Walmart put 200,000 products inside ChatGPT's checkout and watched conversion drop by two-thirds.

Walmart put 200,000 products inside ChatGPT's checkout and watched conversion drop by two-thirds. It didn't pull out. It rebuilt, keeping the AI app as a place to be found and routing the actual sale back to its own site, where the same shoppers now spend roughly a third more per order than everyone else. At this point the working thesis is that it can serve as a genuine distribution channel while you keep the transaction.

This is the third in a short series on the different types of agents shaping how, and whether, brands get chosen. The first two looked at the model owners themselves, ChatGPT, Gemini, Claude, Perplexity, and the closed marketplaces, Amazon, Shopee, Lazada. This one is about outside companies building their own agent experiences inside somebody else's AI, using the app-building tools OpenAI opened up to developers in late 2025.

Walmart joined as one of the first major retailers on in-chat checkout at launch. It disclosed in March that conversion inside ChatGPT ran three times worse than Walmart.com. Rather than treat that as a failure, it changed the model entirely: Sparky, Walmart's own assistant, now sits inside both ChatGPT and Gemini, but login, cart and checkout all route back through Walmart's own systems. Walmart's US CEO has said Sparky shoppers spend around 35% more per order than other customers, and Walmart Connect, its advertising arm, grew 46% last year partly on that behaviour. For a retailer this size, the AI app isn't a sales channel to optimise for conversion. It's a new front door that happens to bring in higher-value customers, built deliberately alongside the store and the app rather than instead of them.

L'Oréal, announced in June, isn't chasing checkout at all yet. It's doing three narrower things: bringing Maybelline's virtual try-on into ChatGPT, a feature no generic feed could replicate because it depends on L'Oréal's own ModiFace technology; feeding OpenAI verified, structured product data for brands like Lancôme and Kérastah so the model answers from L'Oréal's own information rather than scraped reviews and forum posts; and running paid ad pilots with CeraVe, SkinCeuticals and Garnier at the moment someone's actually asking a shopping question. No results disclosed yet, L'Oréal has called it early days.

Neither is a template to copy outright. Both are among the best-resourced companies in their categories, with assets, an owned ad business, a proprietary AR technology, that most brands don't have, and both bets are only months old. But they point the same direction: the commercial case for building here isn't a better checkout. It's controlling what the model says about you, claiming attention and advertising space before it gets expensive, and treating the app as a new distribution channel you shape on your own terms, with the transaction still landing on ground you control.