What Target Learned Building AI Into Ecommerce

The product detail page was supposed to be settled territory. It was the destination, the place where discovery ended and conversion began, and for most of eCommerce’s history it looked the same no matter who showed up. That assumption is now breaking apart.

At eTail Boston, Brad Thompson, SVP of Technology at Target, walked through how a thirty billion dollar eCommerce business is rebuilding its PDP around a simple but disruptive idea: the page itself should change depending on who is looking at it, and it should change in real time.

Thompson’s team has already turned PDP changes into configuration rather than code, letting merchants test layout tweaks, like swapping in a second recommendation carousel, without an engineering cycle. AI pushes that further. Target can now assemble a custom-configured PDP for each individual shopper at runtime, using context like past purchases, browsing history, and intent. Someone who has engaged with video content in the past might see video prominently placed. Someone else, more inclined toward “complete the look” browsing, sees more of those carousels instead.

The page you see is not the page I see

This is not a hypothetical. Thompson described it as already in motion, built on infrastructure Target has spent years assembling. The shift traces back to the earliest days of ChatGPT, when Thompson says the appeal of large language models was obvious, but the risk of hallucination made broad deployment too dangerous. Target’s early AI use cases stayed narrow on purpose: ratings and reviews synthesis on site, internal business processes off it. Both were chosen for low risk and high learning value, not immediate impact.

That caution eased as the ecosystem matured. Protocols like MCP gave Target a consistent way to expose its own digital services to large language models, and later frameworks like ACP from OpenAI and UCP from Google gave it a path to facilitate transactions agentically. Confidence followed capability. Target became the first major retailer to build an app inside ChatGPT, a move Thompson frames not as a revenue play but as a chance to co-develop with a frontier model provider and teach its own engineers how to deploy MCP at scale.

Shoppers are describing problems, not typing keywords

The underlying reason for all of this is a change in how people search. Thompson pointed to a shift away from keyword queries, “organic snacks,” toward constraint-heavy problem statements: a parent needing lunches for three kids at swim camp, under twenty dollars a day, with an allergy to account for. Solving that requires a different kind of technology than a search box ever did, one that can hold multiple constraints at once and act on them.

For Target, that shift is also a data opportunity. Thompson described how a single prompt like that can turn into an institutional memory about a household, its dietary restrictions, its interests, even something as specific as competitive swimming, that a personalization engine can act on months later. A nudge about swim gear when it goes on sale becomes possible because of a conversation that happened somewhere else entirely.

Off-platform traffic still has to land somewhere trustworthy

Much of that shopping behavior is now happening off Target’s own properties, inside ChatGPT, Gemini, or Claude. Thompson noted that OpenAI has already pulled back from letting transactions complete inside the chat interface itself, instead directing users back to a retailer’s own site to close the purchase. That makes the PDP the landing point for intent formed somewhere else, which raises the stakes on getting the product feed right. If the data going to the frontier model is not high fidelity, the page a shopper lands on will not match what they were actually looking for.

Onsite AI is a different cost calculation entirely. Thompson was direct about the economics: an agentic shopping experience at Target’s scale means billions of turns with a large language model, and every one of those turns costs money and adds latency. Search has trained shoppers to expect sub-second responses. A conversational assistant cannot deliver that, and Target has had to get comfortable asking guests to accept a slower, richer interaction in exchange for a better one.

Nothing ships until a model proves it did not get worse

To manage that tradeoff, Target built its own internal test harness. Before any change to a prompt or underlying system goes to production, it runs against roughly five hundred test prompts, with a separate large language model scoring the response quality against the production model’s output, while also tracking latency and token consumption. If quality drops, latency spikes, or token burn climbs past an acceptable threshold, the build fails and does not ship. It is a discipline borrowed from software engineering, applied to a system where the components do not behave with the same predictability as traditional code.

What this means for senior marketers

Thompson’s closing advice: fix your product data, and push your teams to deploy MCP everywhere. Neither recommendation involves a new model or a flashy feature. Both are prerequisites everything else depends on.

The retailers still treating the PDP as a fixed asset are optimizing a page that is quietly becoming obsolete. The ones paying attention are already treating personalization not as a layout choice, but as a runtime decision made fresh for every visitor.