Open the ad account and retargeting is sitting at the top of every report. Ten times return on spend, sometimes more. Cost per conversion that makes every other channel look broken by comparison. It’s tempting to just shift the whole budget there.
Before you do, it’s worth asking the same question that applies to branded search: is this channel creating the sale, or just showing up right before a sale that was already going to happen?
Why does retargeting always look like the best channel in the report?
Because the audience is pre-qualified. Retargeting only shows ads to people who already visited your site, viewed a product, or added something to a cart, so the conversion rate is inflated by definition, not by the ad’s persuasive power.
How is the retargeting audience different from cold traffic?
It’s a fundamentally different population. You’re not finding new customers with retargeting, you’re re-approaching people who already raised their hand.
Comparing retargeting’s conversion rate to a cold prospecting campaign’s is like comparing the batting average of someone who only swings at pitches down the middle to someone who swings at everything. Of course the selective swinger looks better. The comparison was never fair to begin with, and neither is comparing retargeting’s ROAS to the rest of your account.
What does a retargeting test actually measure?
A fair test compares people who saw the retargeting ad against a similar group who didn’t, not against your cold-traffic campaigns.
This is where incrementality testing earns its keep. The question isn’t “did people who saw this ad convert at a high rate,” it’s “did they convert at a meaningfully higher rate than people who browsed the same products and never saw the ad at all.” Most retargeting reports only answer the first question.
How much of retargeting’s reported revenue is real?
It varies by business, but studies and holdout tests across advertisers consistently find a meaningful share, often a large one, of retargeted conversions would have happened anyway.
Someone who added a $40 item to their cart and closed the tab was often already planning to come back. The ad might nudge the timing. It might not change the outcome at all. The platform’s reporting can’t tell the difference between “the ad caused this” and “this was going to happen and the ad happened to run first,” because last-click attribution gives credit to whatever touchpoint came last, and a retargeting ad is almost always the last thing someone sees before converting.
When is retargeting worth the spend?
It’s genuinely useful for recovering abandoned carts, for long consideration cycles where a reminder actually changes behavior, and for staying visible against competitors who are retargeting the same shoppers.
The cases where it adds real lift tend to share a pattern: the purchase decision wasn’t fully made yet. A $2,000 purchase someone is still comparing across three vendors is a different situation than a $15 impulse buy they already decided on. The former is where a well-timed reminder can plausibly change the outcome. The latter is where you’re probably just paying for a sale you already had.
How do you test your own retargeting?
Run a holdout: exclude a portion of your retargeting-eligible audience from seeing the ads for two to three weeks, and compare their conversion rate to the group that still sees them.
Most ad platforms support this directly through a geo or audience holdout. Keep everything else about the campaign the same, and resist the urge to call it early just because the first few days look flat. If the holdout group converts almost as well as the exposed group, you’ve found your redundant spend. If there’s a real, sizeable gap, you’ve found a channel that’s earning its budget.
The takeaway
A great ROAS number on a retargeting campaign tells you the audience was warm. It doesn’t tell you the ad caused the sale. Those are different claims, and the report in front of you is only answering the first one.
The fix isn’t to abandon retargeting, it’s the same fix that applies anywhere in the attribution problem: stop trusting the number that platforms hand you by default, and run the one test that actually isolates cause from coincidence.