AI in retail media: the shelf is becoming a decision system

A sponsored position can put a product in front of a shopper. It cannot answer the shopper's question for it. Retail media is moving from buying visibility to supplying the evidence behind a choice.

Retailers are adding AI to product discovery, campaign management, recommendations and customer service. A shopper can ask for a product that fits a situation, budget or constraint, and the system must compare the available options before it returns an answer.

Paid media still matters. Its performance now depends more visibly on product information, price, availability, reviews and the experience after the click. Those signals shape whether a product remains a credible option once the ad has done its job.

Treating AI only as a creative tool misses the change in how retail systems evaluate products. The better question is: what information does the retail system need in order to choose us?

The old shelf was bought. The new shelf is also evaluated.

Retail media has always contained an algorithm. Amazon, Walmart, Instacart and other retailers have long used relevance, predicted performance and commercial value to decide which products appear where. Bidding only ever controlled one part of the equation.

AI makes the equation more visible because the shopper may no longer begin with a short keyword. They may ask for a quiet dishwasher for a small kitchen, a running shoe for wet trails, or a birthday gift under a specific budget. The system has to interpret the need, find suitable products, compare trade-offs and explain the recommendation.

Amazon's Rufus is an early public example. Launched in beta in February 2024 and later rolled out to US customers, Rufus uses Amazon's product catalogue, customer reviews, community questions and web information. Amazon says shoppers can use it for product research, comparisons and recommendations. The important detail is the input. Rufus does not need another banner. It needs usable evidence.

The implications extend beyond Amazon. Deloitte's 2025 report on agentic commerce recorded increased AI referrals to US retail sites during 2025. The rate will change as measurement improves. The report points to more product research happening before shoppers reach a retailer's standard search results.

Three phases of retail media

Phase one: placement

The first retail media playbook borrowed from search advertising. Brands bought terms close to purchase, watched click-through rate and conversion, then adjusted bids. Teams tracked catalogue quality and feedback speed alongside click-through rate and conversion.

Phase two: performance

As the channel matured, teams added richer campaign structures, audience signals, retail data and incrementality questions. Retail media became a performance channel with its own version of attribution. That was a useful step, but it also encouraged teams to focus on the number visible in the ad console.

Phase three: assisted choice

AI adds a layer before the click. The system interprets a need and decides which products belong in the answer. It may use a product feed, reviews, delivery promise, price and external information. A product can have a strong ad position and still fail the recommendation if the underlying evidence is thin or contradictory.

This phase can be described as evidence-led retail media.

Evidence-led retail media means managing paid placement, product information and customer signals as one connected system. You still need media buying. You also need to make the product easy for a retailer, a shopping assistant and a customer to understand.

The two ground rules

Titles, attributes, images, descriptions, availability and fulfilment details are not background administration. They affect whether a product appears in a relevant comparison and remains a credible recommendation.

An impression is a delivery event. A decision is a shopper choosing your product, or an agent retaining it as a credible option. The second outcome requires more than attention. It requires a coherent answer to the buyer's question.

Five operating shifts

1. Write product data for questions

The old approach starts with a keyword list. The better approach starts with the questions that determine choice. What is it suitable for? Who should avoid it? How does it compare with the common alternative? What does the customer need to know about size, compatibility, ingredients, noise, delivery or maintenance?

This is not an invitation to fill every field with copy. It is a request for precise, consistent product evidence. The same claim should not appear as a vague promise in the title and a different claim in the specification table. AI systems compare patterns across the information available to them, so contradictions can affect how a listing is assessed.

Operating lesson: build a question set for each category, then map every answer to a source field, a proof point and a customer outcome.

2. Separate paid position from earned position

Blended ROAS can make a weak product look healthy. If a product ranks well only when sponsored, the report should say so. If organic position improves after a campaign changes traffic quality or conversion, that should be visible too.

Report at least four views: paid position, organic position, conversion by traffic source and availability during the measurement period. The exact dashboard will vary by retailer, but the principle holds. Do not let the media console become the only version of reality.

Operating lesson: every weekly report should show what was bought, what was earned and what the product did when people arrived.

3. Treat creative as structured evidence

Retail creative still needs to earn attention, but its job is broader than that. Images and video answer questions quickly. A comparison chart can remove doubt. A clear pack shot can confirm size or format. A demonstration can make a claim believable.

AI shopping systems change the value of this work because they need reliable signals across the listing, not a single headline. Creative, copy and product attributes should tell the same story. When they do, the customer gets a clearer explanation and the listing presents fewer conflicting signals.

Operating lesson: brief creative against the decision it must support, not only the placement where it will appear.

4. Feed the algorithm better traffic

Retail media teams often treat off-platform traffic as a separate acquisition line. That misses the relationship between traffic quality and retail performance. If external activity brings people who are likely to buy, the resulting conversion and sales history may affect the product's future position. If it brings cheap clicks with little intent, the signal may be less useful.

Retail media, paid social, search, creator activity and product data all affect what happens on the product detail page. The retailer sees the behaviour that follows the visit.

Operating lesson: assess external traffic by what it teaches the retail system, as well as by its immediate cost per acquisition.

5. Build a measurement layer above the platforms

Retailers will continue to improve their own reporting. Brands still need a view across the system. That view should connect share of shelf, paid and organic rank, availability, price, reviews, traffic source and conversion.

AI makes this more important because the path from question to purchase is harder to observe. A customer may ask an assistant for a shortlist, visit one retailer, return through a different device and buy later. Perfect attribution is unlikely. Better diagnosis is possible.

Use directional evidence and label its limits. Track changes in recommendation presence, referral quality, product visibility and conversion. Keep a record of catalogue changes and retailer updates so that performance changes can be compared with those events.

Operating lesson: measure the conditions that make demand possible, not only the last event you can attribute.

What AI will not fix

AI can make weak retail media faster. It can create more variants, adjust more bids and produce more reports. It cannot decide whether the product deserves to win without reliable inputs.

It will not solve a product that is out of stock, badly reviewed, overpriced for its category or unclear about its use. It will not make a contradictory listing trustworthy. It will not turn a blended metric into a useful one.

More automation increases the impact of poor foundations because the system can act on poor information at scale.

The core idea

The retail media playbook is moving through three stages: placement, performance and assisted choice. The next model is evidence-led retail media.

The formula is simple:

Start by mapping the questions shoppers ask in your category. Check whether your product pages answer them consistently. Separate paid and organic position. Then connect media decisions to the conditions that shape the retail algorithm.

AI will change the interface again. Teams that track the evidence beneath the interface can adapt as the interface changes.