges, not promises. Google also weighs conversion volume, cycle length, and how restrictive your settings are.
A tight tCPA target, a starved budget, or narrow targeting can leave the campaign fewer auctions to learn from. A long click-to-qualified-lead window delays the answer to the model’s question. If that window is 10 days, judging over three of your own cycles puts you at roughly a month. Checking CPA after 48 hours does not make the missing conversions arrive sooner.
If you get four conversions a month, you do not have a seven-day problem. You have a sample problem. That is the fact to carry into every edit.
Next, ask whether an edit changes the evidence
Not every change sends Smart Bidding back to the same starting point. Google names three triggers for the Learning badge: a new or reactivated strategy, a composition change such as adding or removing campaigns, ad groups, or keywords, and a goal change. It also says budget, target, creative, and targeting edits do not erase history or trigger Learning status.
That does not make every non-trigger harmless. It means you need two separate questions: Did the edit change what the model is learning, and did it change the auctions where the model must use that learning? The first can bring back Learning status. The second can move performance without bringing back the badge.
A strategy switch is not a target adjustment
Switching bid strategies changes the job you are asking the system to do. Moving from Max Conversions to tCPA, or from tCPA to tROAS, is not the same as changing a number within an established strategy. The switch can trigger Learning because the strategy itself is new or reactivated.
By contrast, Google product manager Tal Akabas said on Ads Decoded that changing a target or budget on an established campaign does not trigger Learning because bids can scale from existing auction data. That qualification matters. An established campaign has history to scale from. If you make a sharp target change on day three, the absence of a new badge does not give you much evidence that the new target will work.
I would not treat the status column as permission to keep adjusting. Change the target when you have a reason, then let conversions show you what happened.
A budget move can wobble without resetting
The 20 percent budget-change rule that circulates in forum threads is folklore, not Google policy. Google says budget edits do not trigger the Learning badge. It does not say a doubled budget will leave performance untouched.
More budget can put the campaign into auctions it previously could not afford. Less budget can cut auctions that were contributing both conversions and useful feedback. The model retains its history, but the conditions in which it bids have changed. Same history, new territory.
That is why I would make the budget change deliberately and judge it over incoming conversions, not the 48 hours after clicking Apply. If I change the target at the same time, I have made the result harder to interpret. The badge might stay off throughout. The test is still messy.
A new conversion goal changes the answer
Swap what counts as a conversion and you change the outcome the model is trying to predict. That is a goal change and a Learning trigger. Old rows may no longer answer the new question.
This is especially easy to feel in B2B. A form fill, a qualified call, and a closed deal are different outcomes, recorded at different points. Add an offline import with a 14-day lag, and a quiet first week may tell you more about reporting delay than bid quality. I walked through the cost of getting that measurement wrong in how a B2B SaaS team fixed tracking and doubled pipeline.
Google’s current guidance favors a single primary goal, enough budget to feed it, consolidated structure, and no stacked target changes inside one conversion cycle. That does not mean choosing an easier goal just to make a counter go up. The goal still has to represent the outcome you want the bids to pursue. Fix tracking once, take the reset once, then leave it alone long enough to measure the result.
Why “skip the wait” usually starts another wait
Most shortcuts change the container instead of adding evidence. Start on Max Clicks, then flip to tCPA. Duplicate a campaign to dodge the badge. Rebuild the structure because the first week looks ugly. Those moves introduce a new strategy, a new campaign, or a composition change: the very things Google lists as Learning triggers.
A missing badge on a new object is not proof that the object has a useful conversion history. Nor does a loose target tightened every day give you a clean read on which target worked. You may avoid staring at one status label while making the bids harder to evaluate.
The less theatrical approach is to collect more relevant signal without repeatedly changing what the signal means. Use one primary conversion goal rather than several competing ones. Consolidate where splitting campaigns leaves each with too few conversions. Give the campaign enough budget to enter the auctions it needs to test, then avoid stacking edits inside a conversion cycle.
If you run tCPA, tROAS, or Max Conversions, choose a strategy your conversion volume can support. Hold it long enough to see the outcomes. Stillness is the shortcut, not because doing nothing is clever, but because changing the question every few days makes the answers less useful.
Work the model through three accounts
The same reasoning applies whether you sell purchases, calls, or qualified pipeline. Count the conversions, note how long each takes to appear, and ask what a proposed edit does to the evidence.
- A store with 35 purchases a week doubles its budget. It runs Max Conversions, and the budget change does not trigger Learning status. But the extra spend may reach auctions the campaign was not buying before. Expect some wobble; judge the move after another meaningful batch of purchases, not 48 hours of CPA. Change the target on Monday too, and you have made it harder to tell which move caused what.
- A local services business has three campaigns, each recording four calls a month. Waiting seven more days will not make any one campaign high-volume. Consolidating them could put roughly 12 calls a month into one campaign’s sample. That is better than three thin samples, but it is still thin. Changing the primary goal could raise the count only if it captures more real, relevant outcomes; renaming weak leads as qualified calls fixes nothing.
- A B2B team has a 10-day gap from click to qualified lead. It replaces form fills with qualified pipeline as its primary goal. That is a goal change, so it should expect a new Learning period. About three of its own cycles puts the first serious assessment around a month out. Calling the change a failure in week one ignores outcomes still working their way through the pipeline.
The model does not promise that consolidation, patience, or a cleaner goal will rescue a weak offer. It tells you what you can reasonably learn from the conversions you have. That is enough to stop treating a status badge as a verdict.
Launch for a sample you can actually collect
Suppose your forecast is 12 conversions a month. Split that into four tidy campaigns, and each may see only three conversions a month. One consolidated campaign collecting all 12 still needs time to build a substantial sample, but at least the evidence is arriving in one place. I used to launch segmented on day one because it looked organized. It cost clients time spent guessing.
I would start with the primary goal I actually want, choose a bid strategy the volume supports, and resist setting a target the early data cannot justify. Once the campaign has useful history, I can make an edit and give the next conversions time to tell me whether it helped.
My edit checklist is short:
- Change one thing per conversion cycle when you need a clean read. Do not change budget, target, and structure together.
- Expect movement after a large budget change, even if Learning status never appears. Judge it against incoming conversions, not two days of CPA.
- Treat a conversion-goal swap as a new measurement question. Allow for the conversion lag and avoid piling other changes into that window.
This is the unglamorous part of the job, and it is why groas runs accounts with continuous adjustments inside guardrails, an action log that records the reason for each move, and no appetite for stacked resets that waste the sample you just paid to collect.
Now take a new campaign recording four qualified conversions a month. Day seven comes; the badge disappears. What changed? The visible status. What did not change? The fact that the model has only a handful of answers to its auction question, with more still on the way if your conversion cycle is long. I would check the goal, volume, and cycle length before touching the target. Stop watching the Learning column. Count conversions instead.