Every marketing team eventually has the same argument. Paid social says it drove the sale. Google says it drove the sale. The numbers add up to 140% of actual revenue, and everyone retreats to the model that flatters their channel.
Why the models disagree
They're not measuring the same thing. Last click credits whoever was standing closest to the finish line. First touch credits whoever started the race. Platform reported conversions use view through windows that let two channels claim the same sale in good faith.
None of them are lying. They're answering different questions, and none of the questions is what would have happened if we hadn't run this?, which is the only one worth asking.
| Paid social claims | 62 sales |
|---|---|
| Paid search claims | 48 sales |
| Email claims | 30 sales |
| Actual sales | 100 sales |
Illustrative, built from a typical client audit. Platform reported conversions routinely sum well past 100% because view through windows let several channels claim the same purchase in good faith.
Attribution asks who deserves credit. Incrementality asks what actually changed. Only one of those affects your budget decisions.
Run a holdout
The most reliable tool available to you, and the most underused. Pick a region or audience segment. Turn the channel off there. Leave it on everywhere else. Wait long enough to clear your sales cycle. Compare.
- Geo holdouts work well for anything with regional targeting
- Give it a minimum of four weeks, longer for considered purchases
- Match the regions on baseline volume beforehand, or the comparison is noise
- Expect an uncomfortable result at least once, that's the point
The first time a client runs one on branded search, the finding is usually the same: a meaningful chunk of that spend was buying clicks from people who'd have found them anyway.
Watch the ratio that ignores channels entirely
Total marketing spend divided by total new customers. Blended CAC. It can't be gamed by attribution windows because it doesn't use them.
If blended CAC holds steady while you scale spend, scaling is working. If it climbs, something is saturating, regardless of what any individual dashboard claims.
Drop in click through rate for top ranking pages once an AI Overview sits above them
Ahrefs, February 2026, 300,000 keywords. A useful reminder that the measurement ground moves under you, which is exactly why holdouts beat models.
Just ask people
A single open text field at checkout, how did you hear about us?, routinely surfaces channels no analytics tool can see. Podcasts, word of mouth, a Slack community, someone's newsletter.
It's self reported and imperfect. It's also the only instrument that captures dark social, and the aggregate pattern across a few hundred responses is more useful than the precision it lacks.
What to stop doing
- Reporting platform claimed conversions as though they're incremental
- Cutting a channel on last click data alone
- Rebuilding the attribution model every quarter chasing a cleaner answer
- Treating branded search performance as proof the brand work is working