Retail media on your own site: where the ad pound starts eating the merchandise margin

2026-08-20

Retail media is the most profitable thing most retailers have ever sold, and the mechanism that makes it profitable is the same mechanism that makes it dangerous. It monetises attention the retailer has already paid to acquire. Nothing new is manufactured, warehoused, picked or delivered.

Ecommerce
Retail Media
Strategy
Merchandising

What can be concluded from this is that it is a governance issue and not a media issue, which is the reason why I continue to bring this point up when having conversations with commercial directors. The revenue goes into one line and the costs into another, with no one being asked to reconcile them.

Why the pressure is structural rather than managerial

The advertising revenue derived from a retailer's own channels by BCG amounts to a margin of 70% to 90%, whereas the margin for off-site extensions, after allowing for media and agency costs, is 20% to 40%. When this is compared with the core retail operating margin, which is typically 3% to 4%, the figure speaks for itself. The amount of money earned from one pound of on-site advertising is several times greater than the amount obtained from one pound of merchandise revenue.

Walmart is the best example of this. In its 2026 financial year, global advertising revenue reached nearly 6.4 billion dollars, an increase of 46%, and the chief financial officer stated to investors that advertising and membership together accounted for about one third of the profit in that quarter. He referred to it as a structural change in the way the business earns money, a phrase which chief financial officers do not use casually.

In the UK, the market is younger but moving in the same direction, and here I have to flag a sourcing problem rather than smooth it over. The IAB UK Digital Adspend series put retail media at 3.8 billion pounds in 2025, up 18% year on year. IAB UK's Compass research on retail media, published separately, forecasts UK retail media spend surpassing 6.6 billion pounds in 2025 and on-site spend reaching 8.6 billion pounds by 2030 at a 17% annual rate. Those two numbers come from the same trade body and don't agree because they measure different perimeters: one closer to formally measured online retail media and the other covering the wider on-site, off-site, and in-store ecosystem. Anyone quoting a single UK retail media figure without saying which perimeter they mean is quoting a number they have not checked.

It is not in question what the direction is. For the purposes of this discussion, the size of the profit pool is such that ad load creep ceases to be a choice that anyone deliberately makes and instead becomes a thing that occurs during the quarterly reviews.

What the sponsored slot actually takes

Position bias is the starting point. In ranked lists, attention and clicks concentrate heavily in the top positions, and that concentration is largely a property of the position rather than of the item occupying it. This is the same effect that governs category page merchandising, where moving a slow seller from position 22 to position 4 changes its sell-through without changing anything about the product.

Apply that to a sponsored placement and the accounting question becomes obvious. The advertiser is not buying incremental attention. They are buying a position that already carried attention, and something else was in it. If the displaced item would have been bought, the retailer has traded a full merchandise margin for an advertising fee, and the size of that trade depends entirely on the relationship between the two items and on whether the shopper simply bought the sponsored one instead.

There is indeed a genuinely harmless situation, and it is reasonable to point this out. In this instance, the product being promoted is one that the shopper would not have come across otherwise, and the purchase in question is only a small addition to what they were already going to buy. Retail media is thus carrying out precisely the function that advertising is meant to perform. The issue is not that this kind of situation doesn't exist; the problem is that no standard retail media report treats it any differently from the case where the advertisement appeared before a purchase had already been made.

The credibility discount at the top of the page

The most useful piece of evidence I have found on this is a field study published in Marketing Science under the title 'Do Sellers Benefit from Sponsored Product Listings? Evidence from an Online Marketplace'. The authors ran a large study on a marketplace app using both experimental and natural variation, holding the product and the position constant and varying only whether the listing was sponsored.

Their result is accurate: when a product appears in the top positions, consumers prefer organic listings to sponsored ones of the same product; but when the product is in lower positions, this preference vanishes, and consumers show no preference between the two. The reason lies in the issue of credibility: an organic listing that reaches the top of the page is given an implied endorsement by the platform, whereas the sponsored label takes that endorsement away. Since there is no endorsement to remove when the listing is lower down the page, the sponsored label has no cost.

The name for this in the wider literature is persuasion knowledge. Once a consumer recognises something as a persuasion attempt, they apply a discount to it, and the discount scales with how much credibility the format was carrying in the first place. That is why the same sponsored label behaves differently at position one and position eleven.

For a retailer running its own network, this inverts the intuitive pricing logic. The top slot commands the highest price precisely because it carries the most attention, but it is also the slot where converting organic to sponsored destroys the most response per impression. You are selling the position where your product loses the most value in the act of being sold. That is not an argument against selling it. It is an argument for pricing it with that cost included, which almost nobody does.

Relevance decay, and how much weight the evidence carries

Beyond the single slot, there is the question of what a rising ad load does to the quality of the whole result set. Independent audits of Amazon's search results have compared the quality of sponsored recommendations against organic ones across markets and reported that, in some geographies, a majority of sponsored results were lower quality than the organic results they sat above, with the split varying considerably by country.

