Bureau Lila measures what AI systems genuinely owe the content they consult. Our instruments establish, source by source, what the presence of each one changes in the answer produced. What changes nothing contributed nothing, however closely it resembles the finished text.
Generative systems consume content produced by others. On what that consumption is worth, and to whom it is owed, no instrument exists today: the question is settled by estimate, between parties who never put forward the same figures.
Bureau Lila builds the instruments that are missing. We do not claim to reconstruct the inner workings of the systems we measure, nor to hold a truth nobody could verify. We establish a measurement convention: a method published, sealed before execution, replayable by a third party, and on which parties who agree on nothing else can nonetheless agree.
That is the difference between an opinion and a standard. An opinion is contested; a convention is adopted. We advocate for no one: we measure, and we hand the same measurement to everyone, including the party being measured.
A quantified basis to put on the table, where negotiation still rests on indicators that no longer describe how content is actually used.
A way to check the reliability of what operators report, without depending on their declarations alone or gaining access to their systems.
An adversarial, auditable record — preferable to a negotiation in which each side advances its own figures with no referee.
The bureau is backed by no publisher, no platform and no collecting society, and draws no remuneration indexed to the outcome of a measurement. That is the condition for a record to be worth anything once it is on the table.
First campaign of the TX68 protocol, 17 to 24 August 2026. Every hour, articles published in a dozen French daily newspapers were turned into queries a reader might plausibly have typed, then submitted to Google. Six thousand three hundred and twelve AI Overviews were collected, timestamped and archived.
The figures below are established from the sources Google itself displays beneath its answers, and from nothing else. It is the most conservative reading available, and it is enough to measure the place the press occupies.
of the AI Overviews collected cite at least one news site as a source
of the sources cited are news publications
news domains identified in the single week measured
On news queries, then, the press is not a supplementary source. What remains is the question these figures do not settle: among the sources displayed, which ones actually made the answer, and in what proportion? Counting links does not answer it. That is where measurement begins.
The links displayed beneath an answer indicate that a document supports a statement. They do not establish that it determined its content. These are two distinct quantities, and only the second can ground a payment: what is paid for is what was contributed to the process, not a resemblance to the finished product observed after the fact. A practical difficulty compounds this, since the number of links displayed depends on the layout of the page and on decisions taken by the very party that owes the payment.
Measuring real contribution means giving up the shortcuts. Two come naturally to mind, and both lead to absurd allocations. The case below shows where they give way.
An equitable key must satisfy a few simple requirements: two sources contributing the same thing receive the same share, a source with no effect receives nothing, the whole of the value is distributed with neither surplus nor shortfall, and splitting the problem does not change the result. A classical result in game theory shows that a single rule satisfies all of them at once.
What remains is knowing how to measure what a set of sources actually contributes. That is where the operational substance of the TX68 protocol lies, and it is the subject of the report.
The protocol is versioned and documented at every revision. It runs in repeated campaigns, so as to produce series comparable over time rather than an isolated snapshot, and extends to other engines as well as to other content markets.
Applied to the whole set of Overviews collected, the measurement produces a complete allocation key: each domain receives a share, together with its confidence interval. The shares add up to exactly one hundred per cent, which no allocation pro rata to citations can guarantee.
The gap with a uniform allocation is considerable, and it does not always run in the direction one imagines. Some heavily cited titles weigh little; others, inconspicuous in the interface, carry the bulk of the answer. The resulting distribution is moreover markedly more concentrated than is generally assumed.
This is a fact, not a recommendation. An explicit measurement opens to debate a trade-off that current keys left in the dark: the scope of beneficiaries, the base amount and any pluralism adjustments remain decisions that belong to the parties and that apply after the measurement.
The named ranking is reserved for recipients of the report. We provide it to rights holders, collecting societies, authorities and operators on request.
Request the reportBureau Lila did not come out of a market study. It came out of several years of audits run against platforms that did not declare what they were doing, and out of the finding that an instrument to measure usage was missing.
Method and measurement
PhD in algorithmic auditing. Postdoctoral researcher at the CNRS, affiliated with the médialab at Sciences Po and with AI Forensics.
Ran the tests with Mediapart, in February 2026, establishing that a major French language model reproduced long passages of copyrighted works. Quantified, under Article 39 of the DSA, an influence campaign that reached 38 million people in France and Germany without ever being identified as political. Further work on Amazon's recommendations and on the biases of on-device models.
Instrumentation and scale
Co-founder and ex-chief technology officer of Check First, a Finnish adversarial research company. Contributor to the ObSINT guidelines.
Built the instruments that made it possible to document the Pravda network with the DFRLab — the methodical injection of pro-Kremlin content into Wikipedia and into the answers of conversational agents. Stress-tested the advertising transparency mechanisms of eleven major platforms for the Mozilla Foundation. A developer first and foremost, after thirty years of open source tooling.
Work covered by Mediapart · Le Monde · The New York Times · The Washington Post · The Guardian · Politico
Sixteen pages, published in August 2026. The full version 1 protocol, contribution measurement, an allocation key across one hundred and nineteen domains, worked cases and limitations.