Comparing Offline Sales Between Exposed and Unexposed Audiences in DV360

The short answer: yes, but design matters

One of the most valuable questions a media team can ask is whether digital advertising actually drives sales that happen in the real world — in stores, over the phone, or through offline sales teams. Comparing offline sales between exposed audiences (people who saw your ads) and unexposed audiences (people who did not) is exactly how you answer it.

The method is sound, and Display & Video 360 (DV360) supports it — but the credibility of your results depends entirely on how the comparison is designed. A naive exposed-versus-everyone-else comparison is riddled with bias. A properly randomised holdout gives you a genuine read on incrementality.

This guide explains what's possible, the mechanics behind it, and the pitfalls that quietly invalidate results.

Why a simple exposed vs unexposed comparison is misleading

The instinct is to take everyone who saw your ad, look at how much they bought offline, and compare that to people who didn't see it. The problem: those two groups were never equivalent to begin with.

People become "exposed" for reasons that also make them more likely to buy:

• They browse relevant categories, so they're targeted and served more impressions.
• They're more active online, so they generate more ad opportunities.
• Your targeting deliberately selected higher-intent users.

This is selection bias. Exposed users often look like better customers before any ad influence, so they'll show higher sales regardless of whether the ad worked. Comparing them to a self-selected unexposed group overstates impact — sometimes dramatically.

To isolate the true causal effect, the two groups must be genuinely comparable. That means randomisation.

The gold standard: randomised holdout (conversion lift)

The robust approach is a conversion lift study, where eligible users are randomly split before delivery:

• The test group is eligible to see your ads.
• The control (holdout) group is withheld from the campaign, even though they would have been targeted.

Because assignment is random, the two groups are statistically equivalent at the start. Any difference in offline sales afterwards can be attributed to the advertising itself. This is the difference between correlation and incrementality — the sales you would not have gotten otherwise.

Google Marketing Platform offers conversion lift measurement that operationalises this within DV360 for eligible advertisers, and the same logic underpins ghost-ad and PSA-control methodologies used across the industry.

What you need for offline measurement specifically

Digital sales are easy — a pixel fires. Offline sales require you to connect real-world purchases back to the people in your test and control groups. That usually means:

• First-party customer data (hashed emails, loyalty IDs, phone numbers) collected at point of sale.
• Offline conversion imports or a customer data platform that can match transactions to user records in a privacy-safe way.
• A shared, privacy-compliant identifier that lets the platform reconcile who bought without exposing personal data.

Without a way to attribute offline transactions to individuals or matched segments, the exposed/unexposed comparison collapses. Your first-party data strategy is the foundation of the whole exercise.

Methods compared

Method	How groups are formed	Bias risk	Best for
Naive exposed vs unexposed	Observational, self-selected	Very high	Quick directional gut-checks only
Randomised holdout (conversion lift)	Users randomly split pre-delivery	Low	Proving true incremental offline sales
Geo experiments (matched markets)	Regions randomised on/off	Low–medium	When user-level data is limited
Marketing mix modelling (MMM)	Aggregate, no individual split	Medium	Long-term, cross-channel view

For a clean exposed-versus-unexposed read on offline sales, a randomised holdout is the strongest option. Where user-level offline matching isn't feasible, geo experiments are an excellent alternative: you turn campaigns on in some regions and off in comparable ones, then compare aggregate offline sales.

Designing a valid study

1. Define the offline outcome up front

Decide exactly what counts as a conversion — in-store purchase, signed contract, phone order — and how it will be captured and timestamped. Ambiguity here undermines everything downstream.

2. Size the holdout correctly

A holdout that's too small won't reach statistical significance; too large and you sacrifice reach and revenue. The right split depends on baseline conversion rates and expected lift. Offline conversions are often rarer than online ones, so you generally need larger samples and longer windows.

3. Protect the control group

The control must stay genuinely unexposed for the study duration. Watch for contamination from:

• Other campaigns targeting the same users.
• Retargeting pools that overlap with the holdout.
• Cross-device exposure that leaks impressions to control users.

4. Set a realistic measurement window

Offline purchase cycles are longer than a click-through. A durable good or considered purchase may convert weeks after exposure, so your window must reflect the real buying journey — not a 7-day digital default.

5. Account for privacy and consent

All matching of offline sales to audiences must rest on properly consented first-party data and privacy-safe join methods. Build this into the design rather than bolting it on later.

Reading the results honestly

When the study concludes, you're looking for the difference in offline conversion rate between test and control, expressed as incremental lift. A trustworthy readout includes:

• Incremental conversions: sales that happened only because of exposure.
• Statistical significance: confidence that the difference isn't noise.
• Incremental cost per acquisition (iCPA): media spend divided by incremental offline sales — the number that actually matters for budgeting.

Be wary of any result presented without a confidence interval or sample size. A single point estimate with no uncertainty band is a red flag.

Common traps

• Reporting last-touch attribution as incrementality — they are not the same.
• Stopping early the moment results look good (peeking inflates false positives).
• Ignoring seasonality that affects test and control unevenly.
• Overlapping audiences that blur the exposed/unexposed line.

Where this fits in your measurement stack

Exposed-versus-unexposed offline studies are powerful but resource-intensive. Most mature teams use them selectively — to validate high-spend channels or settle strategic debates — while relying on attribution and MMM for day-to-day steering. Running them well takes careful experiment design, clean data pipelines and platform access that not every account has.

If you're setting this up, it's worth confirming your DV360 access and data infrastructure can support lift measurement before you commit budget. Our DV360 managed services and measurement-focused support can help design and execute studies that stand up to scrutiny, and a co-managed setup works well if your in-house team wants to run experiments with expert oversight.

Bottom line

Yes — you can compare offline sales between exposed and unexposed audiences in DV360, and it's one of the most decision-useful things you can measure. The catch is that the answer is only as good as the design. A randomised holdout with reliable offline conversion matching gives you genuine incrementality; a casual observational comparison gives you a flattering illusion.

Get the experiment design and first-party data foundation right, and you'll finally know how much of your offline revenue your digital media is actually creating.

Ready to measure real offline impact?

If you want to build an offline lift study that proves incrementality — not just correlation — talk to our DV360 measurement specialists. We'll help you design the holdout, connect your offline data safely, and interpret the results with confidence.