Home / Predictive targeting vs. retargeting

Predictive analytics · Practical explanations

Predictive targeting vs. retargeting

How the methods differ, where they overlap, and how to choose a useful starting point.

Updated · Educational guide

The difference is the audience-selection rule

Retargeting selects people because they previously interacted with a business, product, or site. Predictive targeting selects or prioritizes an audience using an estimated future outcome. The first describes eligibility based on past activity; the second describes model-based selection.

Both can use prior activity. Both can operate in scheduled batches or with frequent updates. Calling a system “predictive” does not tell you how quickly it runs or whether its predictions are reliable.

Compare the approaches

Predictive targeting and retargeting
QuestionRetargetingPredictive targeting
Who qualifies?People who meet a previous-interaction rule.Eligible people or accounts selected by an estimated outcome.
What data is needed?A reliable interaction record and eligibility rules.Inputs and historical outcomes suitable for evaluation.
ExampleFollow up after a product-page visit.Prioritize accounts most likely to activate a trial.
Main weaknessA previous visit may no longer be relevant.A model can be biased, stale, or poorly calibrated.
Can they overlap?Yes: the audience can be ranked by a model.Yes: the input audience can consist of previous visitors.

Choose based on the decision

When a simple follow-up rule is enough

Use a clear interaction rule when the next step follows directly from the user's activity and there is little evidence that a model would improve selection. Keep recency, completed outcomes, and contact eligibility current so old activity does not drive irrelevant follow-up.

When predictive ranking is worth testing

Consider a model when the eligible audience is larger than your available capacity and you have enough completed historical outcomes to compare alternatives. First ask whether the existing rule can be improved without adding a model.

When to combine them

Define an eligible audience from previous activity, remove ineligible records, then rank the remaining records. Evaluate the added value of ranking against a simpler ordering such as recency.

An illustrative combined workflow

  1. Start with accounts that used a trial during the previous two weeks.
  2. Exclude accounts that already purchased or are not eligible for contact.
  3. Compare a recency-based list with a model-ranked list at the same daily capacity.
  4. Measure outcomes after the same observation window.
  5. Test whether acting on the list improves results, rather than only predicting who buys.

This is a design example, not a performance benchmark. The strongest approach depends on the quality of your records and the action your team takes.

What about third-party cookies?

Neither label specifies a tracking technology. A prior interaction might be recorded in an account system or a first-party product event. A predictive model might also use those records. “Cookieless” does not mean that there is no personal data or that the data is anonymous.

Document the source and intended use of each input, limit unnecessary collection, and review how the resulting decision affects people. NIST's framework offers a broader structure for considering AI-related risks. NIST: AI Risk Management Framework .

Use the same yardstick

Comparisons become misleading if the model gets a larger budget, a different audience, or a longer outcome window. Specify these conditions before evaluating results, and report both successes and costs.

Separate two questions: “Did the score identify likely outcomes?” and “Did the campaign cause more outcomes?” The measurement guide explains that distinction with numbers.

Sources & further reading