Incrementality: When Retargeting Looks Better Than It Really Is

TLDR: Retargeting is incremental only when customers eligible to receive the ads buy more than a comparable group intentionally withheld from them. Platform-reported ROAS cannot answer that causal question by itself. Run a randomized holdout or another credible lift test, calculate incremental orders and incremental ROAS, and keep funding only the retargeting programs that create profitable additional demand.

The central problem in retargeting incrementality ecommerce measurement is simple: the easiest customers to target are often the customers most likely to purchase without another ad. A shopper who viewed the same product three times, started checkout, and then returned through a retargeting ad may deserve to be counted as an attributed conversion. That does not mean the ad caused the order.

This distinction can turn an apparently excellent bottom-funnel campaign into an average or unprofitable use of budget. The practical answer is not to abandon retargeting. It is to treat every retargeting configuration as a causal hypothesis: did showing these ads to this audience produce additional orders, revenue, or contribution margin?

Attribution and incrementality answer different questions

Attribution assigns credit according to tracking rules: touchpoints, attribution models, identity matching, and conversion windows. Incrementality estimates what happened because advertising ran. Google Ads similarly distinguishes attributed conversions from incremental conversions measured by comparing a treatment group with a control group, and defines incremental CPA and incremental ROAS using the measured lift rather than all credited conversions.

Measurement approach	Question answered	Control group	Main use	Retargeting risk
Standard attribution	Which tracked touchpoint receives credit?	No	Reporting journeys and optimizing delivery	Can credit an ad for capturing demand that already existed
Incrementality test	How many additional outcomes did the advertising cause?	Yes	Budget allocation and causal evaluation	Requires enough data and a well-protected control
Blended business reporting	Did total store performance improve while spending changed?	Not necessarily	Monitoring revenue, spend, and efficiency	Other channels and commercial changes can obscure the cause

Attribution remains useful. Operators need it for campaign diagnostics, creative comparisons, conversion paths, and day-to-day optimization. The mistake is asking an attribution report to prove causality. A creator code creates a similar issue: redemption identifies a checkout event but does not necessarily identify who caused the purchase, as explained in this guide to creator-code attribution.

Why retargeting can look stronger than it is

Retargeting audiences are selected because they have already demonstrated intent. Product viewers, cart abandoners, checkout starters, and previous customers usually have more purchase propensity than a cold audience. If some of them return naturally and buy, a retargeting platform may receive credit whenever the journey satisfies its attribution rules.

Several effects can inflate the apparent value:

Selection effect: the campaign targets people whose earlier behavior already predicts conversion.
Conversion lag: a shopper may need time rather than an ad before completing the purchase.
Channel overlap: email, SMS, branded search, affiliates, creators, and paid social may all touch the same customer.
Promotional substitution: a discount may accelerate or subsidize an order that would have happened at full price.
Audience overlap: multiple campaigns may pursue the same shopper and report credit under different systems.
Existing-customer demand: a previous buyer may repurchase because of product satisfaction or routine replenishment, regardless of the ad.

That does not make retargeting useless. A randomized field study involving an online home-improvement seller found that retargeting increased website revisits and that the effect weakened as more time passed after the original visit. The outcome was revisitation, however, not proof of universal purchase lift. The study summary from Stanford Graduate School of Business is a useful reminder to separate attention effects from commercial outcomes.

Results can also change with the message and offer. A randomized cart-abandoner study reported in a Marketing Science Institute working paper found no measured purchase lift from ads without promotional offers in that particular setting, while outcomes varied across promotion treatments and audience conditions. That is evidence for testing specific configurations, not a universal verdict on cart retargeting.

How to run a credible retargeting incrementality test

1. Define the budget decision first

Write down what the result will change. For example: “Should we continue spending $15,000 per month retargeting recent product viewers?” That is more actionable than trying to prove whether retargeting works in general.

Choose a primary outcome before looking at results. Completed orders are usually clearer than clicks or sessions. Revenue can capture basket-size changes, while contribution margin is better when discounts, shipping subsidies, payment fees, returns, or variable fulfillment costs materially affect economics.

2. Define one eligible audience precisely

Separate audiences with different intent and economics. A seven-day checkout starter should not automatically share a test with a past purchaser or a casual product viewer. Record the inclusion window, purchaser exclusions, geography, device rules, consent requirements, and treatment of customers who move between lifecycle stages.

Testing segments separately helps answer a more useful question: where does another paid impression change behavior? It may turn out that retargeting adds value for product viewers but mostly intercepts natural recovery among checkout starters. The reverse is also possible.

3. Choose the strongest feasible control design

A native platform lift study is attractive when the platform can create randomized treatment and control groups and connect them to a reliable purchase event. Google documents user-based Conversion Lift measurement, while Meta Blueprint discusses Conversion Lift, cross-channel impact, and geo-based approaches. Availability and requirements vary, so operators should verify current options in their own accounts rather than rely on fixed eligibility thresholds.

