Viewpoint

What AI tells shoppers about your product, and who answers

Written by Nicola Colombo | Nov 18, 2026, 1:00:00 AM

The label was our last word to the shopper. It is not any more

For fifteen years the label was the one place where we spoke to the shopper without an intermediary. That is over. Ask an AI assistant about a cereal, a yogurt or a protein bar and it answers in a sentence, before the shopper has seen the pack. It builds that sentence from everything it can find: the pack, the brand site, the retailer page, a marketplace listing, a blog from 2019. The brand wrote one of those sources.

Regulators have noticed. On 11 November 2026 the European Commission proposed the Digital Fairness Act, with a full article on the automated systems that conclude contracts on a consumer's behalf. On 30 October 2026 four UK regulators said in one statement that existing law already applies to AI agents, and that the agent's autonomy does not reduce the responsibility of the organisation that deploys it.

So the question for every regulatory, quality and marketing lead has changed. It used to be "is the label compliant". It is now "what does AI tell my consumers about my product, and who answers for it". This article is about the second half of that question, because the first half has an answer nobody likes: the assistant is already talking, and most of what it says did not come from us.

How the assistant builds its answer, and where the facts drift

An assistant does not read the label. It reads text. When a shopper asks about a product, the assistant gathers whatever text mentions it: the ingredient list a retailer typed in, the claims on the brand page, the reviews on a marketplace, a nutrition table copied by a third party two reformulations ago. It then writes one confident paragraph. The pack, the surface we spend months approving, is one source among five, and often the one the assistant cannot read at all.

The problem is that those sources already disagree with each other. In September 2026 we collected 318 claims made about food products in five markets and compared what the pack, the brand site, the retailer pages and the marketplaces said about the same product. Of the 318 claims, 160, or 50%, disagreed with another surface. A sugar claim worded one way on the pack and another way on the retailer page. A recycled-content figure that changed when the bottle did, but only on the pack. An allergen statement that the retailer never updated. The assistant has no way to know which version is current. It picks one, or it blends them.

So the drift is not an AI problem. It is a product information problem that AI has made visible, and audible, to the shopper. Every time the brand lets a surface fall out of line with the pack, it hands the assistant a wrong fact to repeat.

Numbers above: measured, Prodeen Claim Fit sample of September 2026, published 20 October 2026; desk research against official texts, not legal advice.

Three regulators, one month, the same message

None of the three markets has written a food rule for AI assistants. All three have said, in the last two months, that the rules they already have apply to them, and that a company answers for an agent the way it answers for an employee.

MarketInstrumentDateStatusWhat it says about the assistant's answer
EUDigital Fairness Act proposal, COM(2026) 1100, Article 2611 Nov 2026Proposal; applies 12 months after entry into force (Article 40)Traders who offer consumers an automated system to conclude contracts must let the consumer set its parameters, ask at first use whether every contract needs confirmation, and disclose payments received from traders and conflicts of interest
EUDigital Fairness Act proposal, Article 1111 Nov 2026ProposalEvery paid commercial communication by an influencer carries the word "advertisement" or "commercial collaboration", in a clear and prominent way
EUAI Act, Regulation (EU) 2024/1689, Article 50OJ 12 Jul 2024Applies since 2 Aug 2026A person interacting with an AI system must be told so, unless it is obvious; AI-generated text published to inform the public must be marked
UKCMA, Complying with consumer law when using AI agentsFirst published 9 Mar 2026, updated 30 Oct 2026Guidance in force"If an AI agent you use does something illegal, you are responsible"
UKDRCF joint statement on an agentic AI economy (CMA, ICO, FCA, Ofcom)30 Oct 2026StatementExisting legal frameworks apply to AI agents; an agent's autonomy does not diminish the responsibility of the organisation that deploys it
USFTC chair, remarks at a Reuters event (press report, no FTC text)25 Sep 2026Position, unofficialAgents are tools; whoever instructs the tool is responsible for what it does; existing law is enough

Read the three columns together and the pattern is the same one we saw with claims. In the EU the rule is written down in advance, with an application date. In the UK the regulator says the old rule covers the new tool and tells you to train the tool. In the US the enforcer says the same thing from a stage and lets the first case set the line. Different speeds, one direction: the answer the assistant gives is a commercial communication, and someone will be asked to stand behind it.

What practitioners get wrong here is the word "agent". They read these texts as rules for the companies that build assistants. Article 26 is addressed to the trader who makes an automated system available to consumers. The moment a brand puts a chatbot on its site, feeds a retailer's assistant with product data, or pays for a placement in an assistant's answer, it is that trader.

Who answers

The honest answer today is: it depends on how close the brand stands to the assistant. Three distances, three positions.

At the first distance, a general assistant reads public pages and gets a fact wrong. The brand did not build it and did not feed it. The brand still owns the pages the assistant read, and the easiest defence in any market is that every surface the brand controls said the same correct thing. If the retailer page was wrong for two years, that defence is gone.

At the second distance, the brand steers the answer. It feeds a retailer's assistant with a product data file, publishes structured data for assistants to read, or pays for its product to appear in an answer. In the EU, Article 26 of the proposal asks that payment to be disclosed. In the UK, the CMA calls this "using" an agent. The answer has become the brand's commercial communication, under the same rules as the pack.

At the third distance, the brand runs the assistant: a chatbot on its site, a nutrition helper in its app. From 2 August 2026 the shopper must be told it is talking to a machine. Everything that machine says about the product is a claim the brand made.

What to do before the first case

Three things, and all three are product information work that a regulatory team can start this quarter.

One product information record, rendered per surface. The pack, the brand site, the retailer feed and the structured data for assistants are four outputs of one record, with one owner and one change date. When the recipe changes, every surface changes on the same day, and the assistant has nothing stale to find.

A claims register with evidence. Every claim the brand makes anywhere, with the market rule it meets, the proof behind it and the name of the person who approved it. Half of the claims we checked in September disagreed across surfaces; a register is how that number goes to zero.

A monitor on what assistants say, with a correction trail. Ask the main assistants the questions a shopper asks, once a month, per market. Log the wrong facts, trace each one to its source surface, fix the source, ask again. When the first regulator asks "what did you do when you knew", the log is the answer.

This question is worth more detail than one article can give, so we will go into it live. On Thursday 28 January 2027, at 16:00 CET, Cesare Varallo, lawyer in food law and founder of Food Law Latest, joins me for a webinar. We will show what assistants say today about a handful of anonymised products across the EU, the US and the UK, who answers for each wrong fact, and how a product information record and a correction trail close the gap. Save the date and register, or get in touch if you want to run the test on your own products first.

Save the date: live webinar, Thursday 28 January 2027, 16:00 CET

With Cesare Varallo, lawyer in food law: what assistants say about your products, and who answers for it.

Register for the webinar