How it works
An agent that searches, compares and recommends sits where the shopper’s own judgement used to be. So does an assistant that is only asked ‘which should I get?’. If whoever runs it also sells, is paid by some of the sellers, or takes a fee on what is bought, the system has two principals, and the person can see only one of them.
Steering needs no lies. The order of options, which one is described first, which is called the best fit, and what is left out are enough. Ranking bias on a page of search results took a decade of competition law to address. In a conversation it is quieter still: the person often sees one recommendation, not a page of ten, and has nothing to compare it with. Advertising is arriving in the same voice. In an experiment with 179 people, participants ‘struggled to detect ads’ placed inside a chatbot’s answers ‘and even preferred LLM responses with hidden advertisements’.
Field marks
- The operator sells, or is paid by sellers, in the categories it recommends.
- Its picks favour the operator’s own products or paying partners more often than price or quality would explain.
- It gives one answer where a person searching would have seen several.
- Commercial relationships are disclosed in the terms, or in a label somewhere else on the screen, not beside the recommendation.
Where it shows up
The precedent is search. In September 2024 the Court of Justice of the EU upheld a €2.4 billion fine on Google for favouring its own comparison-shopping service in its results. The Digital Markets Act now says that gatekeepers ‘shall not treat more favourably, in ranking’ their own services than a third party’s, and in July 2026 the Commission fined Google €460 million under that article for preferring its own shopping, hotel, transport and sports results, noting Google’s proposals to apply the same principles to AI Overviews and AI Mode. Platforms are also deciding which agents may act on them at all: in June 2026 the Commission ordered Meta to restore rival AI assistants’ access to WhatsApp, and in March 2026 a US court granted Amazon an injunction against Perplexity’s shopping agent, so that on Amazon the only agent that shops is Amazon’s.
Inside assistants, the commercial layer is being built now. ChatGPT began showing ads to free-tier users in January 2026 with the assurance that ‘Ads do not influence the answers ChatGPT gives you’. Its Instant Checkout ranks products ‘purely on relevance’, merchants pay a fee on completed purchases, and where several merchants sell the same product, whether a merchant has enabled checkout is one of the factors that decides which the person sees. Perplexity sells sponsored follow-up questions; Google places ads inside AI Overviews and beneath AI Mode. Each of these is disclosed. None is disclosed beside the recommendation, in the sentence where the choice is made.
Regulators have started to say what they expect. CERRE warns that ‘an agent operated by a merchant may have an inherent incentive to steer purchasing decisions to its own products’. The UK Competition and Markets Authority says people ‘may be steered towards products and services that are more profitable but less suited to their needs’, that ‘paid endorsements should be properly labelled’ and that ‘how results are ranked and any links with suppliers should be transparently disclosed’. Its guidance to businesses is blunter: they are ‘responsible for what an AI agent does in the same way you are responsible for what an employee does’, even when someone else built it.
Designed out
- Disclose interest at the point of choice. Say beside the option when it is the operator’s own, pays for placement, or earns the operator a fee.
- Show what was left out. Let the person see the alternatives the agent considered and why it ranked them lower.
- Rank by the person’s criteria, and check. Write the ranking rules down, keep commercial signals out of them, and audit outcomes for self-preference.