“Don’t touch anything while it’s learning” is how a new Google Ads campaign learns your mistakes by heart. If the conversion goal is wrong or the search terms are junk, leaving it alone does not protect the budget. It protects the problem.

We published The First 28 Days of Google Ads: A Launch Inspection Checklist to give the first month an inspection schedule. This is a second look because a checklist can tell you when to inspect, while the advice that often comes with it says to leave the campaign alone. That combination can cost you money. Fix bad inputs immediately; save your patience for the bidding decisions that need a real sample.

Smart Bidding learns from the conversions and traffic you give it. It cannot tell that a purchase fired twice, that a low-value action became the primary goal, or that a job seeker has no business seeing your ad. Protect a broken setup for two weeks and you have not avoided a reset. You have paid for calibration on garbage.

The freeze protects bad data as faithfully as good data

I used to tell clients not to touch anything for two weeks. I was wrong. The rule sounds disciplined, especially when the alternative looks like an anxious person changing bids every morning. But it lumps two different jobs together: correcting the information a campaign receives and changing the objective it is trying to hit. Those are not the same intervention.

With Smart Bidding, the conversion feed is the bidder’s account of success. If that account is wrong, the system can pursue the wrong outcome with impressive consistency. An audit of 122 ecommerce accounts found 66 with conversion-tracking problems, including micro-conversions set as primary, double counting and broken purchase tags. The same audit found 75 with unmanaged search terms, including missing negatives and uncontrolled broad match. A two-week freeze does nothing to make either input more useful.

On a small budget, waiting is not some abstract concession to the algorithm. Say you spend $3,000 a month and get eight real leads. In two weeks, you might see three or four conversions. That is a thin sample even when the tracking works. If the goal also counts the wrong actions, you have spent half a month collecting evidence for an outcome you did not want. When someone finally checks in week three, the campaign has not earned patience. It has earned a correction.

This is the distinction I want at launch: do not confuse resisting a nervous bid change with refusing to repair a known fault. One protects the test. The other contaminates it.

A Learning label is not the cost of every edit

Google’s Smart Bidding learning guidance names three triggers for Learning status: a new or reactivated bid strategy; a composition change, such as adding or removing campaigns, ad groups or keywords; and a significant change to conversion goals or actions. It also says budget, ROI target, creative and targeting changes do not trigger Learning status or erase historical data. Smart Bidding adapts to those changes in the background over the conversion cycle.

That does not mean every edit outside the three triggers is free. A new target can change bidding behavior without producing a Learning label. A large change in what traffic you allow can change the mix of auctions without wiping the bidder’s history. Learning status is a useful signal, not a complete account of every consequence. The practical question is whether the edit repairs an input, changes what success means, or merely reflects impatience with early results.

The timeline is narrower than the folklore, too. Google says Learning status lasts no more than seven days, ads keep serving, and performance typically converges within one to two conversion cycles. Its advice is to judge performance over two cycles or two weeks, whichever is longer. The September 2026 reframing described calibration in volume terms: up to around 50 conversion events or three conversion cycles. That is an approximate upper benchmark, not a permission slip you need before fixing a broken tag.

On a steady-state campaign, changing tCPA, tROAS or budget does not erase history or trigger Learning. The bidder has bid-landscape data to work from. A new campaign still has to ramp, and repeated target changes inside one conversion cycle can muddy the result even without a Learning label. So there are two kinds of restraint here: avoid unnecessary restarts, and give consequential changes enough time to show what they did. Neither requires you to keep paying for traffic you already know is irrelevant.

Cut junk traffic without waiting for permission

Adding negative keywords to block plainly irrelevant queries is not the same as rebuilding your positive keyword structure. One stops buying unwanted auctions; the other changes the campaign’s composition. Pull the search terms, sort by cost, and add irrelevant queries as negatives. If job hunters are spending your lead-generation budget, the case for waiting is hard to make. You know that traffic is wrong before you know whether week-one CPA will settle down.

Fixing a broken final URL or a disapproval belongs on the same immediate-action list. A destination that does not work cannot produce the result you launched the campaign to get. A disapproved ad cannot serve normally. Neither problem improves because you let a calendar page turn. Check them, correct them, and keep watching what happens next.

Creative changes do not trigger Learning status either. A new headline or image can change how an ad performs, so I would still record the edit rather than pretend nothing happened. But do not treat a creative fix like a new bid strategy. Even older freeze advice allows that small keyword, ad-copy and negative edits are generally safe while urging people to batch changes weekly. Batching may make a report tidier. It does not refund the spend on a query you could have excluded on Tuesday.

If an edit stops spend on traffic you would never pay for, make it when you see the problem. Save the long wait for a decision that actually depends on performance data.

Break learning when the goal itself is broken

A wrong conversion goal is not a minor housekeeping issue. If a micro-conversion sits as Primary, a purchase fires twice, or a GA4 import and a GTM tag both count the same lead, the bidder is optimizing toward a total your backend will never see. Correcting the goal may trigger Learning. Good. What you asked the system to find has changed, and it needs to adapt to the right answer. Waiting two weeks does not preserve useful learning; it preserves miscounting.

