Remarketing metrics: what is actually worth tracking

Every results meeting has an uncomfortable moment. The media report shows the remarketing campaign with the best cost per conversion in the account, someone asks why not double the budget, and there is no good answer. Doubling almost never doubles the result.

The problem is not the campaign. It is that the most-used remarketing metrics measure who converted after seeing the ad, not who converted because of it. It is a small difference in a sentence and an enormous one in a budget decision.

Remarketing works in many cases, and the criticism is not aimed at the channel. The problem is that the number used to prove it works proves nothing, and there is a simple test that does. That test has a name, a method documented by the ad platforms themselves, and it fits any account, even without paid tooling.

Why does remarketing always look like it performs well?

Remarketing looks like it performs well because it does not draw people at random: it picks the ones who already visited the site, opened an email or stopped halfway through a form. That audience converts above average with or without an ad, so the campaign inherits a result that existed before it.

Caption: remarketing metrics only decide budget when they compare the exposed group with a control group

The technical name for this is selection bias. A remarketing campaign does not create intent, it finds the intent that was already there and appears in front of it shortly before the conversion.

Add the mechanics of attribution. Whoever saw the ad and converted lands in the campaign's column, and whoever would have converted anyway also saw the ad, because the list was built with exactly that profile.

The result is predictable. The smallest-budget campaign in the account shows the best cost per conversion, and the team starts arguing for a budget increase based on a number that does not survive the increase.

There is a hard version of this effect, and it deserves to be said plainly. If you switched the campaign off today, part of those conversions would keep happening, through the site, branded search, email or the sales team itself.

The point is not to distrust the channel. It is to recognize that no educational marketing metric that only looks inside the campaign can separate who converted because of the ad from who was going to convert anyway. That distinction is not a methodological detail: it decides whether the remarketing budget should double, stay flat or move to prospecting.

What is incrementality and how do you test it in remarketing?

Incrementality is the share of conversion that exists because of the ad and would not exist without it. Measuring it requires a comparison: one group sees the campaign, an equivalent group does not, and the difference in conversion rate between them is the real result of the investment.

Incrementality is not an agency concept, it is the method the platforms themselves document. Google describes Conversion Lift as an incrementality tool that measures conversions driven directly by people seeing your ads, splitting the audience into a treatment group, which sees the campaign, and a control group, which does not.

The documentation names the output metrics. In a user-based study you get incremental conversions, Relative Conversion Lift, incremental conversion value, incremental cost per action (iCPA) and, when conversion values are assigned, incremental return on ad spend. A geography-based study reads results by region and supports offline data.

The same material carries an important caveat: the feature is not available for all Google Ads accounts, and access runs through your account representative. The official tool exists, but many operations will not be able to use it. Hence the need for a manual test, which any account can run.

The logic of the manual test is the same: hold back a slice of the list, make sure it is not touched by any remarketing campaign, and compare the two conversion rates at the end of the period. Here is how that test is organized step by step, with the mistake that invalidates each one:

Table: an incrementality test that closes the reading of remarketing metrics without proprietary tooling

One necessary honesty about step 6: with small lists, the difference between the two groups can fall within the margin of error, and the test will conclude nothing. An inconclusive result is information too. A channel whose contribution does not show up in a test tends to be a channel whose budget should not grow without a new hypothesis.

List size constrains the test design before any statistics. Google Ads documents 100 active visitors or users in the last 30 days as the minimum for the Search Network, the Display Network and YouTube. Gmail is not part of that volume rule: the page only covers who gets added to the segment, not a minimum size.

The figure of 1,000 that circulates has an explanation. Customer lists uploaded or refreshed from February 1, 2024 follow the 100-user minimum on Search and YouTube; older lists that have not been refreshed since then are still bound to the previous requirement of 1,000 matched users. Before designing the test, check which of the two cases your list is in.

If holding back 20% of the list drops the audience below the serving minimum, the test will not run. In that scenario, the path is to test by period, switching the campaign on and off in comparable windows, with the caveat that seasonality contaminates that design more.

Which remarketing metrics matter and which mislead?

The remarketing metrics that mislead are the ones describing the audience, and the ones that matter describe the difference the campaign made. ROAS, CPL (cost per lead) and conversion rate belong to the first group: they go up when the targeting gets warmer, even if the campaign changed nothing.

Operational indicators are still necessary. ROAS, CPL and frequency diagnose delivery, creative and audience saturation. What they do not do is decide budget. Here is what each indicator actually tells you and what it is for:

Table: a critical reading of remarketing metrics, separating operational diagnosis from budget decisions

The pattern repeats across every dashboard indicator: it describes the audience. Only the comparison with a control group describes the campaign.

There is a practical side effect of that confusion. When remarketing CPL is read as the channel's CPL, the operation starts optimizing for the easiest number and moves budget to where it produces a good report, not enrollments. Consistent conversion rate optimization work solves this with an explicit hierarchy of metrics, defined before the campaign goes live: operational indicators for weekly adjustments, incremental indicators for budget decisions.

