# Research shows AI agents reject sponsored listings 97% of the time. Nobody is building the advertising infrastructure for what comes next. — retained record

> Retained reddit record from reddit:artificial.

Canonical: https://highsignal.app/data/records/489e093f3e066f2e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:artificial)  published 2026-10-20 00:26 ingested 2026-10-21 02:36

[original source ↗](https://www.reddit.com/r/artificial/comments/1xadsk6/research_shows_ai_agents_reject_sponsored/)

## retained text



```
Two recent papers caught my attention and I think they point to a massive blind spot in how we're thinking about AI agents in commerce.

Wadi & Ma (McGill, Sep 2026) tested what happens when you tell an LLM "you work for the traveler" vs "you work for the booking platform." The consumer-deployed agent chose sponsored listings only 2.8% of the time — a 97.2% rejection rate. The platform-deployed agent? 25%. Nearly 9x higher. This replicated across GPT, Gemini, and Claude.

Qiu & Mei (Michigan, May 2026) mapped out four tiers of advertising influence in generative AI and cataloged the commerce protocols being built — Google's UCP, OpenAI's ACP, Microsoft's Copilot Checkout. But here's what struck me: all of these handle transactions. None of them handle demand generation — how does a brand even get into an agent's consideration set?

Both papers study platform-deployed agents (Booking.com's assistant, Expedia's agent). But the market is heading toward user-deployed agents that work exclusively for you, with no platform allegiance. When that happens:

* Traditional ads are invisible to agents (designed for human psychology, not machine parsing)
* Independent agents over-filter commercial info — even a role-neutral agent penalizes "sponsored" content by 42 percentage points just from training data bias
* The protocols being built to replace this (UCP, ACP) are platform-owned, not neutral

So we have a paradox: the agent that's best for the user (conflict-free, fiduciary) is actually worse in practice because there's no open infrastructure for commercial discovery.

Has anyone seen work on this? Specifically on how brands would present structured, machine-evaluable product data to independent agents — not persuasion, but information provision.

I wrote a longer synthesis [here](https://peaceinmind2.substack.com/p/the-680b-blind-spot-when-your-customer?r=vlytn).

**Papers referenced:**

* Wadi & Ma, "Whom Do AI Agents Serve?" — [https://arxiv.org/abs/2609.17989](https://arxiv.org/abs/2609.17989)
* Qiu & Mei, "Generative AI Advertising as a Problem of Trustworthy Commercial Intervention" — [https://arxiv.org/abs/2605.18673](https://arxiv.org/abs/2605.18673)
```



document fetched 2026-10-21 00:19 · canonical https://reddit.com/r/artificial/comments/1xadsk6/research_shows_ai_agents_reject_sponsored
