Category affinity: how purchase behavior becomes a retail media audience
Most retail media audience plans begin with people who already bought the brand. That works for replenishment, loyalty and win-back. It is a poor answer to a growth brief because the audience is limited to known customers.
Category affinity expands the search. It estimates how strongly a shopper's observed behavior connects to a category or product set, then identifies non-buyers whose behavior suggests that the category is relevant to them.
The important word is estimates. Affinity is a model score, not a fact about a person and not proof that an ad will change behavior. A good affinity product makes its inputs, lookback window, exclusions and validation clear. A weak one gives a familiar label to a broad audience and lets the buyer assume precision.
This guide focuses on the behavioral-propensity question: how to build, activate and test a category-affinity audience. A governed taxonomy and SKU-mapping layer supplies the product universe. The broader predictive-audiences guide covers probability-based segmentation across outcomes; category affinity is the category-specific use case. Shopping missions and moments instead describe the purpose and context of a trip.
What category affinity means
There is no single industry-standard formula for category affinity. The method below is a recommended implementation pattern: define a governed product set, construct transparent behavioral features, score an explicit future outcome and validate the result out of time. A category-affinity score can then represent a shopper's relative connection to a category based on purchases, frequency, recency, spend, brand switching and related-category baskets.
Suppose a shopper has not bought a new protein cereal. Over six months, however, they repeatedly bought Greek yogurt, granola, berries, nut butter and protein snacks. Those purchases may support a high score for a health-oriented breakfast set. The score does not say they will buy the cereal. It says their pattern is more similar to category buyers than the average eligible shopper's pattern.
Affinity can be defined at several levels:
- a broad category, such as pet care;
- a need-state set, such as quick breakfast;
- a price or proposition tier, such as premium coffee;
- an adjacent-category relationship, such as craft beer and premium snacks;
- a brand-entry model restricted to category buyers who have not bought the advertised brand.
The product set must be versioned. A model called "health affinity" is impossible to audit if nobody can list the products that created the label.
Affinity is not retargeting, demographics or declared interest
Retargeting selects people who already performed a defined action. Demographics describe population characteristics. Declared interests come from what people say or choose in another context. Affinity uses observed behavior to estimate a relationship that may extend beyond direct purchase of the advertised product.
Amazon's item-to-item collaborative-filtering paper describes matching each purchased or rated item to similar items and combining those relationships into recommendations (Amazon item-to-item collaborative-filtering paper). Category-affinity modeling works at a different decision level, but the underlying lesson is relevant: behavioral similarity can be built from item relationships rather than from demographic resemblance.
That does not make behavior infallible. A household purchase can represent several people. A gift can look like personal preference. A one-off promotion can distort frequency. The model must therefore be validated against future behavior and used with privacy controls.
Build the product universe before the audience
The first modeling decision is not the algorithm. It is the product universe.
Define:
- the focal category or governed product set;
- direct category buyers;
- advertised-brand buyers to include or suppress;
- adjacent categories allowed as signals;
- products or transactions to exclude;
- the hierarchy and mapping version;
- the eligibility population.
If "premium coffee" is inferred from an inconsistent premium flag, the model will learn catalog errors. If a retailer recently moved ready-to-drink coffee into another branch, a category-only query may omit relevant purchases. This is why the affinity asset depends on the canonical taxonomy and SKU-mapping layer, while remaining a distinct audience product.
Choose signals with a reason
More features do not automatically produce a better audience. Each signal should have a hypothesis.
Direct category behavior
Recency, frequency, spend, units and share of wallet can distinguish routine buyers from occasional purchasers. Normalize for category purchase cycles: recent behavior means something different in fresh food than in mattresses.
Brand switching
Brand switching may be tested as a signal of openness to trial, rather than assumed to prove it. A switching feature can help separate category engagement from stable brand habit if it improves future-window prediction for this retailer and category.
Cross-category behavior
Repeated co-purchase can reveal complementary needs. Baby food and nappies, pasta and sauce, or premium cheese and wine may form useful relationships. Treat the relationship as a hypothesis until it predicts future behavior out of sample.
Basket and occasion context
A category relationship can change with the trip. Health-oriented breakfast behavior on weekday mornings may be more actionable than the same score during a party-shopping trip. Mission and moment layers should be applied after the base affinity is defined, not blended into an unexplainable label.
Negative evidence
Repeated category purchases without response to similar offers, dietary exclusions, incompatible products or recent churn may reduce eligibility. Exclusions can improve precision when they remove demonstrably incompatible cases; each exclusion should be tested for predictive value, reach loss and unfair proxy effects.
A transparent scoring example
A retailer wants to find likely buyers for a new premium cat-food brand. The following is an illustrative workflow, not a performance benchmark.
- Fix the dates. Use the 26 weeks ending on the eligibility date as the observation window. Define the outcome as premium cat-food or launch-brand purchase in a later eight-week window. Keep every outcome-window transaction out of feature construction.
- Define eligibility and suppression. Begin with loyalty-linked households active in the last 90 days. Suppress current launch-brand buyers for a penetration campaign and retain a separate flag for other premium cat-food buyers.
- Normalize transparent signals. Convert cat-food recency to a zero-to-one value relative to the category purchase cycle; convert purchase frequency to a percentile among eligible households; express premium share as premium cat-food spend divided by cat-food spend; and encode adjacent cat-care purchasing and brand switching on the same zero-to-one range. Apply explicit penalties for dog-only baskets and stale or uncertain household identity.
