France is the only country that makes manufacturers publish a durability score and the raw data underneath it. We pulled the complete register, recovered the weights the decree never published, and separated every sub-criterion. Here is what the index measures, the tenth of it that measures nothing, and the spare-parts problem the headline number hides.
France's Indice de Durabilité is the most transparent consumer durability score in the world, and one of the most widely misread. This article is our complete account of it: where the data comes from, how the number is assembled, what each part of it actually contributes, and the specific ways a published score will mislead you if you take it at face value.
It merges and supersedes two earlier pieces of ours — a July note on the warranty criterion drawn from a 274-record sample, and an August analysis of the full register. Where those two disagree, the figures here are the ones we stand behind; several numbers have been corrected against the source data and are marked as such.
An index you can actually audit
Most consumer quality scores are either an agency's trade secret or an algorithmic black box. This one is law.
Decrees of 5 April 2024, made under the AGEC act, prescribe exactly which criteria and sub-criteria are computed and how they are weighted. Manufacturers must publish the underlying data on data.gouv.fr under an open licence, and keep proof of the testing on file for inspection by the DGCCRF. That combination — a mandated methodology plus mandated publication of the inputs — is what makes the index auditable from the outside. You can take the manufacturers' own declarations and rebuild the score yourself.
That is what we did.
What is in the register
The consolidated register we work from holds 2,769 models:
| Category | Models |
|---|---|
| Washing machines (lave-linge) | 1,548 |
| Televisions (téléviseur) | 1,221 |
Those are the only two categories that exist so far; the AGEC framework extends to further product groups over time. Coverage within them is effectively complete — every model legally required to publish a score. 2,768 of the 2,769 carry an EAN, and 2,209 link to a manufacturer detail PDF.
Six records are flagged and excluded rather than read: three placeholder rows left in the official dataset (randomised manufacturer names, exemple1.fr URLs) and three whose declared block totals reconcile to neither of the two published scales. The reliability analysis below therefore runs on 2,764 models.
Published scores span most of the available range — washing machines average 8.23 (4.8–9.4), televisions 6.98 (2.3–9.1).
How the score is built
The index runs 0–10 and splits into two equally weighted blocks: repairability and reliability. Each block has its own criteria.
Repairability (50% of the index) — four criteria at equal weight:
| Criterion | What it covers | Weight in block |
|---|---|---|
| A1 | Documentation availability | 25% |
| A2 | Disassembly, tools, fasteners | 25% |
| A3 | Spare-parts availability | 25% |
| A4 | Spare-parts price | 25% |
Reliability (50% of the index) — three criteria at very unequal weight:
| Criterion | What it covers | Weight in block |
|---|---|---|
| B1 | Resistance to stress and wear | 50% |
| B2 | Maintenance and servicing | 40% |
| B3 | Commercial durability guarantee + quality process | 10% |
These weights are not published anywhere as such. We recovered them from the data by two independent routes — the maximum observed contribution divided by ten, and a least-squares fit of each block score on its raw criteria — which agree to three decimal places. They are identical across both product categories.
One structural detail worth knowing: B1's internal split is inverted between the two categories. For washing machines it leans 0.90/0.10 toward resistance to external stress; for televisions it leans 0.10/0.90 toward wear. This is real, not an artefact — both sub-criteria vary in both categories.
Resistance to wear is measured concretely: washing machines by completed wash cycles under CEN EN 50731, across roughly a 1,400–3,400 cycle range; televisions by testing a minimum of five units of the same chassis under the manufacturer's own declared methodology.
Reading a published number: two scales in one column
This is the single most important thing to understand about the register, and the place where our own July analysis went wrong.
The register publishes criterion values in two different units, and does not say which is which.
- Sub-criteria (
note_X_cN.M) are always on the raw 0–10 scale. - Criteria (
note_X_cN) are normally the weighted contribution to their block. So they cap at their own weight, not at ten. B1 tops out at 5.0; B3 tops out at 1.0 for a television and 0.5 for a washing machine.
A published B3 value of 0.25 is therefore not "0.25 out of 10". On the actual 0–10 scale it is 2.5. Our July piece reported this criterion as sitting between 0.25 and 0.5 "out of ten" and concluded the score was near zero. The conclusion survived; the arithmetic behind it did not, and everything below uses the corrected reading.
The trap has teeth in the other direction too. One manufacturer — 366 washing-machine models — declares its criteria on the raw 0–10 scale instead. Rank the register on the bare criterion column and that manufacturer's 9.5 sits above every rival's structural maximum of 5.0, purely as a unit artefact. It looks roughly twice as good as its competitors on every criterion, and it is not. The right fix is to detect the convention per block from the residual, not to hard-code it to a brand.