A customer-level holdout can work when the brand can assign eligible users consistently and suppress the control group from the tested advertising. Geo testing is an alternative when user-level assignment is unavailable, but regions need to be sufficiently comparable and insulated from spillover. Geography can also introduce differences in seasonality, media costs, inventory availability, shipping promises, and customer mix.

4. Keep everything except the tested exposure comparable

The control must be protected from the advertising being evaluated. At the same time, treatment and control should experience the same ordinary store conditions. Avoid changing prices, landing pages, shipping offers, inventory, or promotional calendars for only one group.

Email and SMS do not always need to be switched off. If the question is whether paid retargeting adds value on top of the existing lifecycle program, keep those flows operating consistently in both groups. Suppress or redesign other activity only when that is part of the predefined test. Document campaign overlap so a supposedly withheld shopper is not reached by another campaign carrying the same retargeting message.

5. Measure purchases in a dependable backend system

Use deduplicated completed orders rather than adding together conversions reported by multiple platforms. Reconcile cancellations and returns according to a rule established before the test. Apply the same attribution-independent order definition to treatment and control.

Do not end a test simply because the dashboard briefly shows a favorable result. The required duration and sample size depend on baseline conversion rate, expected lift, traffic, conversion lag, and acceptable uncertainty. If the experiment is underpowered, the responsible conclusion may be “inconclusive,” not “no effect.”

A fictional iROAS calculation

Consider a fictional store that randomly assigns 50,000 eligible product viewers to treatment and 50,000 to control. The treatment group can receive retargeting ads; the control group is withheld. During the measurement period, treatment produces 1,500 orders and control produces 1,400.

Treatment conversion rate: 1,500 ÷ 50,000 = 3.0%.
Control conversion rate: 1,400 ÷ 50,000 = 2.8%.
Incremental conversion-rate lift: 3.0% − 2.8% = 0.2 percentage points.
Estimated incremental orders: 50,000 × 0.2% = 100.
At an $80 average order value, incremental revenue: 100 × $80 = $8,000.
At $12,000 in media spend, incremental CPA: $12,000 ÷ 100 = $120.
Incremental ROAS: $8,000 ÷ $12,000 = 0.67.

Suppose the platform attributed 400 orders and $32,000 in revenue to the same ads. Its reported ROAS would be 2.67. Both reports can be internally consistent: one applies attribution rules, while the experiment estimates that only 100 orders and $8,000 in revenue were additional. The figures are fictional and not benchmarks.

A 0.67 iROAS would usually be difficult to justify on revenue alone, and contribution margin would make the result less attractive. But the decision threshold belongs to the store. Operators should account for gross margin, discounts, shipping subsidies, returns, payment costs, and any credible downstream customer value without using speculative lifetime value to rescue weak immediate economics.

Turn the result into a budget decision

Observed result	Likely interpretation	Practical response
Positive, sufficiently precise lift with profitable economics	The tested program created commercially useful additional demand	Scale gradually and retest after meaningful audience, creative, or budget changes
Positive lift but iROAS below the required threshold	The ads caused sales, but not efficiently enough	Reduce bids or frequency, narrow the audience, change creative, or stop the configuration
Near-zero purchase lift	Most attributed purchases may have happened without the ads	Cut spend or test a materially different audience, timing, message, or offer
Visit lift without purchase lift	The ads generated attention without enough additional buying	Audit product-page continuity, offer quality, price, delivery proposition, and checkout friction
Wide uncertainty around the estimate	The experiment cannot distinguish useful lift from noise	Improve power, measurement, assignment, or test duration before making a large decision

Do not average every retargeting tactic into one verdict. Dynamic product ads, cart reminders, replenishment messages, win-back campaigns, and promotion-led retargeting address different customer states. One can be incremental while another merely harvests demand.

Pre-launch test checklist

State the exact budget decision and primary success metric.
Define the eligible audience, recency window, exclusions, and lifecycle stage.
Separate product viewers, cart abandoners, checkout starters, and past purchasers when volume permits.
Randomize or match groups before campaign exposure.
Confirm that the control can be withheld from the tested ads.
Document overlapping paid campaigns and audience expansion settings.
Apply email, SMS, search, affiliate, and creator policies consistently across groups.
Keep prices, promotions, inventory, shipping promises, and landing experiences stable where practical.
Use deduplicated backend orders and establish rules for cancellations and returns.
Set the economic threshold using revenue and contribution margin as appropriate.
Allow for conversion lag and assess statistical uncertainty.
Record the decision rule before reading the result.

The operating recommendation

Retargeting deserves neither automatic credit nor automatic dismissal. It should earn budget by producing incremental commercial value for a defined audience under a defined setup.

Start with one meaningful segment and one budget decision. Create a protected control, measure additional orders and revenue, calculate incremental CPA and iROAS, and then evaluate contribution margin. Scale configurations that produce profitable incremental orders. Constrain, redesign, or stop the ones that mostly place an advertising claim on demand your store had already created.