Methodology

Attendance

Attendance measures the share of plenary sitting days on which an MEP was recorded present, based on the European Parliament's official attendance records. It is shown on MEP profiles together with a month-by-month trend.

Rates are computed per calendar month and per parliamentary term. Months with too little underlying data are suppressed rather than shown as misleading 0% or 100% spikes.

Presence is not the same as activity: an MEP can be present and silent, or absent on official delegation business. Treat attendance as one signal among several, not a verdict.

Group loyalty

Group loyalty measures how often an MEP votes with the majority of their own political group in roll-call votes. Only decisive votes — for or against — are counted; abstentions and absences are excluded.

A vote is rebellious when the MEP's decisive vote goes against their group's majority on the same roll call. Loyalty is 100 × (1 − rebellious ÷ decisive). Group-level cohesion uses the Rice index (0–100). Group majorities are recomputed from the votes actually cast on each roll call, and mid-term group switches are handled through dated membership history. Months in which an MEP cast fewer than five votes are suppressed.

Loyalty is not a quality metric. Dissent can be principled and discipline can be routine; the score tells you what happened, not whether it was right.

Legislative stage

Every legislative procedure is tracked from the European Parliament's Legislative Observatory (OEIL) with its full stage history — from Commission proposal through committee, plenary, and adoption. The stage shown is the latest recorded event.

Votes come from official roll-call records, which name every MEP's choice. Lobbying data comes from the EU Transparency Register and published Commission meeting disclosures. National implementation is followed through official journals and parliament sources, linked to the parent directive by its CELEX identifier.

Data refreshes continuously — most sources update within hours of publication. Where a page shows aggregated numbers, the underlying records are linked so you can check them yourself.

Caveats

Roll-call votes only: many Parliament votes are by show of hands and are not recorded per MEP, so per-MEP metrics cover roll calls only
Upstream lag: official attendance and vote records can be published days after a sitting, and upstream corrections propagate on our next refresh
The Transparency Register is self-reported by lobbying organisations; GovLens displays it but cannot independently verify budgets or client lists
National parliament coverage varies: some chambers publish structured data, others require parsing documents, so depth differs by country
Group loyalty depends on how the group majority is defined; we compute it per roll call from the votes actually cast
AI-generated summaries and analyses are always labelled as such and never replace the linked primary sources
Metrics with thin data are suppressed, not extrapolated — an empty chart means insufficient data, not zero
GovLens is not a poll and does not measure public opinion
GovLens is non-partisan: we publish what happened and do not advocate positions

Ask GovLens — how AI answers work

Ask GovLens is an AI research assistant that answers questions using only GovLens data: legislative procedures and their stage history, MEP votes and voting records, and the EU Transparency Register with Commission meeting disclosures. It cannot answer questions outside that corpus — opinions, predictions, legal advice, or non-EU politics are refused with an explanation rather than guessed at.

Citation verification. Every sentence in an answer must carry at least one citation to a GovLens entity. Before a sentence reaches you, each of its citations is re-checked against the database — quoted values must match the stored record. Sentences that fail this check are removed, and the answer tells you how many were removed. If most of an answer fails verification, the whole answer is withheld.

Generated-content labelling. Answers are AI-generated and are labelled as such wherever they appear — in the chat, in copied text, and in dossier exports. Generated text is never presented as editorial fact; the citations, which link to primary GovLens records, are the factual layer. Answers are produced with a large language model via OpenRouter; the model in use is disclosed in our runbook and may change as quality gates are re-run.

Logging. Questions and answers are logged without personal data to measure answer quality (verification rates, refusals) and to build evaluation sets from reported answers. Retention windows are documented in the privacy policy.

Transposition deltas — how national divergence is classified

A transposition delta is a versioned, AI-assisted comparison between one article of an EU directive and a Member State's national measure implementing it. Every delta is classified into exactly one of six types:

Faithful — the national measure matches the directive article without material change.
Gold-plating — the national measure adds obligations stricter than the directive requires.
Carve-out — the national measure exempts cases the directive covers.
Narrowed scope — the national measure applies to a smaller set than the directive intends.
Not yet transposed — no national measure has been found for this article.
Unclear — the evidence available is insufficient to classify with confidence.

Verbatim gating.Every non-"not yet transposed" finding must quote a passage from the directive article and, where a measure exists, a passage from the national text, both located exactly in our stored sources. If a quoted passage cannot be verified against the stored directive or measure text, the finding is demoted to "unclear" rather than published unverified.

Translation caveats. National passages are quoted in their original language as retrieved from the national gazette. Where a stored English translation is available, the comparison is made against that translation; the quoted national passage itself is not machine-translated for you, so read it alongside the original-language source link.

Confidence.The confidence score attached to each delta reflects the model's certainty in its own classification given the available evidence. It is a measure of analytical confidence, not legal certainty, and should not be read as a probability that the classification would hold up in court or before the Commission.

Versioning. Deltas are append-only: re-analysis of an article/country pair (e.g. after a new national measure is found) writes a new analysis version rather than editing the old one. Every screen, export, and API response in GovLens shows only the latest analysis version per article and country; older versions remain in the database for audit but are not surfaced.

Not legal advice.Transposition deltas are an AI-assisted research aid, not a legal opinion. They are not a substitute for legal advice and should not be relied upon as a determination of a Member State's compliance with EU law.

Questions about our methodology? methodology@govlens.eu Read our democratic responsibility statement