LinkedIn
I used to hand every new client the same 28-day launch plan: gather data in week one, cut waste in week two, scale winners in week three. It looked organized in a slide deck. It was also wrong about the most important thing: Google was not following my plan.
There is no launch schedule. There are three overlapping mechanisms. Query matching explores what your keywords can buy. The bidder calibrates its predictions against the conversions it sees. Late conversions backfill the report, changing numbers you thought were final. If you mistake one mechanism for another, you can cut a campaign for a CPA it never really had, or rebuild a structure just as it begins to tell you something useful.
Keep one account in mind: a home services campaign spending $200 a day, eventually producing about four leads a day, with most jobs booking three to five days after the click. I will follow that account through the first four weeks, then try the same model on a slower one. The point is not to give you another calendar. It is to show you what each week’s evidence can and cannot tell you.
Mechanism 1: query matching has to find the edges
The first fact is simple: on day one, Google does not know which queries make you money. Your keyword says emergency plumber near me. The auction could involve someone with a burst pipe, a landlord comparing prices, or someone looking for plumbing work. Those are not equally valuable clicks. The system needs signals from the traffic it buys, and you need to inspect what it bought.
With broad and phrase match, matching tests a spread of related auctions. I used to tell clients tight exact match would let them skip this part. I was wrong. Exact is tighter, but Google still matches within its interpretation of intent. It still needs conversion data to distinguish a useful edge from an expensive one.
That is why a new search terms report can look rough. On day three, our home services account might show 42 different queries across 60 clicks: a few obvious fits, many adjacent searches and two that make you reach for the pause button. Some deserve an immediate negative. Others deserve more than a morning’s evidence. By week three, the report may show fewer new queries because matching has begun to concentrate around searches that produced signals, not just because you wrote a clever negative-keyword list.
Value-based bidding can test an unproven query at a lower expected return, too. That is another form of exploration, not a promise that every new search term will become profitable. A messy first sample is not automatically a failed targeting strategy. Block clear junk; do not mistake every unproven query for junk.
Mechanism 2: bids begin with borrowed guesses
Once a query is eligible, the bidder has to price the auction. A new campaign has little conversion history of its own, so its early predictions lean on other available signals, including account history, time of day and device patterns. Those signals can produce a plausible opening bid. They cannot tell the system everything about the particular jobs this business values.
Suppose an emergency call at 11pm is worth more than a drain-cleaning quote form at noon. Conversions help the bidder distinguish auctions that resemble one from auctions that resemble the other. Early on, each new conversion carries a lot of weight. With three conversions in the bank, another one can change the picture substantially. With 60, a single lead is less likely to move it. Thin data produces jumpy bids; a larger sample makes any one result less influential.
The Learning label describes a status in the interface, not the end of that process. Ads keep serving while the label appears, and predictions continue to change after it disappears. Waiting for the label to clear is not the same as collecting enough conversions to judge a target.
Volume gives you a more useful clock. Guidance commonly points to about 30 conversions for judging Target CPA and 50 for Target ROAS. At 30 or more conversions a week, calibration may take about a week; at 10 to 30, one to two weeks; below 10, three weeks or more. Those are working ranges, not appointments on Google’s calendar. Our home services account, at about four leads a day once reporting catches up, is near 28 a week. Week two is the earliest sensible time to expect a substantial sample. Count conversions, not days since launch.
Mechanism 3: reported CPA is incomplete before it is high
A click happens now. A customer may book days later. Then reporting has to catch up. Recent CPA divides spend by a conversion count that may still be missing conversions. That is a problem if you treat Tuesday’s view of Monday as Monday’s final result.
There can be a 24-to-48-hour processing delay, with longer delays for offline imports. On top of that, the standard Conversions column attributes a conversion to its click date. When a job books later, it can appear against an earlier click. The most recent days can therefore look weak until those conversions arrive. I have watched advertisers cut budgets because Monday showed zero conversions on Tuesday, when Monday’s clicks had not had time to finish their work.
Take 10 clicks from our example account on a Monday. Say they eventually produce four booked jobs: one books promptly, two book Thursday and one books the following Monday. On Tuesday night, Monday’s apparent CPA reflects only the first booking. Later, the same Monday has four. The spend did not change. The denominator did.
To see the gap, put Conversions (by conv. time) beside Conversions, then check Tools > Measurement > Attribution > Path Metrics for the Day Analysis. The mix varies: a low-consideration account may record about 90% the same day, while a higher-consideration account may record closer to 60%, with more still to arrive later. Google likewise advises looking at Days to conversion before setting targets from recent dates. Do not read fresh CPA without asking how much conversion lag is still outstanding.
Put the three mechanisms on the same four-week calendar
The weeks below describe our $200-a-day home services account, not a schedule every campaign must follow. All three mechanisms run at once. What changes is which one most distorts the view.
Week 1: wide queries, jumpy bids, missing bookings
The report is least trustworthy when all three mechanisms are moving at once. Query matching is testing its edges. A small number of conversions can move bid predictions. Jobs that will book on day four or five still count as zero today.
For this account, imagine a first-week view of 400 clicks, 11 counted leads and a $127 CPA on roughly $1,400 of spend. A week later, late bookings bring that week’s count to 18 to 20, and its apparent CPA falls to roughly $70–$78. That is not a sudden turnaround. It is the same week with more complete accounting.