The same principle applies when the landing page is broken. I have paused campaigns on day two because a page took nine seconds to load and its form threw an error on mobile. That pause interrupts the launch, and resuming may require another ramp. But no bid strategy learns its way around a form that cannot submit. You can watch the Learning label or fix the page. Only one of those actions gives the next click a chance to become a lead.

Then there is budget math. If your budget can buy four conversions a month, waiting for a date on the calendar will not manufacture the volume needed to evaluate a target. You may need fewer campaigns, consolidated themes, or a daily cap that can buy a readable sample. Those are substantial changes. Make them deliberately, not as a daily response to CPA. But do not leave a thin structure untouched for a month and call the absence of evidence a learning strategy.

The cost of intervening is real: you may restart Learning, disrupt your comparison with the first few days, or need to wait through another conversion cycle before judging the result. The cost of not intervening can be worse. A clean restart beats continued optimization toward a result you cannot use.

The first four weeks: fix inputs fast, change levers slowly

This is not a license to edit everything at once. It is a way to separate urgent repairs from decisions that need time. I would run the first four weeks on three clocks:

  1. Days 1–3: verify tracking and goals immediately. Submit a test purchase or lead, check that it fires once in Tools > Conversions, and confirm that the conversion used for bidding is set to Primary. Compare the reported result with what the backend shows. If a purchase counts twice, a micro-conversion is Primary, or the backend shows 11 orders against 18 reported conversions, fix the goal that day. Yes, a significant goal change can trigger Learning. A restart on a valid conversion is cheaper than spending the rest of the month optimizing toward an invalid one.

A control panel shows input switches left on and bidding levers taped off

  1. Weeks 1–2: clean inputs daily; leave bidding decisions alone. Pull search terms, sort by cost, and add exact negatives for plainly irrelevant queries. Check Policy Manager and final URLs. Confirm the test conversion still fires once. Do not change the bid strategy, target or conversion goal because three days of CPA look ugly. Change a goal if a check proves it is wrong; otherwise, let the campaign collect a sample. The point is not to keep your hands off the account. It is to keep them off the levers you cannot yet judge.

  2. Weeks 3–4: make one bidding decision, then give it time. Count conversions over the last 30 days before changing strategy. My operating rule is to stay with Manual CPC or Maximize Clicks with a cap below about 15 conversions. At 15 to 30, Maximize Conversions without a target is defensible. I would add a tCPA or tROAS target only after roughly 30 or 50 conversions, respectively. Then judge the change over two conversion cycles or two weeks, whichever is longer. Log the change and its date. Repeated target tweaks inside a single cycle make it harder to tell whether the decision worked, even when the status never says Learning.

Those counts are decision rules, not an excuse to ignore a broken setup until a threshold appears. If week one reveals the wrong goal, fix it in week one. If week three produces a disappointing CPA but the goal and traffic are sound, give the bidding change its evaluation window. Inputs get repaired on evidence of a fault; levers get changed on evidence of performance.

Structure needs its own warning because tidiness can become another form of churn. Adding or removing campaigns, ad groups or keywords is a composition change that can trigger Learning. Drip-feed those edits across week two and you may keep restarting the clock in small increments. If an ad group has zero impressions after 14 days, pause or merge it once. If five converting queries share a theme you did not build for, create one exact ad group for that theme and stop. Group necessary structural changes, make them deliberately, and give the result room to show up. Daily hygiene is for negatives and repairs, not a daily rebuild.

A weekly review cannot do an hourly job

Fast input hygiene needs someone watching between scheduled reviews. A Friday check may find a broken purchase tag after five days of spend, a disapproved headline after 400 lost impressions, or a job-seeker query theme after it takes $280. The inspection calendar tells a person when to look. It does not stop junk spend while nobody is looking.

That is why groas runs spend, disapproval, tracking and delivery checks around the clock while a named strategist owns the bidding decisions that need judgment. The work runs on two speeds: respond to bad inputs as they appear, then make a considered change to the levers when the data supports it. For more on the timing, see the deeper guide to how long learning takes and what resets it. I do not want a person changing targets every hour. I also do not want a person discovering on Friday that the purchase tag broke on Monday.

The split is blunt. Break learning to correct a double-counted action, a Primary goal that measures the wrong thing, a broken checkout, or a structure that cannot gather a readable sample. Do not break it because three days of CPA look high or a new bid strategy sounds more reassuring than waiting. The first set changes what the system can learn. The second interrupts it before you know enough.

The agencies billing a percentage of spend will still tell you to sit still for two weeks. Sitting still makes their inspection schedule easy. It does not make your conversion data true. Fix the input the hour you find it, clean junk traffic every day, and give bidding decisions enough volume to mean something. Otherwise, the Learning label may survive untouched. The budget will not.