The First Fix Changes What Later Truth Can Repair
In a conversion emergency, the controls that look like undo buttons leave very different scars.
The dangerous word in a paid-search emergency is fix.
The purchase tag goes dark while orders keep coming. Someone opens Google Ads and reaches for a fix.
The incident may instead be a real promotion lift, a CRM feed arriving two days late, or a checkout accepting traffic but not money. The same small crowd of plausible responses appears: exclude data, add a seasonality adjustment, correct a conversion, change the target, cut the budget.
Each response can sound like the same request: Please do not learn the wrong lesson from this strange week.
It is not the same request.
One intervention makes a set of clicks unavailable to Smart Bidding while leaving the report alone. One lends the bidder a temporary forecast and then expires. One changes the recorded fate of an actual order. One changes the price or volume the business is willing to pursue from now on. And one can stop the business buying visits to a checkout that cannot take money.
The practical difference is what remains possible afterward.
Every response leaves a different shape across four surfaces: the record, the bidder, future exposure, and the reality the business must still own.
A period can remain visible in the paper trail—the record—even when the bidder cannot learn from it.
A corrected number can arrive later. A site can recover. A promotion can end. But the first response may already have decided whether new truth can repair the record, reach the bidder, change future exposure, or merely explain what can no longer be changed.
That is why the first action in an incident is not a platform action.
Finish the sentence before touching the account
Say what is wrong in one complete sentence.
Not “conversions are down.” Not “the algorithm is seeing bad data.” Those are dashboard descriptions. The useful sentence names the failed relationship.
“Orders are real, but the tag failed to report them.”
“The tag is accurate, and the sale really doubled conversion rate.”
“The CRM outcomes are accurate, but they have not reached Google Ads yet.”
“Google Ads correctly recorded zero purchases because checkout rejected every payment.”
Those sentences can produce nearly identical charts for a few hours. They describe different realities.
Then add a second sentence: The decision that cannot wait is…
Perhaps the next auction needs protection from corrupted measurement. Perhaps a known sale needs more aggressive bids while it is live. Perhaps paid traffic needs to stop because the site cannot sell. Perhaps nothing in bidding needs to change; only the record needs repair.
The order matters. “What happened?” is an evidence question. “What should the next auction do?” is a forecast and a business choice. An emergency can require both answers, but it should not be allowed to smuggle the second into the first.
The bidder can be protected before the record is repaired
Suppose a purchase tag fails for eight hours while orders continue normally.
The immediate risk is not just an ugly report. Conversion-based Smart Bidding may interpret missing outcomes as evidence that similar clicks are less valuable. A data exclusion tells Smart Bidding not to use the affected clicks and the conversions attributed to them. It does not change conversion reporting.
That description contains the first trap: the exclusion applies to clicks, not simply to the hours when the tag was dark. If customers commonly convert several days after an ad interaction, conversions lost during today’s outage may belong to earlier clicks. A tag failure on the 18th can implicate clicks from the 13th when the normal delay is five days. Excluding only the visible outage interval feels precise while leaving part of the corruption available to the bidder.
Google recommends acting quickly, accounting for conversion delay and trying to cover at least 90% of affected clicks. But speed does not make the intervention temporary. Google also advises against removing an exclusion after it has been applied.
Now imagine the missing orders become available from the commerce system. Google’s current guidance permits a later reporting backfill after its prescribed wait of at least five days, while still advising that the exclusion remain in place. Primary conversion columns generally return those outcomes to the date of the ad interaction. The old week can heal on the page.
Only one patient has healed: the record.
Analysts and Finance regain the conversions; Smart Bidding remains deliberately blind to the excluded click cohort, including accurate conversions attached to it later.
A restored conversion is not necessarily restored training.
That one-way consequence should determine whether the exclusion is justified: later evidence can repair reporting but cannot, under Google’s advised handling, make the excluded click cohort and its later accurate outcomes available to Smart Bidding.
The intervention has preserved one future by closing another.
That closure also tells you when not to exclude. If an offline feed is merely late, no exclusion is active and the arriving outcomes are accurate, upload them and restore the regular cadence. Google says Smart Bidding works with offline imports, recommends daily uploads when possible and includes reporting delay in the conversion cycle it learns around. Ordinary delay is not proof of corrupted data. In that path, later truth can still reach both the report and the bidder. Creating an exclusion because the freshest days look frightening would remove the very evidence that resolves the fright.
Surprise is not corruption
Now replace the broken tag with a promotion that genuinely doubles conversion rate.
The chart is still strange. The number is not false. A data exclusion here would turn “unusual” into “unusable” and deny the bidder a real commercial event merely because it will not last.
For a brief event the team knows about in advance, a seasonality adjustment can supply Smart Bidding with an expected conversion-rate change during a defined future interval. Google describes one to seven days as ideal; when the interval ends, the adjustment expires, with no reverse adjustment needed.
The important property is where later truth remains free to travel. During the scheduled interval, the adjustment changes the bidder’s forecast; the record and subsequent learning stay open to the event’s real conversions. It borrows confidence for the sale without asking history to pretend the sale never happened.
