Category-Level vs Basket-Level Purchase Signals for Retail Media: What Each Can and Cannot Answer

Third-party purchase data can reach a retail media plan at different grains, and this piece compares two GSDSI feeds at opposite ends. In the category-level feed, a row says a device bought in the snacks category on Tuesday. In the basket-level feed, a row adds what was paid per unit and how many units were bought. Both are called purchase data, yet they answer different questions, so the grain belongs in the brief.

A few terms first. CPG (consumer packaged goods) covers grocery, household, beauty and beverage products. A MAID (mobile advertising ID) is a resettable identifier that a phone's operating system assigns for advertising; see the MAID glossary entry. IAB (Interactive Advertising Bureau) is the advertising trade body that published the retail media guidelines cited here, and IAB Tech Lab is a related nonprofit consortium that produces technical standards. MRC (Media Rating Council) is the industry body that audits and accredits media measurement products and data sources. A SKU (stock keeping unit) is one specific product at one size.

Where third-party purchase data fits in retail media

Third-party purchase data sits beside a network's own data, not in place of it. How a network's first-party data, off-site extension and measurement fit together is covered in the data layer behind retail media networks. The short version comes from the IAB/MRC Retail Media Measurement Guidelines (January 2024). They treat third-party audience segmentation as optional. They describe it as a way to reach audiences a retail media organization cannot identify with its own data. They also require transparency in how the matching is done. Within that frame, outside purchase data serves audience extension, off-site activation and a cross-retailer view.

Category-level vs basket-level purchase signals

In the two GSDSI feeds compared here, both grains record an observed purchase tied to a mobile device ID. Other purchase data can be keyed to other identifiers; the IAB Tech Lab disclosure schema lists cookie, mobile, platform and user-enabled ID types.

GSDSI's CPG transaction data, tied to MAIDs, comes at both grains. The category-level live feed places each transaction in an IAB-level product category. The basket-level feed adds unit price, product IAB code, quantity and hashed-email linkage, with about three years of history. The table below gives the size of each feed. They are separate feeds counted over different windows. Their counts should never be added together. The CPG transaction feed page lists the fields and delivery options.

IAB Tech Lab maintains the shared taxonomies the industry codes against. They include an Ad Product Taxonomy for describing the product or service being advertised. That is a classification, not a SKU catalog. Taxonomies are versioned, and a major version can break compatibility with the one before it. Ask which taxonomy and version a feed's codes follow before you map them to your own category tree.

What category-level signals can answer

Category-level signals are built for breadth. They show which devices bought snacks, bottled water or hair coloring in a period, and which of them also appear at other retailers. That supports off-site audiences of recent category buyers. It supports category trend reads once you define the denominator and the panel's coverage. It also gives a cross-retailer check on whether a network's reported buyer growth looks like the category's. Our share of wallet guide covers the denominator problems in cross-retailer reads. The daily CPG signals explainer goes deeper on the daily read.

What a category row cannot answer is anything that depends on the item. It will not separate a premium buyer from a value buyer. It will not show what was paid. It will not tell you whose brand was bought.

What basket-level signals can answer

Basket-level signals add the item. Unit price and quantity turn a category buyer into a shopper with a price point and a buying volume. With the product IAB code you can see how a device's purchases spread across product types within beverage or household goods. You can also see whether the price paid moves over time. About three years of history lets you compare the same season year over year instead of guessing at seasonality.

Some popular questions need more than these fields. Brand switching and trade-down to store brand need brand or SKU identifiers on each transaction; an IAB code is a classification and will not supply them. Pack-size trends need package attributes, because quantity counts units bought, not what is inside each one. Same-basket analysis, one of the five cross-shop shapes, needs a key that groups items into one transaction. Ask whether the delivery carries those before you plan on them. The transaction-level primer covers what item-level data supports once it is in place.

The trade-off is reach. For these two GSDSI feeds, the basket panel covers far fewer devices than the daily category feed. Whether it is enough depends on category coverage, the destination's match rate and what the campaign needs, so it suits analysis and seed audiences more readily than large off-site flights.

Limits of third-party purchase data in retail media

Third-party purchase data has four limits that apply at either grain.

• Panel, not census. A third-party feed observes a sample of shoppers. It can support estimates of a category or a share when the sample base is disclosed. It is not a retailer's sales ledger, and its totals should not be expected to match one.
• A device is not a person. Both feeds are keyed to MAIDs. Hashed-email linkage on the basket feed associates a device with an email identifier. It does not verify who was holding the phone, and one household can carry several devices.
• Counting windows are not attribution windows. A daily unique count and a monthly unique count describe different populations. An attribution window is a separate choice, covered in the next section.
• Labels can hide modeling. A segment built from observed purchases is a different product from a look-alike expansion of it. The look-alike modeling entry explains the difference.