I would use week one to verify tracking, protect the budget cap and exclude searches that are plainly irrelevant. I would not rebuild the account because an incomplete CPA looked ugly. If you need the week-by-week checklist for what to touch and what to leave alone, keep it beside this explanation. Fix what is broken; do not edit away uncertainty.
Week 2: backfill makes the first comparison possible
Late conversions from week one land on their original click dates, so last week’s CPA can fall without a single change to the campaign. At the same time, our example account approaches the conversion volume that makes each new lead less dominant in bid calibration. Query variety may start to fall as more spend goes to searches that have produced signals.
Now the report begins to answer better questions. Did the same queries keep producing leads? Which search terms are clearly outside the business? Does the conversion count support the target you set, once you allow for lag? At about four leads a day, $200 of daily spend implies a $50 CPA when those leads are counted. That does not make a fresh $50 reading final; it gives you a figure to test against the fuller data.
This is the week for a first measured round of true negatives and a closer read of volume. Read the backfilled week before judging the current one.
Week 3: concentration becomes more informative
The negatives from week two remove the clearest waste. Matching may also send less traffic to searches that have produced no useful signal. In our example, week one’s $1,400 might have been spread across 120 queries, while week three puts 60% of spend into 15 queries associated with booked jobs. That concentration is useful evidence, though it is not proof that every other query should be excluded.
With more conversions recorded, a single lead should also have less power to swing bids. Our account might settle into a $48-to-$55 CPA range: two steady days, a bad day and a good one. I would not ask the bad day to explain the entire campaign.
Small edits now have a better chance of yielding an interpretable result because the baseline is less noisy. Make one, give it three or four days of data, and look again. Tighten the query mix without replacing the structure you are trying to evaluate.
Week 4: decide whether this account has earned a verdict
By week four, the Learning label may be gone. For this four-lead-a-day account, there should also be enough counted conversions and backfilled bookings to ask whether the setup is working. If CPA remains 40% above target after that fuller read, another week of waiting is not a diagnosis. Inspect volume, the conversion action and whether the target fits what this budget is buying.
That verdict belongs to this account, not to week four itself. The duration people call the learning phase is not a universal 28-day timer. A campaign with fewer conversions or longer lag can still be calibrating or waiting for its reported results to fill in. I have kept campaigns I should have restructured because the status label cleared and I read that as approval. The label clearing does not mean the campaign passed. It means you should look at the evidence rather than the label.
Which edits disturb which mechanism?
Most advertisers talk about an edit as if it presses one big reset button. Our model gives us three more precise questions:
- Does the edit change the queries being bought? Adding or removing keywords, ad groups or campaigns can change the traffic mix. You may need to explore that mix again before comparing it with the old one.
- Does it change what the bidder is predicting? A new bid strategy or a substantial change to the conversion goal changes the task you have given the bidder. Old results may be less useful for judging the new setup.
- Does it change the scale or presentation of the existing setup? Budget and target moves do not, by themselves, erase its history. Creative and audience changes can still affect results, so watch what they do rather than assuming nothing changed.
The interface’s Learning status and the practical cost of an edit are not identical. Discussion of target and budget changes in settled campaigns makes the useful distinction: adjusting a working setup is different from launching a new campaign with no history. Neither is permission to change five things and pretend the next CPA tells you which one mattered.
For our home services account, a clear negative keyword removes traffic we already know is wrong. A wholesale keyword rebuild asks matching and bidding to work on a different campaign. Before you edit, name the mechanism you expect to change. If you cannot name it, you may be reacting to a report that has not finished counting.
Count your account’s clock, not mine
Start with two numbers: your daily counted conversions and your conversion lag. At four conversions a day, you can approach 30 around day eight. At one conversion every three days, reaching 30 takes about 90 days. No 28-day template changes that arithmetic.
Then open Path Metrics and look at when conversions arrive. How much appears the same day? How much arrives after 12 days or more? Allow a full lag cycle before treating the newest CPA as settled. The operational launch schedule for the first two to four weeks can tell you what to check each week; these two numbers tell you how much confidence to put in what you find.
This is also the part I no longer want to do by hand. At groas, query mix, bid shifts and lag backfill get checked continuously. That beats expecting a person to watch a search terms report at 1am, as I used to. A useful read needs enough conversions and enough time for late ones to arrive. The calendar comes second.
Try the model on a slower account
Take the chaotic B2B SaaS account that changed its offer every quarter. Say it spends $300 a day on demo requests, records one qualified demo every two days and takes 14 days to work a lead. The home services calendar will give this account terrible advice.
First, the bid clock: at 0.5 qualified demos a day, 30 take about 60 days. Next, the reporting clock: add a 14-day cycle before assuming recent clicks have had time to show their qualified outcomes. Week two is not a fair read, and week four is not automatically a verdict. The query clock may still be exploring which searches bring qualified buyers rather than merely demo requests, because that distinction has not produced much evidence yet.
So I would hold the structure steady, exclude only clear junk and compare qualified outcomes on a time basis that allows for the delay. I would not declare victory from a cheap front-end demo request, or failure from a recent CPA missing demos that have not been qualified yet.
That sounds slow because it is. A four-lead-a-day home services account can earn a useful read in week two. A qualified-demo-every-two-days SaaS account may need month three. Same three mechanisms, different speeds. Stop asking whether the campaign is out of learning. Ask how many relevant conversions it has counted, which queries produced them and how much of the recent result is still on its way.