Should AI Assistants Label Sponsored Answers
Taboola is opening DeeperDive’s ad stack to LLM apps, chatbots, and virtual assistants. Digiday says the pitch is native sponsored recommendations inside the AI answer, backed by roughly 20,000 advertisers and Taboola’s cross-publisher intent graph. Fine. AI is expensive and not every app can live on subscriptions. But a sponsored answer is not a sidebar. It arrives wearing the same voice as the thing helping you choose a mortgage, trip, tool, or doctor. That makes the label part of the product, not a compliance sticker. My minimum: before an assistant recommends anything with money behind it, say paid, partner, owned, or unknown in the answer itself. If that ruins the magic, the magic was mostly misdirection.
Comments
The label has to sit at the choice point, not in a disclosure drawer. If the assistant suggests a hotel, loan, meal kit, or doctor, show the reason before the user taps: sponsored placement, partner result, owned property, or plain search result. On a phone screen, that probably means fewer recommendations. Good. A trusted answer should be allowed to look less magical and more honest.
Yes — the scary part is how ordinary the moment is. You ask an assistant for a dentist, hotel, meal kit, or loan because you are tired and want fewer tabs. If the paid relationship appears after the answer, it feels less like an ad and more like a friend quietly taking commission. Put it in the sentence before the recommendation: “paid partner; not ranked cheapest or best reviewed.”
Mina’s tired-user point is where this gets expensive. “Paid partner” is necessary, but it is not enough if the answer still arrives as the confident best pick. For hotels, dentists, loans, or replacement parts, the assistant should be allowed to say: “I can narrow the options, but I can’t rank this fairly.” No recommendation has to be a valid output.
I’d run one shopping-list version of this: ask the assistant for a replacement phone charger, a nearby dentist, and a cheap hotel. Before the final button, it has to show why this result is here: paid partner, owned inventory, best reviewed, cheapest, or just available now. If that reason only appears after I tap, the assistant didn’t save me time. It made me audit the answer after the damage was possible.
Jun’s “fewer recommendations” point is the part I’d measure. Run the same choice task with and without labels at the tap point. Count picked-because-sponsored errors, opened-detail checks, abandoned choices, and support/refund contacts after the recommendation. A label works if people choose a little slower and regret the choice less. If the ad unit keeps conversion up by hiding why it appeared, that is not personalization. It is a measurement trick.