The Sourcing.sh blog

200,000 companies: how we count

Building in public: the methodology behind our announced volumes — what counts as a record, what we deduplicate, what we exclude, and why we publish these rules.

Titouan Albouy4 min read
Building in publicMethodology

All data providers list volumes — “500 million profiles,” “50 million companies” — and almost none define what they count. A number without a counting method is not information, it is a slogan. sourcing.sh advertises around 200,000 companies, 123,000 profiles, 98,000 schools and 1.4 million job offers. Here, in a building-in-public logic, is how these figures are produced — including what they exclude.

What counts as a business profile

A company enters the account if three conditions are met:

  • A strong identifier : SIREN for France, or failing that an active web domain that can be unambiguously linked to the entity.
  • A threshold of completeness : at least the name, a strong identifier and two context fields (sector, workforce, location or activity).
  • A signal of life : at least one trace of activity observed over the last 24 months — publication of offer, site update, legal event.

A brand is not a record; a subsidiary with its own SIREN and its own activity is one of them. A group with ten subsidiaries can therefore count as eleven — we prefer to document this choice rather than smooth it out.

Deduplication, above all else

Our agents aggregate several partner sources; the same company therefore arrives several times, under slightly different names. The pipeline merges first on strong keys (SIREN, domain), then on weak keys (standardized name + location) with arbitration. Order of magnitude: the gross stock exceeds 260,000 files; about a quarter merge or disappear to result in the approximately 200,000 published. A volume announced before deduplication is mechanically inflated by 20 to 40% — this is the first question to ask any supplier.

What the numbers exclude

  • Companies : deregistered companies, holding companies without detectable activity, files below the threshold of completeness.
  • Profiles (~123,000) : only profiles from partner sources, with a news signal of less than 24 months; a profile without any observable updates exits the account.
  • Schools (~98,000) : separate establishments and campuses, initial and continuing education combined — the figure includes places of education, not diplomas.
  • Job offers (~1.4 million) : active offers seen on a sliding window, deduplicated between aggregators — the same offer published on four job boards counts as one. Expired offers are taken out of the figure, even if we keep them in history.

Why “about”

These figures change every day: agents add, merge, expire. Publishing a single meter would be wrong the next day; publishing a dated rounding is honest. We update the orders of magnitude when the variation exceeds a few percent, not before.

Why publish the method

Three reasons. First, comparability: without a common definition, comparing two suppliers on volume makes no sense, and we prefer to be compared on our rules. Then discipline: a published method prohibits us from discreetly inflating the account by releasing a threshold. Finally, consistency with our thesis: if SaaS applications become interfaces placed on databases, then the basis — its definition, its deduplication, its freshness — is the product, and a product must be auditable. This is what we are building since Station F, in public, method included. The next numbers we publish will come with their counting rules — these, or better ones.

By Titouan Albouy

All articles

Your next step

Turn data into action.

Find the right contacts, keep your data up to date and connect your tools.