Method
How it works, part by part
A research tool is only useful if you know what it did to produce an answer. These are the ten parts it is built from. Each one exists because of a problem you would otherwise run into.
01
The register
A precedent tool is worth what is in it, and how recent that is.
Every decision comes from the European Commission's own public case register. We add no commentary of our own.
The register is checked every day. New merger cases are added, changed ones are updated, and a decision is read once its text is published. The register lists more than 10,300 notified mergers since 1990. More than 9,900 of them have a published decision that has been read into the scheme.
02
The coding scheme
The text of one decision cannot be compared with the text of another.
So each decision is first read into a fixed scheme, one part at a time: the deal and the parties, jurisdiction, procedure, every market, every theory of harm, the defences, the economic evidence and the remedies. Most variables are chosen from a fixed list, which is what lets decisions be compared and counted.
The parts refer to each other. A theory of harm is attached to the market it concerns, and a remedy to the markets it is meant to fix.
03
Reading in pieces
A language model handed several hundred pages at once starts to skip markets.
Long decisions are therefore read chapter by chapter and put back together. A Phase II decision with a hundred markets is coded market by market, not summarised in one pass.
After a reading, the tool checks whether every chapter that defines a market produced one. If a chapter came back empty, the decision is read again in smaller pieces.
04
The checks
A reading can be incomplete without anyone noticing.
Automatic checks flag a reading that looks wrong, for example a conditional clearance with no remedy recorded, or far more theories of harm than markets. A flag is a prompt for a person to look at the case.
A reading can also be compared with the Commission's own press release on the case, which states the outcome in the Commission's words.
When a decision is read again, the new reading does not silently replace the old one if it would lose more than half of what was there. Both are kept until someone has compared them.
05
The paragraph reference
A point you cannot check is a point you cannot use.
Each recorded value is stored with the paragraph of the decision it was taken from, and theories of harm and remedies with a quotation in the Commission's words.
In the tool, the paragraph opens next to the answer. That is the check: read the source before you rely on it.
06
Comparable cases
To look up a precedent you first have to know which cases to look for.
A deal you describe is read into a profile of the facts known at notification: products, industry, overlap, countries, share and structure. You can correct the profile before anything is searched.
Decided cases are ranked by how close their product markets are to yours, with industry and recency as tie-breakers. Countries, shares and deal structure are shown as reasons for and against each case. They do not move the ranking, because in our tests they did not make it better. If you only want cases from one country, ask for that as a filter.
Each result says how close its best market is. A weak match shares an industry and not a market: it is a neighbour, not a precedent, and the tool says so.
The outcome of a case is never used to find it. What the Commission did is what you are looking for, so matching on it would be circular.
07
The citation test
Anyone can call two cases comparable. Someone has to be the judge.
We let the Commission be the judge. When it defines a market, the Commission cites earlier decisions. For each decided case we check whether the decisions it cited come out on top of our ranking.
Every change to the ranking is measured against that score before it goes live.
08
The Analyst
A question about the case law is half filter and half meaning, and a search box handles neither well.
Your question is first matched to the coded data: which theory, which outcome, which remedy, which sector, which years. Filters that describe the same thing have to hold on the same record, so a theory and its outcome are matched on one theory, not anywhere in the case.
Passages are then retrieved from the matching decisions by meaning, and a model writes the answer from them.
Under each answer is the list of cases it was built on, made from what was actually looked up. Both steps can still be wrong. A model can miss something, or state a point with more confidence than the decision supports.
09
Register facts and readings
A fact from the Commission and a reading by a model are not the same kind of thing.
Some of what you see is the Commission's own data: the deadline, the phase of the investigation, the timeline of the procedure and the links between cases. The rest is read from the decision text by a model.
The two are stored apart and labelled apart, so you can always tell which one you are looking at.
10
Court review
A decision that was annulled is a different precedent from one that stood.
Judgments and orders of the EU courts are linked to the decision they review, together with the Advocate General's opinions. An opinion is shown as an opinion: it decides nothing, and it is not counted as an outcome.
One case number can cover more than one decision, for example a new decision after an annulment. A ruling is only counted against the decision it is actually about.
What it does not do
It does not give legal advice, and it is not a substitute for reading the decision or for professional judgement. It covers merger control only. Cartel and abuse of dominance decisions are coming soon. Not every case has a published decision, and a small number of decisions could not be read: the Cases page in the app shows the status per case.