The reliability half: what each criterion actually does
If a criterion carries a tenth of the weight, you would expect it to explain something like a tenth of the differences between products. Here is the share of variance in the published reliability score that each criterion accounts for:
| B1 wear | B2 maintenance | B3 guarantee | |
|---|---|---|---|
| Washing machines | 39.3% | 60.7% | −0.1% |
| Televisions | 53.3% | 45.4% | 1.3% |
A negative share is not a typo — it means the criterion's small covariance with the total runs slightly against it. In plain terms, B3 contributes nothing and very marginally works against the ordering.
Delete B3 entirely, redistribute its weight proportionally to B1 and B2, and compare the resulting ranking with the published one:
| Washing machines | Televisions | |
|---|---|---|
| Rank correlation with published score | 0.999 | 0.994 |
| Product comparisons that change direction | 0.4% | 1.3% |
| Mean reliability score, published | 7.61 | 7.15 |
| Mean reliability score, B3 removed | 8.15 | 7.64 |
So B3 does one thing: it takes about half a point off everyone's score. It changes the answer in roughly four comparisons in a thousand for washing machines, and thirteen in a thousand for televisions.
Why B3 is inert: a floor and a ceiling
B3 is built from two sub-criteria — the commercial guarantee counts for three quarters of it, the manufacturer's quality process for the remaining quarter. Each one is stuck.
- Guarantee duration is zero for 87.5% of washing machines and 97.5% of televisions. Almost nobody offers a commercial durability guarantee beyond the two-year statutory EU minimum.
- Quality process is full marks for 92.2% of washing machines and 99.8% of televisions. It is a self-declaration that the manufacturer runs a continuous-improvement process, and essentially everyone declares it. Across the entire register this sub-criterion takes exactly two distinct values.
Combine a floor weighted 0.75 with a ceiling weighted 0.25 and the arithmetic has one likely answer: 2.5 out of 10. 86.6% of washing machines and 69.8% of televisions sit on exactly that value.
A criterion on which almost every product scores identically cannot separate products, whatever weight the decree assigns it.
The exceptions are real
104 models — 3.8% of the register — score 5 or better on B3, so the low scores are a commercial choice, not a technical impossibility.
Eleven televisions score a full 10.0; ten of them are Metz Display models. Our July note cited a single Easypitch record at 10.0. That citation was wrong: the record's declared reliability block reconciles to neither published scale — the value exceeds the criterion's own ceiling by a factor of ten — so it is a malformed entry, not an exceptional product. It is one of the six records we now flag.
On the washing-machine side no model exceeds 5.0, and the makers that reach it are Miele (34 models), Westpoint (11), Qilive (11) and Essentielb (3). Miele is the largest of them, not the only one — our August analysis said otherwise, and that was wrong.
Repairability is not a proxy for durability
France replaced its earlier repairability index with this durability index, and the register shows why that mattered. The two halves of the score are only loosely related:
| Correlation, repairability vs reliability | Variance explained | |
|---|---|---|
| Washing machines | +0.33 | 11% |
| Televisions | +0.52 | 27% |
| Pooled | +0.41 | 17% |
A product that is easy to repair is somewhat more likely to be reliable, but the two are far from interchangeable. Between 73% and 89% of the variation in reliability is invisible to a repairability score. Anyone using a repairability rating as a stand-in for how long a product will last is filling in most of the answer with noise.
What the headline hides: spare parts, by channel
The repairability half rests on spare-parts availability, and the register carries far more detail on this than the score exposes. The source files contain a full parts matrix — four supply channels (repairer, producer, distributor, consumer) by up to fifteen standardised part slots. Extracted, that is 115,979 model-by-part observations across 2,214 declaring models and 30 distinct part types.
The finding it yields does not appear anywhere in the published score. The consumer channel is a lockout, and almost entirely for televisions:
| Share of parts declared unobtainable | Repairer | Consumer |
|---|---|---|
| Televisions | 4.2% | 17.0% |
| Washing machines | 2.0% | 3.4% |
Per manufacturer it is starker. Toshiba declares 0% of parts unavailable to repairers and 80.6% unavailable to consumers across 70 television models; Techwood, 0% against 74.3% across 23. The parts exist and are supplied — consumers simply cannot buy them. Most other market-placers declare identically in both channels, with a gap of exactly zero.
For televisions this gap reaches the published score: the correlation between consumer-channel unavailability and the overall index is −0.56. For washing machines the gap averages 1.4 percentage points and is uncorrelated with the score.
The parts hardest for a consumer to obtain are revealing in themselves:
| Category | Part | Unobtainable to consumers |
|---|---|---|
| Washing machines | Software and firmware, including reset | 44.0% |
| Televisions | Panel (dalle) | 37.1% |
| Televisions | External power supply | 32.4% |
| Televisions | Internal power supply | 22.0% |
| Televisions | LED backlight bar | 19.6% |
The washing-machine entry is the software-reset lock: on nearly half of models the one component needed to clear a fault code is declared unobtainable to the owner. That is an obsolescence mechanism sitting in plain sight inside an official durability dataset.
