James Taylor explores the real issue at the heart of the “AI in retail media” debate: who controls the decision at ranking time?
When a shopper asks an AI assistant for “moisturiser for sensitive skin under £30”, the first shelf they see may no longer be assembled on a retailer’s website. It may be assembled inside an external AI interface.
That changes where commercial control actually lives. A retailer may still own the catalogue, the stock and (potentially) the customer relationship, yet lose influence over the few products that actually reach the shopper. As discovery moves from pages of results to a handful of recommendations, each ranking decision becomes business critical.
Retailers must own the layer that decides which products are eligible, relevant and a commercial imperative. Expose a raw catalogue to an LLM and you have not modernised the storefront. You have turned the catalogue into input for someone else’s ranking system – and potentially someone else’s auction.
That is the real issue underneath the “AI in retail media” debate: who controls the decision at ranking time? If a retailer does not, then their margin is destined to slip completely out of hand.
Stop treating retail media AI as one machine
Much of the overpromising in the retail media market comes from collapsing five distinct functions into a single box marked “AI”:
- Causal measurement determines whether a campaign produced an outcome or merely coincided with it. It is the discipline that separates incrementality from flattering correlation.
- Predictive ML estimates what is likely to happen next: the probability of a click, a conversion or longer-term value. It lives on signals and shapes our digital experience.
- Decisioning combines those signals with relevance controls, inventory, margin, supplier funding, campaign pacing and merchandising rules to determine what is shown now. This is where bottom line monetisation actually happens.
- Generative AI creates or interprets language, imagery and powers conversational interfaces. It can make an experience more useful – a unified brain for example, but it does not govern business profit.
- Agentic orchestration organises multi-step workflows through tools and APIs. An agent may call the systems involved in a decision; whilst agents reason and make decisions, it should not be mistaken for the system that governs the ranking output decision.
These five functions work together, but they do not surrender their distinct roles. Predictive models shape the experience. Generative AI can act as a unified brain across language and interfaces. Agents can reason, choose and act – but the commercially binding ranking output still needs to operate inside retailer-defined relevance controls, funding rules and profit objectives.
That is the distinction that matters. The agent may decide what it wants to do; the retailer must govern what the commerce system is permitted to show, fund and prioritise. It must do this in a way that is indistinguishable from a raw organic relevance, else an agent may elect not to trust or attend that retailer’s endpoint again.
Put a control layer between chat and the catalogue
The Model Context Protocol is useful infrastructure because it gives AI applications a standard way to access external tools and data. But MCP is a protocol, not a decisioning engine.
Raw catalogue exposure to an LLM functions technically at a minimum viable level, but it is the fastest way to give up profit potential.
A retailer-controlled ranking layer manages profit risk through economic decisioning. It can evaluate the shopper’s intent alongside relevance controls, stock, margin, supplier funding, campaign pacing, private-label priorities and merchandising rules—then return the best permissible ranking for that moment.
That becomes more important as the visible shelf collapses. A desktop results page might expose dozens or hundreds of products. An AI shopping agent may consider only a handful. Moving from fourth to third place can determine whether a product enters the consideration set at all.
Ranking is therefore the power layer. Paid influence should sit inside that same ranking logic. Supplier funding, sponsored bids and merchandising priorities need to work together.
The aim is not to make an irrelevant product win an auction. It is to make organic and paid priorities work inside one engine, with commercial influence proportional to relevance and every paid placement clearly identified.
Intent capture expanded
Moving to semantic AI search introduces relevance curves and AI supports materially more ad opportunities for suppliers, and more conversions for every search.
At a pet retail leader, moving from exact-match keyword bidding to transformer-led relevance ranking increased the share of search queries with monetisation coverage by approximately four times. This was a sequential before-and-after comparison rather than a randomised result, so it is strong directional evidence rather than a claim of perfect causality.
At a leading high end beauty retailer, bid-only auctions replaced signal-led AI relevance. Ad impressions increased by approximately nine times, but retail-media click-through rate fell by 87.3% and sitewide sales fell by 55.6% during the switch. The auction produced more exposure while the business produced less value.
The point is not that reach is bad, it is that reach detached from relevance is commercially dumb. Optimising the auction while degrading the shelf is simply a faster route to the wrong answer.
While humans can tolerate and circumnavigate irrelevance, AI agents will not: if they can’t trust a retailer’s end point for quality, they will shop elsewhere.
Optimise for retailer profit, not proxy metrics
Search, recommendations and retail media are converging into one ranking and economic optimisation problem. Treating them as separate brains forces each system to win its own metric: search chases relevance, advertising chases yield, merchandising chases strategic priorities and pricing chases conversion.
The retailer, in reality, experiences the combined financial result. In order to optimise a complex multi-input system, it needs to be measured in one high capacity system.
