EU regulators fine fake test sites. Anthropic now names them in Claude’s rules

Fake test sites now break EU consumer law, Google’s spam rules and, from 12 November, Claude’s usage policy. Each rulebook catches a different part of the same trick. On 15 October, Denmark’s Consumer Ombudsman said Obsidian Media will pay DKK 400,000, about €54,000, for five sites that posed as independent product tests. Nobody at the company had tested the products.

Obsidian Media earned commissions from tests it never ran

The five sites were forbrugsguiden.dk, forbruger-test.dk, senge-testeren.dk, bomagasinet.dk and bedste10.dk. They had “pros and cons” boxes and a section called “our assessment”. That design suggested someone had used each product. According to the Ombudsman’s press release, the company only read other people’s tests, reviews and product specs.

The money came from affiliate deals. In such a deal, a shop pays a commission for each visitor or sale that a partner site sends its way. Readers were not told about these deals. The Ombudsman found that the sites were built to send traffic to partner shops and that their layout hid this goal.

Consumer Ombudsman Torben Jensen called it a “gross violation” of the Danish Marketing Practices Act (translated from Danish). The decision cites the bans on misleading marketing and hidden advertising. The Ombudsman reported the company to the police in the summer of 2025. Obsidian Media sold the sites in September 2025. We covered the fine on Obsidian Media when it was announced.

The release also says affiliate marketing is usually legal. It becomes illegal when people cannot see that a deal stands behind a link, review or recommendation.

Italy fined Trustpilot €4 million over reviews it did not check

Denmark is not the only EU country acting on fake reviews this year. On 23 March 2026, Italy’s competition authority, the AGCM, fined three Trustpilot companies €4 million. The AGCM press release says the firms did not check well enough that reviews were real. That included reviews marked as verified.

The AGCM also found that paid tools let businesses choose which customers got review invitations. That can make a rating look better than the full customer base would rate it. Trustpilot did not explain clearly how the site works or which firms paid for its services. Silicon UK reported that Trustpilot disagreed with the decision and planned to appeal.

In the Netherlands, the consumer authority ACM said in August 2024 that it had started investigating firms that sell fake reviews. It can fine up to €900,000 per violation, according to the ACM statement.

All three cases rest on the same EU rules. Since 28 May 2022, the Omnibus Directive, an update of EU consumer law, bans posting or ordering fake consumer reviews. It also bans claiming that reviews come from real buyers without checking it. National laws must allow fines of at least 4% of yearly turnover, or €2 million when turnover data is missing.

Before the rules took effect, authorities from 26 EU countries, Iceland and Norway checked 223 shops, booking sites and comparison sites. At least 55% may have broken EU law. Some 176 sites did not say whether they ban or label rewarded reviews, according to the Polish consumer portal konsument.gov.pl.

Google treats fake sources as spam in its AI answers since May

Google fights the same sites through its search rules. Between 13 and 15 May 2026, it added one phrase to its spam policies, Search Engine Roundtable reported. Spam now includes “attempting to manipulate generative AI responses in Google Search”. This covers AI Overviews and AI Mode, the AI answers shown in search results.

The current spam policies page, last updated on 28 August 2026, also targets “thin affiliation”. These are pages full of shop links that add little beyond the shop’s own text. Google says good affiliate pages offer things like original reviews, testing or comparisons. Obsidian Media’s sites claimed testing that never happened.

Google’s September spam update finished rolling out on 8 October. Earlier, Google also called fake author profiles deception in its content guide. It does not ban AI-written text as such, as our review of Google’s AI content rules showed.

Anthropic, OpenAI and Google write very different rules for their AI tools

We compared the usage rules of three AI makers. Each one bans deception in general terms. Only one names fake reviews, fake site networks and AI answer sources directly.

Anthropic published a new policy that takes effect on 12 November 2026. A section called “Do Not Engage in Deceptive Campaigns or Artificial Activity” bans fake personas, fake media outlets and fake reviews. It bans networks of “fake or ostensibly independent” websites. It also bans efforts to “manipulate the sources from which search engines or AI systems draw answers”. Building tools for such work is banned too. We reported on the new Claude usage policy on 10 October.

OpenAI uses broader wording. Its usage policies, in force since 29 October 2025, say its services may not be used “to manipulate or deceive people”. They list deceit, fraud, scams, spam and impersonation. The page does not mention fake reviews, fake site networks or search and AI answers.

Google has a separate policy for its AI tools, last changed on 17 December 2024. It bans frauds, scams and other deceptive actions. It also bans claiming that AI content was made only by a human in order to deceive. It has no line on fake reviews. Google covers that part in its search spam rules instead.

LKM Company used Claude for 8,913 articles on about 70 fake sites

Anthropic’s September 2026 threat report shows the kind of operation the new rule targets. Investigators traced about 70 fake news sites to LKM Company, which the report calls a France-based digital advertising agency.

Claude wrote and rewrote articles in a fixed format. Each one had three to four internal links. The report says the articles were “specifically designed to boost their site’s authority rankings on search engines”. The network published at least 8,913 articles in about 20 languages. Most drew little real response. Anthropic banned the account and the company.

The case shows why a usage policy matters. Consumer authorities fine the company that runs the sites, after a long investigation. Obsidian Media’s case took more than a year from police report to fine. An AI maker can act earlier, by cutting off the tool that writes the content.

GEO agencies face the Claude ban from 12 November

GEO, or generative engine optimisation, means work aimed at getting a brand into AI answers. From 12 November, an agency that uses Claude to fill several unrelated-looking sites with praise for one client can break Anthropic’s rules. The policy says Anthropic may limit, suspend or end access after a breach. It covers developers who use the API and people who use apps built on Claude. In the EU, national laws must also allow fines for such sites of at least 4% of yearly turnover.