Video: the performance monitoring that organizes the hierarchy of remarketing metrics, on the mkt4edu channel (in Portuguese)

How does the attribution model distort remarketing results?

The attribution model distorts remarketing because it distributes credit according to where the touch sits in the journey, and remarketing is almost always a final touch. Models that favor the last click inflate the channel. Models that favor the first contact almost erase it, with the same data.

In Google Ads, the list has shrunk. The attribution models documentation records only two models in operation, last click and data-driven, and states that first click, linear, time decay and position-based are no longer supported. Conversions were migrated to the data-driven model, which is the default for most conversion actions.

A short list reduces the chance of deliberately picking a bad model, but it does not solve the underlying problem. The data-driven model uses the account's own history, so it also operates on a remarketing audience selected by intent.

In the CRM, the variety is wider, and so is the variation in reading. HubSpot documents several attribution models with explicit credit-splitting rules: First Interaction and Last Interaction concentrate all credit at one end, the U-shaped model splits it between the first touch and the lead conversion, the W-shaped model adds deal creation, and Time Decay favors touches closer to the conversion.

The conclusion is uncomfortable. Switching from First Interaction to Last Interaction, the same remarketing campaign can go from irrelevant to decisive in the report, without a single impression changing. That is why the model choice has to be recorded and kept: comparing months with different models produces a trend that does not exist.

And the two questions have to be separated for good, because they get confused constantly. Good marketing attribution work answers who deserves credit for a conversion that already happened. Only an incrementality test answers whether that conversion would have happened anyway. Attribution is accounting, incrementality is causality, and a serious report shows both without pretending they are the same thing.

How do you read remarketing inside the enrollment funnel?

In the enrollment funnel, remarketing changes role at each stage, and incrementality tends to fall as the applicant advances. Someone who only visited the program page may genuinely need a nudge. Someone who already applied and has an invoice issued normally finishes with or without an ad.

That curve is counterintuitive, and it explains a lot of badly read reports. Bottom-of-funnel campaigns show the best apparent ROAS exactly where the real contribution tends to be smallest, because the audience is about to convert on its own.

In the middle of the funnel, the effect inverts. The lead who downloaded an asset, talked to the team and disappeared is the case where paid reinforcement has the most room to change the outcome, even though the cost per conversion looks worse on the dashboard.

A second caution is the window. A program decision cycle is long and does not fit inside a closed month of media reporting. Reading remarketing by calendar month mixes applicants at different stages and produces a meaningless snapshot. The useful reading is by cohort: following everyone who entered the list in the same week through to the end of the process, exposed or not.

None of this works without a unified database. The educational marketing strategies that manage to measure remarketing properly have one thing in common: the same contact is traceable from the first click to enrollment, with the stage recorded.

When data is split between the ad platform, a spreadsheet and the academic system, no holdout test holds up, because you cannot even guarantee the control group stayed out of everything. Marketing data analysis in HubSpot solves the operational side of that, because it lets you mark the control group as a contact property and filter any report by it.

Which remarketing report should you take to the budget meeting?

The report that works in a budget meeting has six blocks and a numeric recommendation at the end. It opens with the holdout result, not with campaign performance, because the question around the table is always the same: what happens if we cut this.

A format that supports the conversation is organized like this:

Table: a report structure built for budget decisions, with one business question per block of remarketing metrics

Three things are deliberately left out of that report. Standalone ROAS, a screenshot of a platform dashboard and a cross-channel CPL comparison with no context adjustment do not belong, because none of them answers a question from the table.

Declaring uncertainty out loud is part of it too. If the test did not have the volume to conclude, the report says so and proposes a better design, instead of presenting a fragile number as a finding.

Running this routine depends on where the data lives. A CRM for educational marketing that records stage, source and test group on the same contact turns that report into a saved view, not manual month-end work.

On the media side, the same logic applies to campaign structure. Paid media work designed to be measured is born with exclusions, naming and windows already prepared for the test.

Video: the paid versus organic comparison that remarketing metrics have to support in the budget meeting, on the mkt4edu channel (in Portuguese)

Frequently asked questions about remarketing metrics

How do you decide whether the remarketing budget continues?

The remarketing budget continues when an incrementality test shows conversion that would not have happened without the channel. The platform dashboard does not answer that: through selection bias, it credits remarketing with the enrollment of someone who had already decided, and the channel looks great precisely because it harvests a result that was coming anyway.

Remarketing is still one of the most efficient uses of media available, precisely because it talks to people who already showed interest. What does not hold up is deciding budget with an indicator that rises whenever the audience gets warmer.

Swapping campaign ROAS for cost per incremental conversion moves the conversation. Instead of defending a pretty number, the team presents what happens to enrollment if the channel goes dark.

Sometimes the answer is uncomfortable, and the test shows half the budget was buying conversions that were already coming. That is good news in disguise: there is idle money available to fund new prospecting.

If you want to build this measurement into your operation, with test design, campaign structure and a report that holds up in the budget meeting, talk to a specialist at mkt4edu for an assessment.