I want to be careful here. That work is an algorithmic audit published as a preprint rather than a peer-reviewed field experiment, the quality measure is constructed by the researchers, and it examines one marketplace. It is suggestive rather than conclusive, and I would not build a business case on it. It is useful for framing the right internal question: whether anyone at your organisation is measuring the relevance of the blended result set at all, rather than measuring the click-through rate of the ads inside it.

There is also a structural finding worth knowing from the marketplace design literature. Work published in Marketing Science on designing an online retail marketplace shows that the bids sellers place in the ad auction reveal private information to the platform, which the platform can then use to improve its organic placements, a practice the authors call strategic listing. It also shows that doing so creates an externality between the sponsored and organic sides which intensifies competition in the auction and reduces sellers' incentive to participate at all. In other words, the interaction between the paid and organic sides is not a leak to be plugged. It is a design variable, and it can be set well or badly.

Why the cost never appears in the numbers

There are three reasons, and they add up.

The first point is attribution. Since on-site retail media is closer to the actual transaction than any other form of advertising, last-click and last-touch models assign it a huge amount of credit. A sponsored impression given to a customer who is already three-quarters of the way through the process of making a purchase will be given the credit for that purchase. The reported return on ad spend appears to be extraordinary and is mostly purely mechanical.

The other point is that no counterfactual is set up. In order to find out how much a sponsored placement cost you, you have to know what would have happened in that position in the absence of the placement, and for that you need a holdout. Few retail media networks have one, mainly since the advertiser is paying for reach and a holdout reduces that reach, and also because the result is commercially inconvenient for the team responsible for the measurement.

The third factor relates to organisation. The retail media team is assessed according to advertising revenue, while the trading team is evaluated on merchandise contribution. If ad load increases and category conversion drops by a small amount, the first team can clearly demonstrate its increase, whereas the second team can only notice a decline that it cannot clearly assign to any particular cause. In any internal discussion, precise figures always win over imprecise ones, no matter which of the two is bigger.

What I would measure instead

The contribution per session, averaged. To calculate it, take the total gross profit from a session. This includes the merchandise contribution and the advertising revenue after deducting the costs of serving, and divide that by the number of sessions on the relevant surface. It is a rough figure, but it is the only one that makes the trade noticeable, since a change in ad load causes the two components to move in opposite directions and this metric cancels out those movements.

Set up an ad load ladder and, keeping all other factors the same, change the number of sponsored placements per surface among the matched cohorts or geographical areas. The blended contribution per session will increase as ad load rises and then start to decline, the point at which this happens being the figure you actually require. In all the versions of this test that I have either seen or heard about, the optimal value lies below the level that the advertising team would have chosen and above the one that the trading team would have selected, which is the normal result when two different functions are each trying to optimise half of an equation.

There are also two supporting measures: the search refinement rate, which is the proportion of sessions in which a shopper carries out a second search after having seen a result set, since an increasing rate is an early indication that relevance has deteriorated before conversion becomes significant enough to be detected; and the exit rate from the first result set when dealing with high intent queries, for the same reason.

The incremental effects on the ad side must be tested rather than predicted; geographic holdouts are the practical method used by most retailers, and the level of discipline required is the same one that eventually resolved the brand bidding issue in paid search: the figures provided by the platform reply to a question different from the one that the business actually wants to answer.

Who should own the ad load decision

Not the retail media team, and not the trading team either. Both are structurally conflicted in opposite directions.

The decision belongs with whoever owns total online contribution, which in most organisations is the ecommerce director or the commercial director above them. What the retail media lead owns is the pricing, the yield mechanics, the advertiser relationships and the measurement standard. What the trading team owns is the relevance floor and the merchandising rules the ad system must operate within. The ad load itself, the number of paid positions per surface and where they sit, is a P&L decision and needs to be made by the person whose P&L it is.

It's a rather dull aspect of governance, and it's the factor that decides whether retail media becomes a sustainable second source of profit or instead turns into a slow tax on the shopping experience which funded it.

Where this goes next

The governance associated with it is not keeping up with the speed at which the interface is changing. When Kroger introduced its AI shopping assistant, it included advertising from the very beginning rather than adding it in later, and OpenAI has been testing product carousels within ChatGPT. Since a conversational interface has no page, no fold, and no visible ranked list, the shopper has much less evidence to determine which of the recommendations was paid for and which was earned.

Which means the credibility discount that currently protects shoppers on a search results page, the thing the Marketing Science study measured, may not operate in the same way. If it does not, the short-term monetisation opportunity is larger and the long-term trust cost is larger with it. That is worth deciding deliberately now, while the surfaces are still being designed, rather than discovering the answer in a conversion report in two years.