Methodology: SIREN matching and confidence
Reachable matches a company to its SIREN, the French company registry number, by checking its website, name and address against SIRENE, then shows the method and a confidence score. At the default 0.95 threshold, our 10 November 2026 test on 200 real places gave 96 correct matches out of 97.
By Alexandre Rastello · Published on
Measured on 10 November 2026, 200 places in the Alpes-Maritimes
| List threshold | Correct | Wrong | Recall |
|---|---|---|---|
| 1 (“certain”) | 21 | 0 | 12% |
| 0.95 (default) | 96 | 1 | 56% |
Recall: correctly matched rows out of the 171 rows whose operator was established by hand.
What is SIREN matching?
It means finding the SIREN, the 9-digit number of every French company in INSEE's SIRENE registry, for a row that does not have one.
A row from Google Maps, a web search or your CSV carries a name, a domain or an address, rarely a SIREN. Without it, checks that read official registries (BODACC legal notices, labels, public contracts) cannot run.
A row does not need a SIREN to exist: your data is never filtered out or dropped for lack of a match.
Which methods does Reachable use?
Three methods, from the most direct to the most indirect: the SIREN read on the website, the VAT number read on the website, and the name and place compared with SIRENE.
- SIREN or SIRET read on the website (legal notice, terms, footer): method
WEBSITE_LEGAL_NOTICE. - French VAT number read on the website: method
WEBSITE_VAT; it contains the SIREN. - Name and address or position compared with active SIRENE establishments, with a single candidate: method
NAME_PLACE. - A name alone is refused: two companies can share the same name.
How is the confidence score computed?
Confidence depends on how many pieces of evidence agree, not on the method alone: a SIREN read on a website is not certain by itself.
Read alone, a website's SIREN was right in 48 cases out of 60: a chain store's website often shows the franchisor's SIREN, not the store's. Hence this grid:
| Evidence | Confidence | Measured |
|---|---|---|
| SIREN on the website = only name + place candidate | 1 | 21/21 |
| SIREN on the website + active establishment at the row's place | 0.95 | 23/24 |
| Only name + place candidate, nothing contrary on the website | 0.95 | 52/52 |
| SIREN on the website, row without a place (CSV of domains) | 0.8 | 48/60 |
| SIREN on the website, no establishment at the row's place | none | 4/10 |
Why is the default threshold 0.95?
Because at 0.95 the test gave 96 correct for 1 wrong and matched 56% of rows, against 12% at the “certain” threshold.
You set the threshold per list. At 1, only the strongest evidence passes and only one row in eight is matched. You can always require 1. The default moved from 1 to 0.95 on 10 November 2026, after this measure.
How was the test sample built?
200 OpenStreetMap places in the Alpes-Maritimes, drawn at random by stratum, whose real operator was established by hand for 171 of them.
Each place carries what a Google Maps row would: domain, name, address, postcode and position. Breakdown:
- independent shops: 55
- independent offices (agency, practice, real estate, insurance): 30
- tradespeople, restaurants, hotels, health: 25
- other independents: 25
- chains or networks (domain shared by several places): 40
- subdomains (store locator, directory): 25
The truth was established without seeing Reachable's output: reading the website, an active establishment in SIRENE, and the Recherche d'entreprises API for head offices outside the department. A truth was kept when at least two signals agreed; a franchisee is the operator, not the franchisor. 29 places stayed undetermined (closed place, two possible operators, dead website).
What happens below the threshold?
The row stays in your list, but columns that need a SIREN do not run and cost you nothing.
The match is shown with its method and confidence, never presented as certain when it is not. Credits are detailed on the pricing page; where each data point comes from is on the data sources page.
What are the limits of this measure?
One department, a 56% recall and one method not yet available: the measure will be repeated on a wider sample.
- One department (Alpes-Maritimes): SIRENs of head offices outside it were checked in the national SIRENE registry.
- 56% recall: 45 runs read no page at all (dead website, access refused, redirect, service error).
- A domain declared in an official source is not yet a method: no imported source provides it.
- The one wrong match at 0.95: a former operator whose establishment was still active at the address.
- A VAT number read alone was seen only once: it follows the levels of the SIREN read on the website.