Published
A check tells you what the REP register said on the day you ran it. Monitoring tells you when that changes — a day later at best, stream by stream, in the source's own words. Here is how it works, what it reports, and what it will not do.
A one-off check against the REP register answers a narrow question well: on the date of the snapshot you queried, the register carried, or did not carry, a registration for this producer, in this stream, under this identifier. That is a real fact from the official source, and it is a fact about one day. It says nothing about the following morning, and nothing in it changes when the register does. The record in your supplier file goes on saying what it said on the day it was made, which is exactly what makes it good evidence and exactly what makes it a poor guide to the present. Most supplier files hold one answer of this kind, taken at onboarding, and far fewer say how old it is or which version of the register it came from.
Between two checks, the register keeps moving. ADEME adds producers, records withdrawals, corrects entries and reissues identifiers, and none of it is announced to the companies buying from the producers concerned. A supplier checked at onboarding in January can have had a registration withdrawn in May, and the only trace in your own file is still the January answer. Checking again before every order does not close the gap either. It costs time, it is the first step dropped when the team is busy, and even done faithfully it tells you the state of the register on the day of the order, not the day the change happened. Between a check that was right and a check that is out of date lies a stretch of time nobody looked at, and that is the stretch an auditor asks about.
Monitoring is therefore not a faster check but a different object. A watch compares the register with itself, one day against the day before, for the identifiers you name, and records each difference as an event tied to the two snapshots it was read from. What you get is not a status kept up to date somewhere, but a dated sequence: this registration was present on one date and absent on the next, with whatever the register published about why. That sequence answers the question that tends to follow an audit finding, namely when you knew, and not only what is true today. It also keeps running when nobody remembers to look, which is the point: the change you most need to hear about is often in a supplier nobody has thought about since onboarding.
ADEME publishes the register as open data and refreshes it daily, so that is the rhythm a watch can follow, and no faster. Repère takes a snapshot of both of ADEME's datasets every day and compares each new snapshot with the one before it. A change becomes visible when the pair that contains it is compared, which means an alert is at best a day old by the time it reaches you, and older still if the publication itself runs late. Nothing here is real-time, and a service that claimed otherwise would be describing how often it polls rather than how often the source changes. Detection is daily because the source is daily, and the honest way to read an alert is as news from the latest publication, not from the last few minutes.
Querying the source more often would not help. Read the register every hour and you would mostly fetch the same file again, then a new one, and the change would still date from the day it was published. The trail a watch leaves behind has the same resolution: a day, not an hour. Each event says that the register differed between a snapshot taken on one date and a snapshot taken on the next, and it names both, so anyone reading it months later can see which two versions of the register were compared and when each was taken. The comparison is also idempotent over a given pair of snapshots, so a job that has to re-run after a failure does not report the same change twice. An event also carries the stream it concerns and the IDU the register showed, so it can be filed without going back to the source to work out what it was about.
On a day when the register did not move for a supplier, a watch says nothing by default: an event is raised only when there is a difference to report. Some teams would rather have a dated line for every day, so the file shows an unbroken trail instead of a gap, and a watch can be set to send a daily notification for that purpose. That notification reports one thing only: that the comparison of two named snapshots produced no change for that supplier. It is not a statement that the supplier is compliant, and it should never be filed as one. On a day when the register did move for that supplier, the change event is that day's notification and no daily line is sent beside it, so the file never says on one date both that something changed and that nothing did. A register that did not change overnight can still be one in which the supplier was never registered for the stream you buy in; the daily line records what the source contained on two dates, and that is all it records.
ADEME publishes the register as two datasets, not one. The first, rep-producteurs-idu, lists producers currently registered and holding an identifiant unique; it held 308,665 rows on 11 October 2026. The second, rep-plus-en-conformite, lists producers no longer in conformity, and held 108,230 rows the same day. A withdrawal moves a row from the first to the second. The row does not stay where it was with a status switched to revoked, which is the shape most people assume when they first build against the register. A monitor written on that assumption watches a field that never changes value and says nothing on the one day it matters, and a monitor that reads only the first dataset sees every withdrawal as a blank where a row used to be.
The second dataset is also the only one that says anything about the withdrawal itself. Its rows carry two fields the active dataset never has: date_fin_conformite, the date conformity ended, and motif_desattribution, the motive. Those two facts are what make an alert worth filing, because they turn a blank into an event with a date and a stated reason, for one producer and one stream. So when a registration leaves the active dataset, Repère looks for it in the withdrawn one and quotes both fields on the event, as the register published them. The motive needs care of its own: it is not a clean enumeration, values combine, as in AUTRE || SIGNAL, and so it is parsed as a list rather than compared with one expected string.
