The short version
- US consumer protection rules already require paid ads to be labeled “Ad” or “Sponsored,” and Wen and colleagues (opens in a new tab) argue these rules should be extended to AI answers.
- In an economic model by Zhang and colleagues (opens in a new tab), always showing ads in AI answers ended with the lowest cumulative payoff in all eight market conditions tested across 160 simulation runs.
- In an experiment with 4,927 US adults, Li and Aral (opens in a new tab) showed that small cues, such as links or helpfulness counts, change trust in AI answers.
- Undisclosed commercial interest is already in AI citations: in our self-ranking lists study, 24.2% of cited “best X” lists with an identifiable publisher ranked that publisher first.
- The direct question, whether labeled AI ads beat disguised ones for the brand, has not yet been tested with real users.
Why does labeling matter more in AI answers than in search results?
Because an AI answer blends everything into one voice, so an unlabeled ad looks exactly like advice. On a traditional results page, sponsored links sit in marked slots. Wen and colleagues (opens in a new tab) point out that US Federal Trade Commission rules require paid ads to be labeled “Ad” or “Sponsored,” so people can tell promotion from neutral information.
In an AI answer, they argue, commercial influence can move into the evidence the assistant reads and the reasoning it writes. Persuasion then “operates through the model’s reasoning itself,” and the line between advice and marketing collapses. They call for clear markers when an answer reflects a material commercial connection.
Regulators are paying attention. An audit by Uberti-Bona Marin and colleagues (opens in a new tab) notes that on 31 August 2026 the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act. That brings duties on consumer protection risks and independent audits.
Does a clear label help the advertiser or hurt it?
It costs some short-term persuasion but likely protects long-term trust. Erickson (opens in a new tab) explains the short-term cost. Earlier advertising research shows that people who know they are being persuaded become more skeptical. A disguised ad avoids that skepticism, which is why it can work better in the moment.
The long-term picture is different. Erickson argues that when users do recognize promotion inside an answer, it can reduce trust and worsen their view of the brand. His conclusion: “In the long run, a clear delineation of advertising may benefit companies since it will not breed the same level of distrust that disguised advertisements might.” This is a reasoned argument built on earlier advertising studies, not a test in AI search.
Wen and colleagues add a market-level warning. If hidden promotion outperforms labeled ads, firms that hide their motives win, and the whole channel drifts toward covert tactics. That raises the risk of “trust erosion” for everyone when the influence comes to light.
What does economic modeling say about ad-heavy AI answers?
A model of AI platforms finds that leaning hard on ads wins early revenue but loses users and long-run value. Zhang and colleagues (opens in a new tab) built a model in which an AI engine decides, question by question, whether to show an answer with ads or an ad-free one. Ads earn money now. Ad-free answers build user experience, which supports retention and paid subscriptions later.
Their simulations used 500 users over 20 periods, and a sensitivity check covered eight market conditions in 160 simulation runs. In every condition, always showing ads produced the lowest cumulative payoff of the four policies tested. When users were highly sensitive to ads, the always-ads policy ended at 1.40 in the model’s payoff units, against 9.19 for the best policy. That best policy served ad-free answers to non-subscribers 92% of the time.
The authors add that the damage is hard to undo: switching to ad-free answers later did not recover the gains of serving them from the start. These are simulations with assumed user behavior. They describe the platform’s incentives, but an advertiser’s reach depends on that platform keeping its audience. Ads are already arriving in real assistants, as our guide to ads in AI assistants explains.
How much do small design cues change trust in AI answers?
A lot, which is why the design of a label matters as much as its presence. In the experiment by Li and Aral (opens in a new tab), 4,927 participants matched to the US adult population saw search results presented either as AI answers or as ordinary search. Our guide on whether people trust AI search less covers its overall trust finding.
Several cues moved trust:
| Cue added to the AI answer | Effect on trust in the experiment |
|---|---|
| Reference links | Raised trust, even when the links were wrong or invented |
| “Users who found this helpful” shown between 65% and 95% | Raised trust |
| The same count shown between 5% and 35% | Lowered trust |
| Highlighting how certain the AI was | Lowered trust and willingness to share |
The study did not test ad labels. But it shows that people respond to small signals around an AI answer. That cuts both ways. Wen and colleagues warn that labels can backfire through over-labeling, and they suggest platforms test the presence, wording and placement of disclosures with controlled experiments.