If the lift is discovered only after the event, the control has missed its moment. A seasonality adjustment is not a time machine. Let the real outcomes arrive. If the business now wants a different efficiency–volume trade-off, change the target and own a new goal; do not disguise hindsight as damaged data.
Even in an awkward overlap, the distinction holds. If the promotion is real but its purchase tag fails halfway through, the scheduled adjustment can govern bids during the sale. The exclusion can cover only the click cohort whose outcomes became unreliable. A later backfill may restore the sale orders to the report while those particular clicks remain unavailable to bidding.
The final state is intentionally uneven. The restored sale remains visible in the record—the ink beneath the amber veil—while the affected cohort remains unavailable to the bidder. “Weekend fixed” would conceal the most important fact.
Some facts really do change after conversion
The record is not sacred merely because it was first.
A customer can return a product, cancel a booking or reveal that an order was worth less than initially reported. A conversion adjustment can restate that conversion’s value; a retraction can remove it. Here the thing being corrected is the fate of a specific conversion, not the credibility of a period of clicks.
When that conversion is used for bidding, the correction reaches reporting and Smart Bidding together. This is sharper than a spreadsheet edit: after a conversion is retracted or restated to zero, Google says subsequent adjustments to it are ignored. For the narrow case of certain offline imports configured to count every conversion, Google describes a different path: create a new conversion with a changed timestamp. The original conversion does not become editable again.
A data exclusion changes what evidence a click cohort may contribute while leaving reporting open. A retraction changes the recorded fate of one conversion and can close ordinary correction. Both can be reasonable. Neither deserves casual use because the team dislikes how a number looks.
A target or budget change does neither. It changes the exposure the business authorizes next.
An accurately measured bad week should remain bad. If the business is no longer willing to buy the same exposure, change its target or budget. Google says Smart Bidding starts responding to a target change quickly, while achieved performance still needs completed conversion cycles to evaluate. That move changes future auctions. It does not acquit the past.
A looser target can authorize more expensive volume; a tighter one can trade volume for efficiency. A budget change widens or narrows available exposure. The record and bidder remain open to later evidence, but auctions entered or skipped while the new instruction is active cannot be replayed.
The honest zero is the real test
The difficult case begins when measurement works and checkout does not.
Google frames data exclusions as protection from conversion-data problems. Its guidance also mentions outages and scheduled site maintenance, but it does not clearly settle the narrower case in which a site failure is measured correctly and produces real zero sales.
There is no corrupted number to identify. The zero is accurate; what it should teach Smart Bidding about the next auction is not settled.
There are two forecasts competing for custody of one accurate fact.
The first says: The outage was a genuine commercial failure. Let the bidder learn that paid traffic sent to this experience did not convert. This keeps the failure inside optimization, but a repaired site may then spend cautiously while new evidence slowly outweighs a condition that no longer exists.
The second says: The repaired checkout is materially the healthy checkout again. Exclude the outage clicks because their zero conversion rate is true but not predictive of the next auction. This protects recovery, but it deliberately prevents Smart Bidding from learning an accurate cost of the business’s operational failure.
Calling one option “truth” and the other “bad data” avoids the decision. Both begin with the same true zero. The question is whether the failed site and the repaired site should inform the same future auctions—whether they belong to the same forecasting regime.
Consider a constructed case. For 47 minutes, a retailer’s checkout rejects every payment while healthy purchase measurement records zero sales. The operator pauses ads, and the incident log confirms the rejection rate. After the repair, synthetic and live test orders pass before ads resume. Price, inventory, creative, demand and traffic mix are unchanged. Only then does the operator restore the prior budget and exclude the affected click cohort because the failed mechanism has been removed.
That is an owned forecasting judgment, not a data correction.
The report keeps the zero-sales interval. Smart Bidding is prevented from treating those clicks as evidence about the repaired checkout. The budget pause changed exposure only while the business could not sell. The incident review keeps the lost revenue and wasted media visible.
No one gets to call the loss fake merely because the bidder was told not to carry it forward.
Another business could reasonably refuse the exclusion. If outages recur, the repair is uncertain, or operational reliability is a persistent part of the customer experience, the zero may still describe the next auction. In that case the bidder should be allowed to learn it, even if recovery spend becomes uncomfortable.
The open choice is not whether the failure happened. It is whether the failure is over.
Write the scar before pressing the button
An incident is not closed when the chart looks calm. It is closed when the team can state, without merging them, the four consequences that will survive it:
The record now says…
The bidder may remember…
Future exposure changed because…
What truly happened was…
Use the four sentences once, at approval, while someone can still challenge the intervention. They are not a recap ritual. They are a preflight. If the team cannot complete one of them before acting, it does not yet understand the scar it is about to create.
For the constructed checkout case, the sentences are plain. The record retains clicks, cost and zero purchases during the outage. The bidder may learn from the surrounding healthy periods but not the excluded cohort. Future exposure stopped during the failure and resumed after validation. The business bought traffic that could not purchase for 47 minutes.
The sentences do not have to agree on the consequence. They do have to agree on the incident. That is their value.
So do not ask only, “Which control fits this incident?”
Ask: “After we use it, what will accurate information still be able to repair?”
A safe intervention is one whose scar you can name before you make it.