When AI starts selling, what happens to trust?

Google's latest personalised-ad test raises a question beyond whether advertising works.

The trust worth preserving is confidence that they understand whose interests the recommendation serves.

Author: Ian Ash

Published: October 17, 2026

Category: Analysis

An AI assistant can help someone choose a product while leaving them unclear about whose interests shaped the choice. Google's latest advertising test makes that tension more immediate: personal context used to answer a question may also help select an ad. The evidence does not establish that AI advertising inevitably destroys trust. It does show why an accurate answer, a visible label and a successful placement cannot be treated as interchangeable proof that users understand the commercial relationship.

Google's new test adds personal context to advertising

On October 16, Search Engine Land reported that Google was beginning to test ads in AI Mode for users who connect their apps through Personal Intelligence. The report describes a change from that experience's initially ad-free approach. It does not establish universal availability, a rollout denominator or new advertiser targeting controls. [1]

Google's help documentation draws a distinction that matters. Ads will not be based directly on a user's Gmail inbox or Google Photos library. However, summaries, excerpts and inferences from connected sources may be used both to personalise the experience and to show ads. Google says advertisers will not have access to the inbox or Photos, excludes sensitive categories from personalised advertising, and lets users choose whether to connect apps and disconnect them later. [2]

That is not a claim that advertisers can inspect someone's email. It is a documented expansion of the possible use of derived personal context. The strategic question is whether a user understands that helping the assistant answer and helping it select advertising can be two uses of the same information.

Platforms are making different promises

OpenAI says ChatGPT ads operate on separate systems from its chat model, are clearly labelled, and cannot be used by advertisers to shape, rank or alter responses. Its current FAQ describes a gradual rollout to eligible Free and Go users in select regions, rather than advertising for every account. It says Plus, Pro, Business, Enterprise and Edu accounts will not have ads, and accounts identified as belonging to under-18s are excluded. [3]

Anthropic takes a different position. Its February 4 announcement says Claude will remain ad-free, without sponsored links alongside conversations, advertiser-influenced responses or unrequested product placements. The company argues that even separately displayed ads could compromise the clear space to think and work it wants to provide. That is Anthropic's rationale, not a measured comparative trust advantage. [4]

These are company commitments and product positions, not independent audits of user response. Maintaining answer independence and making that independence understandable are separate tests. A platform could preserve the boundary it promises while users remain unsure where the boundary sits.

Trust concerns and commercial usefulness can coexist

Clutch's October 13 report describes a September survey of 747 people, including 628 who considered themselves regular AI-assistant users. It reports that 75% of users agreed ads in an assistant's answers would make them trust those answers less, while 45% agreed ads would make it easier to find and buy what they wanted. These are different questions, not mutually exclusive attitudes. [5]

The figures describe anticipated reactions and perceived convenience, not experimentally observed trust loss or purchases. The public report does not disclose country, recruitment, weighting, full questionnaire or separate item-level counts. It cannot support a claim that three-quarters of all consumers will abandon AI when they see advertising.

Equativ's September 21 article reports an August survey of 1,000 US and Canadian consumers, with 58% open to seeing sponsored product recommendations in AI. The advertising-technology company's public materials do not provide full sampling methods, country allocations or item-level counts. That finding measures stated openness, not actual buying. Clutch and Equativ should not be averaged into a market forecast: their samples, questions and disclosed methods differ. [6]

A useful chatbot is not necessarily an understood chatbot

The peer-reviewed December 2025 study Ads that Talk Back retained 179 US-based Prolific participants after exclusions. Its six conditions crossed GPT-4o and GPT-3.5 with no advertising, embedded personalised advertising, or embedded advertising with disclosure. It tested a researcher-built system, not a production ChatGPT or Google advertising product. [7]

The authors reported no significant ad-versus-control differences on aggregate participant sentiment and several perceived-quality scales. That is not an objective finding that answer correctness was unchanged, nor a universal finding of no harm. Qualitative responses included trust concerns even when participants still regarded the chatbot as helpful. Human perceptions and the paper's separate model benchmarks must not be conflated.

Its disclosure figures also deserve care. The results section says 31 of 60 disclosure participants noticed the cue and four clicked it, but the published condition counts add to 58. That internal discrepancy prevents treating those figures as precise, settled percentages. The broader design question remains useful: providing a disclosure is not the same as establishing that people noticed, understood or acted on it.

