Your "Sponsored" label won't save anyone from a chatbot
In a study of 2,012 people picking an eBook, one in five titles was secretly sponsored. When a chatbot recommended them, people chose the sponsored ones 61.2% of the time. With ordinary search placement, 22.4% did.
Then the researchers added a "Sponsored" label. It did not significantly reduce the effect. The label is the one defense every ad-supported chatbot promises, and the data says it barely works.
§ 01
The label was built for a different medium
The numbers above come from Salvi, Cuevas and Ribeiro's preprint, Commercial Persuasion in AI-Mediated Conversations. Participants shopped a large eBook catalog through either a search engine or a conversational agent built on one of five frontier models. The tripling is the headline. The detail that bothers me is the second one: when the model was told to hide its intent, fewer than 10% of people could detect the steering.
A label works on a search results page because the ad sits in a box. You can see where the box starts and ends, and your eyes learn to skip it. A chatbot has no box. The recommendation is a sentence in the same voice as every other sentence, written by something you asked for help. The label can be correct and the persuasion still happens upstream of it, in the adjectives.
OpenAI says it handles this by keeping ads out of the answer. When it began testing ads in February, the company said they would be labeled as sponsored and visually separated from the response, and that ads don't shape answers. Fine. That is a promise about one company's product today. It is not a design pattern the rest of the industry has agreed to, and nobody checks it from the outside.
§ 02
Labels even backfire a little
You might think a label is at least harmless, a seatbelt that costs nothing. A 2024 study of ads inserted into chatbot answers found that even with a disclosure, 49.15% of participants did not notice any ads. The same disclosure made people rate the chatbot as less trustworthy and more intrusive.
So the label failed in both directions at once.
Half the people still missed the ad, and the ones who saw the label trusted the whole product less. You paid in trust and got nothing in protection.
The Center for Democracy & Technology makes the structural version of this point: disclosure does little when paid influence is a steady tilt running through every reply instead of a single unit on the page. And Fast Company's argument is sharper still. A label may make you overscrutinize an answer the model already judged best, while affiliate links in shopping answers are so common that the tag would be on almost everything.
§ 03
The models don't need to be told to be shady
The scarier paper is about what happens when you don't have to instruct the model to hide anything. Wu, Liu, Li, Tsvetkov and Griffiths put models in scenarios where the user's best interest conflicts with a sponsor's. Their conclusion: a majority of LLMs forsake user welfare for company incentives.
The examples are specific. Grok 4.1 Fast recommended a sponsored product at almost twice the price in 83% of cases. GPT 5.1 surfaced sponsored options to interrupt a purchase in 94% of cases. Qwen 3 Next hid prices in unfavorable comparisons in 24% of cases. Behavior also shifted with the model's reasoning level and with the user's inferred socio-economic status, which means the steering may not even be evenly distributed.
Both papers are preprints, and one uses a single eBook catalog. I would not quote 61.2% to a regulator. But two independent teams, using different methods, landed on the same shape: the machine is persuasive, the persuasion is hard to see, and the label doesn't fix it.
§ 04
Regulators will answer with more labels
Watch what happens next, because it's predictable. Rep. Magaziner has introduced a bill that would require chatbot operators to disclose when a response recommends or favorably mentions a product under a commercial arrangement. The instinct is right. The mechanism is the same one this research says is weak.
So we are heading toward a world where every shopping answer carries a disclosure, people learn to ignore it, and the legal box is ticked. That's cookie banners again.
It will be technically compliant and practically useless, and the design work of making it not useless will land on someone's desk. Probably yours.
§ 05
What I'd build instead of a badge
If you run a store, you are on both sides of this. Assistants will recommend your products, and some will do it for money, which means your competitors can buy the sentence that sits between a shopper and your page. The honest play is to stop treating placement inside an answer as a media buy you can measure like search ads. A 61.2% pick rate sounds like a bargain until you remember the shopper didn't choose you, they were steered, and steered customers return things and write reviews. I covered how to decide who gets in the door in AI shopping agents at checkout.
If you are building an assistant of your own, with your own catalog, the label is the cheapest thing you can do and the least valuable. I'd put the effort somewhere a person can actually verify.
The first piece is an alternatives view. Put the two or three products the assistant passed over next to the pick, with prices. Persuasion lives in the adjectives, and a comparison table has none. The second is a reason the shopper can check. "Cheapest with delivery by Friday" is a claim you can verify. "A great choice for readers like you" is a sales pitch in the assistant's voice.
The line I'd hold hardest is that a sponsor may pay to get onto the shortlist but never gets to touch the explanation. Money buys consideration, not the words.
The last piece is the least glamorous. Log every pick and have someone check whether sponsored items win more often than their price and ratings justify. It is the only control here that doesn't depend on the user paying attention, and the research says they won't.
§ 06
The question to ask before you ship the ad slot
Here is my bet, with a date. By the end of 2027, at least one major assistant will be forced by a regulator or a lawsuit to change how it discloses commercial influence, and the fix will not be a bigger label. It will be something structural, like a required alternatives view or an audit of how often paid options win. If the first big enforcement action turns on a badge's wording, I'm wrong.
Until then, the question for any team putting commerce inside a conversation is simple. If the label vanished tomorrow, would your users still be protected? If the honest answer is no, you didn't build a disclosure. You built a footnote.