The Chatbot Closed the Sale. The Publisher Created the Demand. That's a Problem.

Criteo bought its way into ChatGPT. Advertisers can now place sponsored products inside artificial intelligence responses, the same way they buy display inventory across the open web. It's a logical move. It's also the clearest signal yet that the ad industry is about to repeat its worst mistake at scale.

Picture this scenario: a user asks ChatGPT for a trail running shoe recommendation. The model synthesizes dozens of sources, generates a response, and serves a Criteo-powered placement. The user taps it and buys the shoe.

Criteo takes its margin, the brand gets its sale, and OpenAI captures the ad revenue.

The publisher whose expert review shaped that user's preferences, established their buying criteria, and built the brand trust that made the click possible? It gets nothing. No attribution or revenue share. There’s zero sign that it contributed to that transaction at all.

If that sounds familiar, it should. This is the original sin of digital advertising, now supercharged and baked into the infrastructure of AI.

Attribution Was Always a Useful Fiction

The industry has been trying to answer "Who deserves credit for a sale?" for 20 years. The honest answer is we don't know. But we have plenty of models that let us act like we do.

First-touch attribution handed credit to the first click, structurally enriching Meta and the walled gardens. Last-touch attribution handed it to the final click, enriching Google. Linear models split it evenly and satisfied no one. Data-driven MTA tried to weight each touchpoint by statistical inference across thousands of journeys, and worked reasonably well in a world where most of the path was at least partially visible.

The operative word is "visible." Different devices, consent opt-outs, cross-browser gaps — all of it created holes that attribution models papered over with assumptions. The industry spent hundreds of millions building MTA vendors, incrementality frameworks, identity graphs, and clean rooms to manage those gaps.

That entire apparatus assumed a world where touchpoints existed and were hard to measure. The chatbot era has produced a world where touchpoints don't appear in the measurement layer at all.

This Isn't an Attribution Gap. It's an Attribution Absence

When a user gets a product recommendation inside ChatGPT, there's no cookie or session ID or pixel on the content sources the model drew from because the user never visited those sources. The publisher's review didn't appear as a URL the user clicked. It appeared, transformed and synthesized, inside the model's output, with no thread connecting it to the conversion event downstream.

The old attribution problem was a measurement problem: tracking was imperfect, but the touchpoints existed. Statistical models filled the gaps with reasonable accuracy for most use cases.

The new problem is structural: the interface itself is designed in a way that makes content contributions invisible. You cannot retrofit a pixel onto a training weight. You cannot add a session ID to a synthesized response. The architecture doesn't permit it.

Criteo deserves credit for moving fast. However, what it has built is a mechanism for monetizing the transaction layer of a journey that the content industry built, with no pathway for that content industry to participate in the outcome. That's not just unfair to publishers. It's unstable for the ecosystem and ultimately for the brands investing in it.

Licensing Deals Won't Solve This

There's an emerging framework that publishers and AI companies are calling "pay per demonstrated value." Rather than flat licensing fees for training access, publishers get paid in proportion to how often their content is cited in AI responses. High-quality specialist content earns more than commodity content the model rarely draws on.

The logic is sound and the direction is right. However, this framework solves a different problem. Pay-per-demonstrated-value is a licensing negotiation. It compensates publishers for their contribution to training. It doesn't create a real-time signal connecting a specific piece of content to a specific purchase.

The publisher whose trail running guide influenced a Brooks shoe sale gets a share of a licensing pool, not a performance signal saying this content drove this conversion, at this margin, from this reader. Those are fundamentally different things.

Licensing fees are negotiable, and they tend to resolve in favor of whichever party has more leverage. Performance signals are mathematical and auditable. One is a negotiation that the smaller party usually loses. The other is infrastructure that doesn't require anyone's goodwill to function.

The industry keeps reaching for the negotiation when what it actually needs is the infrastructure.

The Intent Didn't Form in the Chatbot

Here's the thing about the moment a user types "best trail running shoe for marathon training" into ChatGPT: by that point, most of the decision is already made.

The intent didn't form in the model. It formed in the months before the query, in the content the user read, the writers they trusted, the reviews that shaped their criteria. The chatbot is confirming a preference and facilitating a transaction. The education happened somewhere else, in content, long before the query was typed.

That means the most strategically important commerce moment is not inside the chatbot. It's upstream of it, inside the content experience that formed the intent in the first place.

A publisher who knows their reader has spent three months reading trail running content knows something ChatGPT doesn't know at query time: the full journey. It has the relationship, the context, and the trust. If the publisher can capture the commerce moment inside that content experience, surfacing the right product at the right point in the reader's journey with a conversion path that lives in its own environment, it owns the attribution entirely. No statistical model. No licensing negotiation. A direct, clean, mathematical connection between content engagement and purchase outcome.

That's not a gap in the market. That is the market. And most publishers haven't built for it yet.

The Play is Upstream, Not Inside the Chatbot

Criteo getting into ChatGPT first is a real win for the company. It will capture meaningful margin on conversions at the transaction layer, and that revenue is real.

But the brands and publishers building for the long term should be asking a harder question: If the transaction layer is increasingly owned by AI interfaces, where does demand actually get created? Where does brand preference form? Where does consideration happen?

The answer, as it has always been, is in content. The publishers and platforms that invest in genuine reader relationships, understand behavior over time, and build commerce into the content experience rather than bolting it on afterward will have something no attribution model can replicate: provable causality between content and conversion.

The chatbot era didn't invent this problem. It just made the stakes obvious. Every conversion that happens inside an AI interface is a conversion the content industry influenced and didn't get paid for. That gap will keep widening until publishers stop waiting for the attribution problem to be solved upstream and start owning the commerce moment themselves.

The infrastructure to do this exists. The business case has never been clearer. The only question is whether the open web will move fast enough to claim it.

Peter Wilmot is the chief product officer of Shopsense, an AI-powered content-to-commerce platform.