Is undisclosed commercial content already in AI answers?
Yes, though not as paid ads: vendors’ own rankings already appear among AI citations. In our self-ranking lists study, we checked numbered “best X” lists cited by six AI surfaces in the US in September 2026. Of 269 lists with an identifiable publisher, 65 (24.2%) ranked their own publisher first. Publishers that included themselves almost always went first: 92.9% of self-including lists did.
These lists were a small part of all citations, 1.1%. But a reader of such an answer may be reading a vendor’s ranking of itself without being told.
Fabricated claims are a sharper version of the same problem. In a skincare test with three commercial AI models, Chu and Hou (opens in a new tab) found that a made-up clinical citation had the same effect as 0.17 rating points of real product improvement. They class such invented claims as potential false advertising.
What should you do about it?
Choose disclosure by default, both in paid placements and in your own content. Practical steps:
- In any AI ad program, require the platform’s sponsored label to be visible next to your mention, not hidden behind a link.
- Disclose commercial interest in your own comparison content. If you publish a “best X” list that includes your product, say so near the top.
- Never use invented statistics, reviews or endorsements in copy that AI assistants may read and repeat.
- Ask platforms for evidence on how their labels affect user trust, and prefer those that test label wording and placement.
- Track both reach and brand sentiment after AI ad campaigns, since the risk shows up in trust, not clicks.
If you want help building honest, citable visibility in AI answers, see our generative engine optimization service.
What does the research not tell us yet?
No study we reviewed has measured whether labeled AI ads outperform disguised ones for the advertiser. In detail:
- Erickson’s case for clear labels rests on earlier advertising research, not on tests inside AI answers.
- The economic model describes platform incentives with simulated users; its payoff numbers are not money.
- The trust experiment tested links, feedback counts and certainty cues, not ad labels.
- Wen and colleagues note that disclosure can carry a trust penalty and can over-label; the right design is still an open question.
- Our self-ranking lists study observes citations, not whether readers notice or care about the self-ranking.
Frequently asked questions
Do ads in AI chatbots have to be labeled?
US consumer protection rules require paid ads to be labeled “Ad” or “Sponsored,” but how these rules apply inside AI answers is unsettled. Wen and colleagues argue the FTC-style rules should be extended to AI answers explicitly.
Does labeling an ad make it less effective?
Possibly in the moment, because people who know they are being persuaded become more skeptical. Erickson argues that clear labels still pay off in the long run because disguised ads breed distrust.
Are AI platforms better off without ads?
In one economic model, always showing ads gave the lowest cumulative payoff in all eight conditions tested. Mixed policies that mostly served ad-free answers did best, but this is a simulation, not observed behavior.
Is it disguised advertising to publish a list that ranks my own product first?
It is not a paid ad, but readers of an AI answer may not know the list is self-ranked. In our study, 24.2% of cited “best X” lists with an identifiable publisher ranked that publisher first, so disclosing your interest is the safer practice.
Sources
- Erickson, J. (2025), Fake Friends and Sponsored Ads: The Risks of Advertising in Conversational Search (opens in a new tab), arXiv:2506.06447.
- Wen, Y., Zhang, N., Yuan, H., Chen, X., Zhang, H. and Guo, H. (2026), Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots (opens in a new tab), arXiv:2606.12439.
- Zhang, L., Jiao, C., Li, B. and Xiong, C. (2026), An Economic Framework for Generative Engines: Advertising or Subscription? (opens in a new tab), arXiv:2603.29071.
- Li, H. and Aral, S. (2025), Human Trust in AI Search: A Large-Scale Experiment (opens in a new tab), arXiv:2504.06435.
- Uberti-Bona Marin, L. G. and colleagues (2026), "If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations (opens in a new tab), arXiv:2609.18729.
- Chu, X. and Hou, Y. (2026), Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems (opens in a new tab), arXiv:2606.17443.
- Underneath (2026), How many “best of” lists cited by AI rank their own brand first?