Paid search · A worked comparison
Before cutting the loser, date the evidence
The reporting window tells you when the clicks happened. It does not tell you whether their outcomes are complete.
- An early ranking can reverseTwo campaigns of the same age can have different proportions of conversions still to arrive.
- Keep cost attached to the right clicksSwitching to conversion-time counts does not complete the original click cohort.
- Treat the correction as an assumptionA historical delay pattern can guide a forecast, but a changed audience or broken import can invalidate it.
Campaign A costs $50 per conversion. Campaign B costs $100. Both spent the same amount and started on the same day. Moving money from B to A seems like the disciplined decision. It can also be a decision based on which customers finished buying first.
Google Ads explains that conversion lag can make recent CPA appear inflated. The cost is already visible while some outcomes have yet to arrive. The less obvious consequence is comparative: two equally recent campaigns can be unequally complete. Giving both the same calendar window does not necessarily make their apparent efficiency comparable.
The example below is invented, with its event-level arithmetic executed locally. It is not an account export, a campaign experiment or evidence that a particular advertiser should increase spend. Its purpose is to identify exactly what must remain comparable before a budget decision has a defensible denominator.
The same clicks produce a different winner
Assume two Search campaigns each spend $1,000 on eligible clicks on September 1. Each counted purchase has equal value, one click receives full credit, and there are no tracking failures, refunds or later attribution changes. We inspect the same September 1 click cohorts on September 3 and again on September 15.
One spend cohort, two observation dates
| Campaign | Spend | Seen Sept 3 | Seen Sept 15 |
|---|---|---|---|
| A | $1,000 | 20 purchases $50 CPA | 25 purchases $40 CPA |
| B | $1,000 | 10 purchases $100 CPA | 40 purchases $25 CPA |
To reproduce the counts, assign A twenty purchases on September 2 and five on September 11. Assign B ten on September 2 and thirty on September 11. Every one of those purchases belongs to a September 1 click. Nothing about the spend changes between the two readings.
The first report favors A; the completed example favors B. A had already revealed 80% of its eventual purchases at the first reading, while B had revealed only 25%. Equal age did not mean equal completeness. A decision that rewards the early conversion pattern can systematically favor faster journeys over slower ones.
Those completion percentages are known here because we constructed the full event ledger. An advertiser looking at September 3 does not possess September 15’s results. Nor does the example establish that fourteen days is an adequate wait for a real account. Its ending is stipulated so that the arithmetic can be checked.
The conversion-date column changes the question
There is an attractive shortcut: switch to Conversions (by conv. time) so that the report contains sales that happened during the selected week. That is a legitimate view of outcomes, but it is not a repair for an incomplete click cohort.
Google’s current conversion-data documentation distinguishes the two date bases. Standard conversion columns associate the conversion with the click date. The conversion-time columns group by when the conversion happened. In a click-based Search example, the standard view keeps attributed outcomes alongside the spend on those clicks.
Extend campaign B into September 8–14. It now spends $1,200 on a new set of clicks, which produces twelve observed purchases on September 10. The thirty delayed purchases from the September 1 cohort arrive on September 11. The conversion-date view therefore contains 42 purchases for September 8–14.
$1,200 ÷ 42 = $28.57 looks reassuring. But thirty of those purchases came from clicks whose cost belongs to the previous week. For the new click cohort, the observed calculation is $1,200 ÷ 12 = $100. Its final result remains unknown.
The $28.57 ratio is arithmetically correct and mismatched as a cohort CPA. It combines one week’s click cost with outcomes from two click cohorts. Conversion-time reporting can help reconcile sales dates; it cannot, by itself, establish the efficiency of the spend incurred on those dates.
A maturity correction can reverse again
Historical delay patterns offer a way to estimate what remains. Google’s guidance on time to conversion describes estimates based on historical performance, including predictions of conversions still to be reported. Those are estimates, not observed purchases.
A deliberately simple calculation exposes the assumption. Suppose B has ten observed purchases and historical evidence suggests that 25% of eventual purchases usually arrive by this age. Then estimated total purchases are 10 ÷ 0.25 = 40, giving a projected CPA of $25. That happens to reproduce our constructed ending.
Change only the assumed completion rate to 50%. Estimated total purchases become twenty and projected CPA becomes $50. Against A’s completed $40 result, the preferred campaign flips. The same ten observed purchases support opposite rankings under different delay assumptions. This is a sensitivity calculation, not Google’s forecasting algorithm.
A new audience, different product or changed lead-qualification process may alter the delay pattern. Even a perfectly measured old distribution can be a poor description of the new traffic. If the decision changes across plausible completion rates, the honest output is an unresolved ranking rather than a precise-looking corrected CPA.
Build the waiting rule from older evidence
To inspect historical delay, Google recommends a reporting range ending at least thirty days ago, or longer for a longer conversion window, then segmenting Conversions → Days to conversion. That is guidance for obtaining sufficiently complete history, not an instruction to ignore every current campaign for thirty days.
Use a comparable conversion action and segment. Record the click-date range, export time, observed cost and conversions, and the maturity assumption beside any projected result. If you compare early snapshots, preserve those snapshots; a later export of an old date range may already contain outcomes absent from the original review.
Hold the meaning of “conversion” still as well. Our custom-goal analysis shows why an action’s Secondary label alone does not settle its role in bidding. A comparison that changes which actions count is no longer isolating delay.
For an actual account change, retain the proposal and applied change separately from the outcome window, as in the Ask Advisor approval audit. A mature before-and-after report can be more comparable without proving that the change caused the difference. Auction conditions, demand and traffic composition may have moved too.
Waiting cannot repair a missing signal
Customer decision time is only one source of delay. Google’s data-freshness documentation describes differing processing schedules and later revisions. A completed purchase may still be waiting for an import; a failed import may never arrive without repair. Neither condition should be diagnosed from a high CPA alone.
If transaction records show completed sales but the conversion pipeline has stopped, investigate collection and delivery now. If tracking is healthy and customers have not yet completed the journey, preserve the cohort and allow a justified observation period. These require different actions even though the dashboard can look the same.
Later numbers are not guaranteed to improve. Refund adjustments can reduce reported value, and altered measurement can change what is counted. A sufficiently mature conversion total also says nothing by itself about margin, lead quality or incremental demand.
The useful budget review therefore has two dates on the page: when the clicks occurred and when their outcomes were observed. Alongside them sits one explicit uncertainty: how much of this cohort remains unresolved. That makes “cut the loser” a decision that can be examined rather than a reaction to whichever customers happened to buy first.
Google documentation and example assumptions
Primary sources checked 22 September 2026. The worked example contains 77 invented purchase events: 65 from September 1 clicks and 12 from September 8 clicks. Five assertions checked grouping, counts and sensitivity; CPA arithmetic was independently recalculated. No account, API or auction experiment was performed.