How a registration is recognised from one day to the next matters just as much. The obvious candidate is the IDU, which on 11 October 2026 was always fifteen characters long: FR, the six-digit immatriculation, an underscore and six more digits, FR-prefixed even for the 156,427 Chinese producers on the register that day. But ADEME reissues IDUs, and a monitor keyed on the IDU reads a reissue as one registration ending and another starting, which is a false withdrawal and a false new registration on the same morning. Repère therefore identifies a stream by the pair of stream code and immatriculation, and reports beside it the IDU the register showed on each date. For the same reason, the prefix says nothing about where the producer is established.
There is a third case, and it is not rare enough to ignore. A registration can leave the active dataset without any matching row appearing in the withdrawn one. The usual reasons are mundane: ADEME corrects an entry at the source, or the producer registers again under another SIRET, so the old immatriculation drops out and a new one takes its place. Neither is a withdrawal in the register's sense, and the register describes neither of them. All anyone can see from outside is that a row was there on one date and is not there on the next, with nothing in either dataset to explain it. A monitor that reads only one dataset cannot tell this case from a withdrawal, and one that reads both has to decide what to say about it.
Repère reports it as its own event type, distinct from a withdrawal, and never guesses which of the two it was. The reason is practical rather than pedantic. Without an end date and a motive there is nothing to cite, and an alert that called the event a withdrawal would put a fact in your file that the register never stated, which is precisely the kind of inference an auditor is paid to take apart. The same rule covers a bad day upstream. If the withdrawn dataset is behind when the comparison runs, the registration that left the active dataset is still reported, as a disappearance, rather than held back until a reason turns up: an alert that says less than you would like is better than no alert at all.
Every event names one stream, because the evidence is per stream. A producer registers once for each stream it is concerned by, and a supplier can be registered for household packaging and absent from electrical equipment simply because it sells into one and not the other. A watch therefore reports stream by stream and never collapses the result into one yes or no for the company, and it can be limited to the streams you actually buy in. Read a supplier's events together, too. Registration is per establishment, so one SIREN can hold several immatriculations, as SIREN 384560942 did with three on 11 October 2026, and a disappearance under one immatriculation can arrive on the same day as a new registration under another. That opposite movement is reported too: a registration that appears for a watched supplier raises its own event, and it is often the other half of a story that began with a disappearance.
A watch takes a SIRET or a SIREN, and nothing else. A name match is a guess: company names are not unique, legal and trading names differ, and one group often registers several entities under nearly the same name. A guess made once can be checked by eye; a guess re-run every day raises false alarms about the wrong company, or stays silent about the right one, and nobody notices which. So a watch cannot be keyed on a name. A SIRET is resolved to its SIREN, so the watch follows every establishment the company has registered rather than a single site. Get the number from the supplier, from an invoice or from its official documents, and if all you have is a name, find the identifier first and confirm it before you watch anything.
Each watch delivers to one destination you configure: a webhook URL or an email address. Webhooks are signed with HMAC-SHA256 over the timestamp and the body together, so your endpoint can verify that a delivery came from Repère and was not altered in transit, and a captured delivery cannot be replayed forever, because its timestamp soon falls outside the window the receiver accepts. Delivery is best-effort, with a named retry budget: when an attempt fails, the queue retries after 1 minute, then after 5 minutes, 30 minutes and 2 hours — five attempts in all, spread over roughly two and a half hours — and then stops. An event that could not be delivered is not lost with the delivery; it stays readable through the API, which remains the record of what the watch saw, so a broken endpoint costs you the notification but not the trail. Push delivery is a convenience on top of that record, not a replacement for it: you can also read the events through the API on your own schedule, and reconcile what arrived against what was recorded.
When an alert arrives, treat it as what it is: a dated statement of what the register recorded, not a verdict on the supplier. A withdrawal with its end date and motive is a fact to file and, usually, a question to put to the supplier, which may hold a new registration under another establishment or have an explanation the motive field does not convey. A disappearance with no withdrawal record calls for the same question, asked more carefully. File the event as it came, with the two snapshots it names, rather than a summary of it: the snapshots are what let someone else check it later. Whether any of it changes the relationship is a judgement that belongs to you or to the authorities, and Repère does not make it for you. How many suppliers you can watch depends on your plan, and there is a free tier to try it on real suppliers before deciding how much of your list deserves the same attention.
Figures quoted in this guide were read from ADEME's published datasets on the date above. The register changes every day, so treat them as a dated observation rather than a current count.
Query the register by SIRET or SIREN and keep what it said, with the day it said it. No credit card to start.