- Choose the method. For a first release, average the normalized positive signals and subtract the documented penalties. This is a rules score, not a trained probability model. A trained model may replace it only if out-of-time validation and monitoring show a material operational benefit.
- Create score bands. Rank eligible non-buyers into fixed quantile bands, preserve an uncertain or ineligible state and store the product-set, feature and score-version IDs.
- Set the activation policy from holdout evidence. In the future window, compare category and brand purchase rates for each band with the eligible-population baseline. Activate only bands that show stable predictive separation, sufficient reachable scale and acceptable suppression effects. Do not invent a universal cutoff or call the top band "high intent."
This sequence lets another analyst reproduce the path from eligible transactions to score bands without treating illustrative mechanics as measured results.
Validate the model before attaching media claims
As a recommended model-development control, use a time-based holdout. Build features from an observation window, then test whether higher score bands predict the pre-specified target behavior during a future window. Do not validate on the same transactions used to create the score, and validate the pattern separately for the retailer, category and intended activation population.
Useful diagnostics include:
- category purchase rate by score band;
- lift over the eligible population baseline;
- precision among the highest-ranked shoppers;
- coverage and reachable audience size;
- calibration, so a score increase corresponds to a meaningful probability increase;
- stability across stores, regions and seasons;
- performance for new, infrequent and high-value shoppers;
- overlap with existing-buyer and retargeting audiences.
An affinity model can be good at predicting category purchase and still be a poor campaign audience. The top band may consist mostly of people who were going to buy anyway. Predictive lift is not incremental lift.
Activation depends on the campaign job
For retention, direct buyers may be the correct audience. For penetration, suppress recent brand buyers and reach category-affine non-buyers. For trade-up, begin with category buyers whose baskets and price mix indicate the relevant proposition. For cross-sell, use a clearly defined relationship between a source category and target category.
IAB's playbook recommends matching in-store content placement to the campaign objective (IAB Digital Out of Home (DOOH) & In-Store Retail Media Playbook). The audience name is not enough. The buyer needs the source data, logic, addressable match rate and media channel.
Channel choice also matters:
- Offsite can reach an affinity audience before an active shopping session.
- Onsite can combine the score with current browsing or search.
- In-store can use a cohort or location-level activation where individual linkage and consent permit it.
The creative should reflect the audience logic. A category non-buyer needs a reason to enter. A frequent category buyer may need a product-level distinction. A switcher may respond to proof of difference rather than a generic category message.
Measure the campaign separately from the audience model
There are two different questions:
- Did the affinity score identify people more likely to perform the target behavior?
- Did media delivered to that audience cause incremental outcomes?
The first is predictive validation. The second needs an experimental or credible counterfactual design.
IAB/MRC guidance distinguishes attributed outcomes from causal measurement and requires clear audience and product scope (IAB/MRC Retail Media Measurement Guidelines). Where possible, randomize eligible shoppers into treatment and holdout groups inside the same affinity bands. This keeps the audience definition constant while testing the effect of exposure.
Report:
- eligible audience size;
- matched and reachable size;
- treatment and holdout counts;
- exposure and frequency;
- category and brand purchase outcomes;
- new-to-brand rules;
- incremental lift with uncertainty;
- results by score band;
- suppression logic and overlap with other audiences.
If the highest-affinity band has high conversion but little incremental lift, the plan may be overpaying to reach inevitable buyers. A middle band may produce lower conversion but more incremental customers.
Common failure modes
Affinity becomes a renamed category-buyer audience
If every shopper already bought the focal category, the product may be useful but it is not an expansion model. State that clearly.
The model leaks future information
Features created after campaign eligibility or from the outcome window make validation look stronger than it is.
The label hides incompatible signals
A single "family affinity" audience may mix baby, school-lunch and household-cleaning patterns. Split the commercial use cases.
The product set drifts
If category assignments change, audience size and score distributions change. Bind the model to a taxonomy and product-set version.
Correlation is treated as persuasion
The fact that two categories co-occur does not prove that advertising one will create demand for the other. Multi-category basket research also warns that products can co-occur because they are complementary, share purchase cycles or reflect unobserved household factors (Marketing Science shopping-basket study).
Buyer checklist
Ask the platform:
- What exact product set defines the affinity?
- Which observation window and purchase cycle are used?
- Are current brand buyers included, suppressed or scored separately?
- Which features drive the score?
- How is household or shopper identity handled?
- How was the model validated out of time?
- What baseline does reported lift use?
- How much does the audience overlap with retargeting?
- Can treatment and holdout be randomized within score bands?
- How do taxonomy updates invalidate or refresh the model?
FAQ
Is category affinity the same as purchase intent?
No. It estimates a behavioral relationship to a category. Intent is more immediate and may not be observable from historical transactions alone.
Can a non-buyer have high category affinity?
Yes, if adjacent behavior strongly predicts future category entry. That prediction should be validated rather than assumed.
Is a high-affinity audience always best?
No. It may convert well because it contains likely buyers. Incremental efficiency can be higher in a lower band with more room for persuasion.
What is the minimum viable affinity model?
A governed product set, a defined eligibility population, recency-frequency features, brand-buyer suppression where relevant, time-based validation and a campaign holdout.
Bottom line
Category affinity is useful when it turns product and transaction data into a transparent, testable audience hypothesis. Define the product universe, explain the signals, validate against future behavior and test media incrementality separately. Otherwise "affinity" is only an attractive name for an audience nobody can audit.
Ready to see how this works in practice?
Footprints AI helps brands and retailers measure what matters. See our customer success stories or get in touch to discuss your retail media strategy.