A decisioning layer should therefore answer a harder question: what should this shopper see next to maximise long-term profit and trust? That means balancing relevance, conversion, margin, ad yield, supplier funding, inventory position and strategic priorities in one governed choice at ranking time.
The objective cannot be an isolated metric. A system told only to maximise sponsored reach may sacrifice relevance and trust. A system told only to maximise conversion may favour products that would have sold anyway or only the most popular brands. A campaign can improve CTR or ROAS while making the retailer’s underlying economics worse.
Profit per session is a more useful north star because it forces the systems to account for the whole commercial outcome, not merely the number each subsystem is paid to improve. Causal measurement must remain independent of the optimisation loop so retailers can distinguish additional value from demand that has merely been claimed by an ad.
Decisioning is also bigger than conventional sponsored bidding. Supplier funding can support dynamic discounts, inventory movement, category growth, private-label strategy and cross-category conquesting. These are not separate customer experiences requiring separate brains. They are commercial inputs into the same question: what gets shown, where and to whom?
The choice retailers are actually making
The practical route remains sequential. First, establish transformer-based search and real-time signals. Vectorise the catalogue, understand natural-language intent and move beyond exact-match keywords. A raw product feed attached to an LLM is not a shortcut around this foundation.
Second, build a unified decisioning layer across search, recommendations and retail media. Combine semantic relevance with inventory, margin, supplier funding, campaign pacing and retailer-defined controls. This is where the retailer decides how commercial priorities may influence – but not destroy – the customer experience.
Third, expose that governed capability through APIs and MCP tools. Every meaningful action should be machine-operable, policy-constrained, auditable and reversible. Agents can then reason and act autonomously without the retailer surrendering the rules, financial contracts or ranking logic under which those actions become real.
Rapidly emerging tech companies like Particular Audience provide an off the shelf solution. Skipping to the agentic surface is seductive because it produces the most impressive demonstration. It is also how a retailer can hand away the economic core of its business while congratulating itself for modernising the interface.
A retailer will not control words produced by a third-party AI platform or auctions operated outside its walls. It can still control the ranked, policy-bound commercial answer supplied by its own stack. The strategic failure is not that another interface exists; it is allowing that interface to bypass retailer intelligence and consume only a raw feed.
Return to the shopper looking for moisturiser under £30. Behind that simple request sits a choice about relevance, stock, margin, supplier funding, retailer priorities and shopper trust. Someone will govern that choice.
Retailers that own or meaningfully influence the ranking layer retain commercial control of their monetisation surface. Those that expose inventory without decisioning become inventory inside someone else’s auction.
The shelf more than ever is algorithmic. Ranking is the real power layer. Retailers must decide whether it remains theirs.
Author
James Taylor is Founder & CEO, Particular Audience
You are in: Home » Retail Media » GUEST COMMENT Retailers are about to find out whether they still own their own shelves
GUEST COMMENT Retailers are about to find out whether they still own their own shelves
Paul Skeldon
James Taylor explores the real issue at the heart of the “AI in retail media” debate: who controls the decision at ranking time?
When a shopper asks an AI assistant for “moisturiser for sensitive skin under £30”, the first shelf they see may no longer be assembled on a retailer’s website. It may be assembled inside an external AI interface.
That changes where commercial control actually lives. A retailer may still own the catalogue, the stock and (potentially) the customer relationship, yet lose influence over the few products that actually reach the shopper. As discovery moves from pages of results to a handful of recommendations, each ranking decision becomes business critical.
Retailers must own the layer that decides which products are eligible, relevant and a commercial imperative. Expose a raw catalogue to an LLM and you have not modernised the storefront. You have turned the catalogue into input for someone else’s ranking system – and potentially someone else’s auction.
That is the real issue underneath the “AI in retail media” debate: who controls the decision at ranking time? If a retailer does not, then their margin is destined to slip completely out of hand.
Stop treating retail media AI as one machine
Much of the overpromising in the retail media market comes from collapsing five distinct functions into a single box marked “AI”:
These five functions work together, but they do not surrender their distinct roles. Predictive models shape the experience. Generative AI can act as a unified brain across language and interfaces. Agents can reason, choose and act – but the commercially binding ranking output still needs to operate inside retailer-defined relevance controls, funding rules and profit objectives.
That is the distinction that matters. The agent may decide what it wants to do; the retailer must govern what the commerce system is permitted to show, fund and prioritise. It must do this in a way that is indistinguishable from a raw organic relevance, else an agent may elect not to trust or attend that retailer’s endpoint again.
Put a control layer between chat and the catalogue
The Model Context Protocol is useful infrastructure because it gives AI applications a standard way to access external tools and data. But MCP is a protocol, not a decisioning engine.
Raw catalogue exposure to an LLM functions technically at a minimum viable level, but it is the fastest way to give up profit potential.