Persuasion is not the same as a purchase

The April 2026 preprint Commercial Persuasion in AI-Mediated Conversations, by Francesco Salvi, Alejandro Cuevas and Manoel Horta Ribeiro, reports two preregistered experiments with 2,012 final participants. In the first experiment, an agent instructed to promote sponsors produced an adjusted sponsored-book selection estimate of 61.2%, compared with 22.4% under search placement. Those arms contained 404 and 402 participants respectively; the raw counts were 247 and 90 selections. [8]

The endpoint was choosing an eBook in a controlled task, not an observed retail purchase. The authors paid a bonus equivalent to the book price because the books could not be supplied as planned. The paper's separate keep-book-versus-$1-cash proxy did not produce significant corrected pairwise condition differences. The result establishes a task-specific selection effect, not a live advertising conversion rate.

An additional arm combined an advance warning with a visible sponsorship label. It significantly improved the study's composite bias-detection measure, while its reduction in sponsored selection was not statistically significant after adjustment. That is not proof that labels alone never work. It is evidence that recognising a commercial arrangement and changing a choice are different outcomes, in a single product domain and short experimental setting.

Disclosure explains a conflict; it does not remove it

A 2018 Management Science article compared a disclosed, sponsor-biased online recommendation agent with a neutral agent. Its accessible abstract reports lower trust and higher distrust for the disclosed biased agent. Within biased-agent conditions, trust improved only when both sponsorship disclosure and explanations for organic recommendations were supplied; the interventions did not significantly lower distrust. The study predates generative AI, and the accessible records do not provide its participant counts or effect sizes. It offers an analogy, not a quantified estimate for today's assistants. [9]

Counterevidence also matters. A Nuremberg Institute for Market Decisions research-magazine summary describes a tablet-shopping experiment with 555 participants from Germany, Austria and Switzerland. It varied interface interactivity and sponsorship labelling. NIM reports no direct disclosure effect on recommendation adherence, while the negative effect on perceived integrity depended on interactivity and disappeared in its high-interactivity condition. The summary does not publish cell sizes, effect estimates or inferential statistics. [10]

Together, these sources reject a universal formula such as "advertising reduces trust by X". They involve different tasks, systems, populations and interventions. Nor is lower trust necessarily a design failure: recognising that a recommendation is commercially biased may correct excessive confidence rather than damage a previously well-calibrated judgment.

Personalisation adds a question about permission

Consider a hypothetical travel request. An assistant uses a booking confirmation to suggest restaurants near a hotel, then selects a relevant sponsored placement using the inferred destination. The recommendation and ad could both be useful. A user could still be surprised if they thought connecting the information served only the original request, rather than advertising too.

No study reviewed here measures the long-term trust effect of Google's October test or this particular data-use scenario. The possibility that relevance improves while purpose expectations are disappointed is a hypothesis to investigate, not a demonstrated consequence. Platforms should test comprehension of data use, not assume an opt-in click proves understanding of every downstream purpose.

Why the issue reaches beyond paid visibility

For AEO practitioners, the potential spillover is uncertainty about recommendations that were not paid for. If people cannot distinguish sponsorship from an assistant's independent judgment, brands earning organic representation may face a credibility problem as well. That is a plausible risk, not an observed outcome of Google's announcement. It could also be limited by clear separation and credible explanations.

The distinction builds on AEO Updates' analysis of paid and organic AI visibility and its separation of source exposure, brand representation, recommendation and choice. Buying a placement should not be reported as earning an independent recommendation. Equally, a user can distrust an assistant's commercial incentives while still finding a particular brand or offer appealing.

Marketers should ask platforms what is paid, where it appears, what information selects it, and what advertisers can influence. Campaign reporting should keep paid exposure, organic representation, brand evaluation, confidence in the answer and trust in the platform separate. Clicks and conversions matter commercially, but neither proves that the commercial relationship was understood.

What evidence would settle the argument?

A useful next study would compare clearly defined experiences: an unsponsored answer, an answer with a separate labelled ad, an embedded sponsored recommendation, and a personal-context-based ad with an explicit explanation. Labels, advance warnings and data-use explanations should be tested separately where feasible, with answer quality and commercial relevance controlled rather than allowed to explain every difference.

Measurement should distinguish ad recognition, understood independence, data-use comprehension, answer confidence, brand evaluation and actual choice. Follow-up should test whether reactions persist, whether users change assistants or verification habits, and whether any apparent benefit survives a fully explained commercial arrangement. Higher-stakes decisions such as mortgages, insurance or investment-provider selection require separate research; an eBook or tablet task cannot supply their response estimate.

AEO Updates Takeaway

AI advertising could succeed commercially before platforms establish that people understand how it works. The evidence supports a research and governance question, not an inevitable trust-collapse story: can users recognise what is sponsored, understand how personal information was used, and judge an answer's independence without losing the usefulness of the experience? The trust worth preserving is confidence that they understand whose interests the recommendation serves.