Your retargeting line item is buying customers you already had

Retargeting reports beautifully because it takes credit for demand created elsewhere. A geo-holdout test, which you can run without a data science team, tells you what share of that spend is genuinely incremental. It is usually less than the dashboard implies.

Every paid account we inherit has the same shape on first inspection: retargeting and branded search show extraordinary returns, prospecting looks mediocre, and budget has quietly migrated toward the former over several quarters. The account looks healthy. Revenue is flat.

This is not a targeting failure or a bidding failure. It is an attribution artefact, and it is structural.

Why the best-performing line item is often the least useful

Retargeting audiences are, by construction, people who already visited your site. Branded search captures people who already typed your name. Both are populations with elevated purchase intent that you did not create in that moment — you created it earlier, somewhere else, or it arrived on its own.

When those people convert, the last touchpoint claims the sale. The dashboard shows a 9x return, and the natural conclusion is to spend more there. Do that for a few quarters and you arrive at an account that is highly efficient at harvesting demand and no longer generating any.

The tell is simple and worth checking today: if platform-reported revenue substantially exceeds actual revenue, you are paying for conversions you would have received anyway. We have seen accounts where the sum of platform-claimed revenue was over twice what the business actually booked.

Geo holdouts: the test that settles it

Incrementality testing has a reputation for requiring a data science function. Multi-touch attribution models and media mix modelling do. A geo holdout does not — it needs a spreadsheet and the discipline to leave it alone for a few weeks.

The logic is straightforward. Split comparable regions into test and control. Turn the channel off in control. Compare total business outcomes — not platform-reported conversions — between the groups. Whatever difference emerges is what that spend was actually producing.

Running a geo holdout

Pick your geos. Group regions into pairs with similar historical revenue and seasonality. Cities or states work; countries are usually too coarse.
Check the pre-period. Confirm the groups tracked each other for at least eight weeks before you start. If they did not, your comparison will be noise.
Turn the channel off in control. Off, not reduced. A partial cut produces a result you cannot interpret.
Run it long enough to clear your sales cycle. Four weeks minimum, and longer than your median time-to-purchase. Ending early is the most common way these tests get ruined.
Compare total revenue, from your own systems, not the ad platform's reporting. The platform is the thing being tested; it cannot also be the judge.

The uncomfortable part is step three. Turning off a line item that reports a 9x return feels reckless, and someone senior will object. The counter-argument is that if it really is producing 9x, four weeks of holdout will prove it beyond dispute and you will spend into it with far more confidence than you have now.

What the results usually look like

In our experience the pattern is consistent, if uncomfortable:

Branded search is the least incremental line in most accounts. A meaningful share of those clicks would have arrived organically. It is rarely zero — competitors bid on your terms — but it is rarely the number reported either.
Retargeting retains real value at the margins, particularly for considered purchases with long cycles, but the window is usually far shorter than the one being run. Thirty-day retargeting windows are mostly paying to reach people who were coming back anyway.
Cold prospecting almost always under-reports. It creates the demand the other two lines then claim credit for.

The account is not broken. The measurement is. Fix the measurement and the budget reallocates itself.

Rebuild the account around what you learned

Once you know the incremental share of each line, the reallocation is usually obvious — move budget toward prospecting and shorten the retargeting window. But moving budget alone will not work, because prospecting fails for a different reason: it demands far more creative than harvesting does.

An audience that already knows you will tolerate a mediocre ad. An audience that has never heard of you will not. Accounts that shift budget to prospecting without also increasing creative volume simply find a new way to waste the money.

The cadence that works for us is a new concept batch every two weeks, judged on cost per incremental acquisition rather than click-through rate. CTR measures whether an ad is interesting. It does not measure whether it created a customer.

The reporting change that makes this stick

None of the above survives contact with a monthly report built on platform ROAS. If the dashboard everyone looks at still credits retargeting with the revenue, budget will drift straight back within two quarters.

Re-base reporting on contribution margin — revenue net of cost of goods, shipping, payment fees and media — and report it at the account level rather than per channel. Channel-level ROAS invites exactly the double-counting that created the problem. A single blended number that reconciles to the P&L does not.