Retargeting Needs a Counterfactual: The B2B Holdout Playbook

Retargeting is the channel most likely to look efficient while doing the least incremental work. It reaches people who already visited, searched, opened an email, or appeared in a sales conversation. When one of those accounts converts, the retargeting platform is nearby enough to claim credit. The report looks tidy. The budget decision is not.

The fix is not replacing last-click with a more elaborate attribution dashboard. It is giving retargeting a counterfactual: what happened to comparable accounts that were eligible for the campaign but deliberately held out from exposure? That result should validate the platform signal, shape the CRM readout, and calibrate the broader measurement model. It is the same discipline behind calibrating MMM with experiments.

Why retargeting needs a counterfactual

Platform attribution answers a useful operational question: which audience, creative, or campaign was associated with a conversion? It is good for diagnosing delivery, comparing creative paths, and spotting broken tracking. It does not answer the budget question: would this account have progressed without the ads?

That distinction matters more in B2B because the audience is already self-selecting. A target account may return through direct traffic, speak with sales, attend a webinar, or search for the brand after seeing an ad. A last-touch rule can assign the eventual form fill to retargeting even when the buying motion was already underway.

A holdout makes the question harder to evade. Google’s geo-experiment research describes the core design clearly: assign non-overlapping regions to treatment or control, keep the assignment intact, and compare the outcomes. For an account-based B2B program, the unit can be an account or a carefully defined account cluster rather than an individual cookie. The design is less glamorous than a multi-touch model, but the decision it supports is stronger.

Design the test around accounts, not clicks

Start with an account list that is stable enough to remain addressable during the test. Freeze eligibility rules before launch. Do not let sales priority, website activity, or late-stage intent move accounts between treatment and control after assignment; that leaks outcome information into the experiment.

Define treatment as eligibility to receive the retargeting program, not proof that an individual saw an impression. Meta’s Conversion Lift documentation uses the same intent-to-treat principle: compare people eligible for ads with a control group that is not eligible, rather than conditioning the result on exposure. Judge the campaign assignment, then report delivery as a diagnostic.

Choose an outcome that matches the buying motion. A form fill is easy to count and easy to overvalue. For a longer sales cycle, use qualified opportunity creation, progression to a sales stage, or closed revenue when the observation window allows it. Pre-commit the primary outcome and decision rule, or the team will select the most flattering metric after the test ends.

Protect the control group from contamination. Suppressing one retargeting campaign is not enough if brand search, display partners, paid social, or sales-triggered audiences can reach the same accounts. Record every activation that could expose the control. If a clean account holdout is impossible, label the comparison observational—not experimental.

Reconcile three measurement layers

The test result should not replace the platform report. It should tell you how much confidence to place in it.

The first layer is delivery and tracking. LinkedIn’s Insight Tag documentation describes post-click and view-through conversion tracking, audience creation, and campaign associations. Use that layer to inspect whether the tag fires, whether conversion rules are attached to the intended campaigns, and whether the account audience is being populated. These are implementation checks, not causal proof.

The second layer is the holdout readout. Compare treatment and control on the pre-committed outcome, then inspect the result by account tier, buying stage, and creative family. A weak overall result can hide useful lift among accounts that had not entered an active sales cycle. A strong platform report with no separation between treatment and control is a reason to reduce confidence, not increase spend.

The third layer is the planning model. Google’s Meridian announcement describes using incrementality experiments as priors for marketing mix modeling. Use the holdout to anchor a causal estimate, then use an MMM for cross-channel allocation and scenarios. A platform dashboard should not answer an allocation question, and an MMM should not decide which creative to pause today.

If the test has too few usable regions for a conventional geo analysis, Google Research’s time-based regression work provides an alternative for estimating a counterfactual response. Disclose the assumptions that make the estimate credible.

What to do on Monday

Create a measurement brief before changing the campaign. Name the eligible account population, treatment rule, control rule, primary outcome, observation window, contamination risks, and decision rule. Have marketing operations own the assignment table and have finance or revenue operations sign off on the outcome definition.

Then build a reconciliation sheet with one row per account: assignment, delivery status, platform conversion, CRM stage, opportunity value, and final outcome. Keep the platform report beside the causal readout instead of blending them into one score. The gap between those columns is not an embarrassment. It is the evidence you need to decide whether retargeting creates demand, harvests existing demand, or does both for different account states. That reconciliation follows the logic in the last-touch attribution rebuild.

Retargeting earns more budget when it survives that test. Until then, treat its attributed conversions as a useful signal—not a verdict.