A retailer-controlled ranking layer manages profit risk through economic decisioning. It can evaluate the shopper’s intent alongside relevance controls, stock, margin, supplier funding, campaign pacing, private-label priorities and merchandising rules—then return the best permissible ranking for that moment.
That becomes more important as the visible shelf collapses. A desktop results page might expose dozens or hundreds of products. An AI shopping agent may consider only a handful. Moving from fourth to third place can determine whether a product enters the consideration set at all.
Ranking is therefore the power layer. Paid influence should sit inside that same ranking logic. Supplier funding, sponsored bids and merchandising priorities need to work together.
The aim is not to make an irrelevant product win an auction. It is to make organic and paid priorities work inside one engine, with commercial influence proportional to relevance and every paid placement clearly identified.
Intent capture expanded
Moving to semantic AI search introduces relevance curves and AI supports materially more ad opportunities for suppliers, and more conversions for every search.
At a pet retail leader, moving from exact-match keyword bidding to transformer-led relevance ranking increased the share of search queries with monetisation coverage by approximately four times. This was a sequential before-and-after comparison rather than a randomised result, so it is strong directional evidence rather than a claim of perfect causality.
At a leading high end beauty retailer, bid-only auctions replaced signal-led AI relevance. Ad impressions increased by approximately nine times, but retail-media click-through rate fell by 87.3% and sitewide sales fell by 55.6% during the switch. The auction produced more exposure while the business produced less value.
The point is not that reach is bad, it is that reach detached from relevance is commercially dumb. Optimising the auction while degrading the shelf is simply a faster route to the wrong answer.
While humans can tolerate and circumnavigate irrelevance, AI agents will not: if they can’t trust a retailer’s end point for quality, they will shop elsewhere.
Optimise for retailer profit, not proxy metrics
Search, recommendations and retail media are converging into one ranking and economic optimisation problem. Treating them as separate brains forces each system to win its own metric: search chases relevance, advertising chases yield, merchandising chases strategic priorities and pricing chases conversion.
The retailer, in reality, experiences the combined financial result. In order to optimise a complex multi-input system, it needs to be measured in one high capacity system.
A decisioning layer should therefore answer a harder question: what should this shopper see next to maximise long-term profit and trust? That means balancing relevance, conversion, margin, ad yield, supplier funding, inventory position and strategic priorities in one governed choice at ranking time.
The objective cannot be an isolated metric. A system told only to maximise sponsored reach may sacrifice relevance and trust. A system told only to maximise conversion may favour products that would have sold anyway or only the most popular brands. A campaign can improve CTR or ROAS while making the retailer’s underlying economics worse.
Profit per session is a more useful north star because it forces the systems to account for the whole commercial outcome, not merely the number each subsystem is paid to improve. Causal measurement must remain independent of the optimisation loop so retailers can distinguish additional value from demand that has merely been claimed by an ad.
Decisioning is also bigger than conventional sponsored bidding. Supplier funding can support dynamic discounts, inventory movement, category growth, private-label strategy and cross-category conquesting. These are not separate customer experiences requiring separate brains. They are commercial inputs into the same question: what gets shown, where and to whom?
The choice retailers are actually making
The practical route remains sequential. First, establish transformer-based search and real-time signals. Vectorise the catalogue, understand natural-language intent and move beyond exact-match keywords. A raw product feed attached to an LLM is not a shortcut around this foundation.
Second, build a unified decisioning layer across search, recommendations and retail media. Combine semantic relevance with inventory, margin, supplier funding, campaign pacing and retailer-defined controls. This is where the retailer decides how commercial priorities may influence – but not destroy – the customer experience.
Third, expose that governed capability through APIs and MCP tools. Every meaningful action should be machine-operable, policy-constrained, auditable and reversible. Agents can then reason and act autonomously without the retailer surrendering the rules, financial contracts or ranking logic under which those actions become real.
Rapidly emerging tech companies like Particular Audience provide an off the shelf solution. Skipping to the agentic surface is seductive because it produces the most impressive demonstration. It is also how a retailer can hand away the economic core of its business while congratulating itself for modernising the interface.
A retailer will not control words produced by a third-party AI platform or auctions operated outside its walls. It can still control the ranked, policy-bound commercial answer supplied by its own stack. The strategic failure is not that another interface exists; it is allowing that interface to bypass retailer intelligence and consume only a raw feed.
Return to the shopper looking for moisturiser under £30. Behind that simple request sits a choice about relevance, stock, margin, supplier funding, retailer priorities and shopper trust. Someone will govern that choice.
Retailers that own or meaningfully influence the ranking layer retain commercial control of their monetisation surface. Those that expose inventory without decisioning become inventory inside someone else’s auction.
The shelf more than ever is algorithmic. Ranking is the real power layer. Retailers must decide whether it remains theirs.
Author
James Taylor is Founder & CEO, Particular Audience
Read More
You may also like
Subscribe to our email community