Which attribution window and grain fits each retail media question

An attribution window is the period over which an ad exposure can be credited with a sale. The IAB/MRC Retail Media Measurement Guidelines (January 2024) require each window to be empirically supported and disclosed before a campaign runs. They note that windows of 3, 7, 14, 28 or 30 days are typical in digital measurement, and they decline to prescribe one. They encourage day-level detail so reports with different windows can be reconciled. When a short campaign is projected to a full year, the forecast basis must be disclosed. The table below pairs common questions with a grain and a window rule.

Where each grain enters a retail media measurement test

An incrementality test compares a test group that was eligible to receive the ads with a comparable control group that was not. In a randomized holdout, analysis follows the original assignment (intent to treat), so a test person who never saw an ad still counts in the test group. A matched design is a different method: when no holdout is available, the control is built from unexposed people matched to those exposed, and it needs stronger assumptions. Holdouts and pre-registered methods are covered in the parent post. Purchase grain matters at three points in the test.

1. Matching test and control. When a holdout is not available, the IAB/MRC guidelines describe synthetic or forensic matching on pre-campaign purchase behavior for the brand, the category and the retailer. That works only when the matching variables account for the confounders, and the guidelines suggest validating the match with A/A and B/B tests before the A/B read. Category-level data can add an outside view of category buying to the match. It does not rescue a weak match.
2. Reading reach. Purchase records show who bought, not who saw an ad. Counting off-site reach among category buyers also needs exposure data matched on the same identifiers.
3. Reporting the sample. The guidelines ask for relative sample sizes, sampling and non-sampling error, and guidance such as confidence intervals. A third-party read should carry its own sample base for the same period.

The guidelines list a clean room as one route for matching third-party data. A clean room is a controlled environment where parties join data and run approved analyses without seeing each other's raw records; see the clean room glossary entry. The guidelines treat deterministic data, such as login and loyalty records, as the baseline for incrementality, and the deterministic vs probabilistic entry explains the difference.

What to ask a purchase-data vendor before delivery

A purchase-data vendor should be able to answer these questions before any sample ships. The IAB Tech Lab Data Transparency Standard already defines many of them. Its disclosure schema asks a seller to declare the audience precision level, the data sources, the inclusion methodology, any look-alike or cross-device expansion, the refresh cadence and the source lookback window. The precision options are individual, household, business, device, browser and geography. Loyalty card transaction, online ecommerce and offline transaction are separate source options.

• Which taxonomy and version do the product codes follow, and has that changed during the history delivered?
• Which fields arrive on each row: price, quantity, product code, brand or SKU identifiers, and a key that groups items into one transaction?
• What is the counting window for each unique-device figure, and how does it differ from the refresh cadence?
• What is the sample base for each period, and how does coverage vary by retailer type?
• What does the hashed-email linkage associate, and what does it not establish?
• Which match path reaches the destination you will activate in?
• Can you test on a sample file cut to your exact category and period?

IAB Tech Lab's taxonomies README also describes an Audience Taxonomy whose top tier says whether a segment is demographic, interest-based or purchase-intent based. Ask which label a purchase segment carries.

Category vs basket purchase data decision table

A worked example: one brief, both grains

Take a bottled water brand planning a summer flight on one network. The IAB Ad Product Taxonomy shows why the version matters here. Version 1.1 lists Bottled Water under Water; version 2.0 lists only Water under Beverages, and the official 2.0-to-1.1 mapping has no Bottled Water entry. The network can reach the buyers it identifies. The brand also wants recent category buyers it cannot see, plus a pricing question: in the panel, are observed bottled water purchases carrying a lower unit price than in past summers?

The reach half is a category-level job. Recent bottled water buyers, keyed to device IDs, feed the off-site audience, with the match method written down as the guidelines require. The pricing half is a basket-level job: unit price and quantity on observed panel purchases under the bottled water code, compared season against season with about three years of history. That works if the delivery uses the 1.1 Bottled Water node consistently across seasons; otherwise run it at the Water level. Report it with the sample base for each season, meaning how many devices and purchases sit behind each figure, because the result describes the panel's purchases, not every bottled water buyer. Asking how the brand's own buyers moved would also need brand or SKU identifiers on each transaction.

How to choose the grain for a retail media brief

Choosing the grain starts with the question: reach, depth or measurement. Then name the identifier, the counting window and the codes you need, and test a sample on that exact cut. Retail teams can see how these feeds fit wider programs on the retail industry page. If you have a specific category and question in mind, send it to GSDSI's team to scope a sample at the right grain.