Five ways to misread the register
- Assuming a criterion column is out of ten. It is usually a weighted contribution. See the two-scales section above — this is the error that most distorts cross-manufacturer comparison.
- Treating scores as continuous. They are banded. Among washing machines, the wear sub-criterion takes three distinct values across 1,548 models and the quality-process sub-criterion takes two. Treat them as ordinal; differences of 0.1 are not meaningful.
- Reading
nom_metteur_sur_le_marcheas the brand. It is the legal market-placer, and there is no brand column at all. Groupe SEB covers Rowenta and Moulinex; Electro Dépôt covers Valberg; Vestel covers Continental Edison. Brand-level aggregation on this field silently merges unrelated products. - Aggregating to a brand average. Within-brand spread is large enough to destroy the signal: Haier's 302 washing machines span 6.5 to 9.4, Indesit's 78 span 6.0 to 9.2. A brand mean is a weaker guide than the score of the specific model you are looking at.
- Trusting the published precision. 1,746 of the 2,769 scores are published as unrounded computed values such as 6.84399350649351, while shops display them to one decimal place. That is the source file's own behaviour, not an import artefact, and the extra digits carry no information.
What this means if you use the index
Read the sub-scores, not the headline number. An overall 8/10 can hide excellent repairability sitting on top of ordinary resistance to wear, or the reverse.
On the reliability half, only two of the three criteria carry information. Resistance to wear and ease of maintenance are where products genuinely differ. The warranty criterion tells you something true and useful — that almost no manufacturer will stand behind its product past the legal minimum — but it tells you that about the entire market at once, not about the product in front of you.
That is a finding about policy, not about a washing machine. It belongs in an argument for extending guarantee obligations. It does not belong in a per-model score presented to a shopper as a comparison tool, where it functions as a near-constant deduction dressed up as a measurement. Ten percent of a legally mandated score is not a small thing to spend on a criterion that separates 3.8% of the market from the rest.
On the repairability half, check the parts position for the channel you will actually use. A television whose panel and power supply are supplied to repairers but not to owners is repairable in the index's sense and not in yours.
What we do with it
Every French-index record in our database carries the full breakdown into individual sub-criteria on a single reconciled scale, the test methodology for its product category, and the spare-parts matrix with delivery bands and years of coverage preserved as intervals rather than flattened into point estimates. The two declaration conventions are normalised so that manufacturers are comparable. The goal is that one summary number never hides what is underneath it.
The full methodology writeup is on the Methodology page.
Method
Source: the consolidated indice de durabilité files for lave-linge and téléviseur published on data.gouv.fr under Licence Ouverte, retrieved 27 August 2026 — 2,769 models, six flagged as malformed, 2,764 in the reliability analysis. Criterion weights recovered empirically by two independent methods agreeing to three decimal places. Variance shares are covariance-based and additive across criteria by construction. Spare-parts availability is measured on the declared delivery-delay field, where "non disponible" marks a part as unobtainable through that channel; availability bands are category-specific by decree and are not comparable across categories. Scores are banded rather than continuous and are treated as ordinal throughout.
Two limitations we would rather state than hide. The four sub-weights inside repairability criterion A3 are not identifiable from the data — the two declaration lists are pairwise collinear — so they are flagged unstable and never used. And the source CSVs carry no raw declared values behind the B criteria: there is no guarantee-years column, only scores. The underlying declarations live in the 2,209 manufacturer detail PDFs, which we have not yet parsed.
The ranking-stability figures here (rank correlation, and share of comparisons changing direction) were recomputed for this article and differ from those in our August note, which did not reproduce; the variance shares and mean scores reproduce exactly. The brand attributions for perfect B3 scores are corrected here for the same reason.
Sources
Decree of 5 April 2024 on durability index criteria for televisions and washing machines (Légifrance); schema.data.gouv.fr (etalab/schema-indice-durabilite); economie.gouv.fr; ADEME. Consolidated manufacturer datasets on data.gouv.fr, retrieved 2026-08-27. This article supersedes and replaces "The hidden gap in France's Durability Index" (6 July 2026) and "A tenth of France's reliability score measures nothing" (27 August 2026).
Further reading
What the Secondhand Appliance Market Reveals — and What It Doesn't
We analysed 15,221 secondhand appliance listings from seven marketplaces in three countries. The most-resold brand in Czechia is not Bosch or Samsung — it is Mora. Appliances lose half their value within three years, and expensive brands lose a larger share. And the one thing we wanted most from this data, it refused to give us.
Do Testing Agencies Agree on What Quality Is? A Digest of Our Research
We compared 16,978 test results from four European consumer testing organisations. Four of our studies: how strongly agencies agree (ρ = 0.84–0.92), whether higher prices mean higher quality, why the market rewards brands more than tested performance — and when the crowd agrees with the experts. With links to the full texts.