90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-079
Travel · Australia · 2022 · Misrepresented algorithm

Trivago advertised the best hotel deal while its algorithm favoured the advertiser that paid it most, and the penalty was $44.7 million

By Ellie Harris · Filed Conduct December 2016 to September 2019; ACCC proceedings from August 2018

Alleged: Trivago N.V. (Expedia Group) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

Trivago advertised the best hotel deal while its algorithm favoured the advertiser that paid it most, and the penalty was $44.7 million

What happened

It was reported that in April 2022 Australia’s Federal Court ordered Trivago to pay $44.7 million in penalties for misleading consumers about what its hotel comparison site actually did. From late 2016 to September 2019, Trivago’s television and online advertising presented the site as the quickest way to find the cheapest available hotel room. The court found the site’s ranking algorithm did something different: in selecting which offer to display as the top result for a hotel, it gave significant weight to which online booking site paid Trivago the highest cost-per-click fee. The result, on the evidence before the court, was that higher-priced offers were selected as the top position over available cheaper offers in 66.8 percent of listings. The trial judge described the television advertising as highly misleading, and the full court dismissed Trivago’s appeal.

It was reported that the court heard there were about 111 million click-outs on the Trivago website in the period, roughly 104 million of them on the top position offer, and that Trivago admitted it earned approximately $58 million in fees from clicks on offers that were not the cheapest available, with the ACCC putting the amount consumers overpaid for rooms at about $38 million; the judge, adjusting for cancelled bookings, worked with a loss figure of about $30 million. Announcing the penalty, the ACCC chair said consumers were misled into thinking they were getting a great hotel deal when that was not the case, and called the outcome a message to other comparison websites that they must not mislead consumers when making recommendations. The case has become a standard Australian citation for a now common pattern: a recommendation algorithm whose marketed objective, serving the user, differs from its actual objective, serving the platform’s revenue.

What an auditable version would have shown

It took litigation, from an August 2018 filing through a liability judgment and an appeal to a 2022 penalty, to establish what the algorithm optimised for. An auditable version makes an algorithm’s objective a matter of record rather than litigation: signed records of each ranking decision, capturing the candidate offers, their prices, the fees attached, and the score that chose the winner. Aggregated, those records answer the exact question the court had to reconstruct, how often does the top position go to something other than the cheapest offer, and why, as a running number rather than a forensic exercise over three years of conduct.

Where the gap was

The gap here is between the claim made to consumers and the objective given to the algorithm. A MetricRecord computes the divergence continuously: percentage of rankings where the displayed choice was not the cheapest, revenue weight versus price weight in outcomes, published internally as a signed aggregate someone must reconcile with what the ads say. A ConductRecord preserves individual ranking decisions with their inputs, so that a regulator’s question about any given listing has an answer in the record. Together they turn “what does this algorithm actually do” from a trade secret into an auditable fact.

What governance should have looked like

When a company tells the public its algorithm serves them, that claim is a specification, and someone should be testing the system against it. The gap between marketing and objective function is one of the quietest failure modes in consumer AI because nothing breaks: the site works, the results look plausible, and the divergence is visible only in aggregate. Governance for recommendation systems should treat advertised behaviour as a constraint, measure conformance to it continuously, and keep records fit to show a court, because the alternative, as here, is that the truth about the algorithm emerges only through one.

Failure Pattern: a recommendation algorithm’s actual objective, advertiser revenue, diverged from its advertised objective, the best deal for the user, and the divergence was visible only in aggregate.

Governance Principle: an algorithm’s advertised objective is a specification: conformance between what the system is claimed to optimise and what it actually optimises must be measured continuously and kept on the record.

The reference implementation of MetricRecord and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Trivago N.V. (Expedia Group) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →