Desk research vs direct partner interviews in France

The Sirene register, published by INSEE as open data under the Licence Ouverte and refreshed monthly, lists every registered company and each of its establishments with its identifiers, registered address, declared activity code and recorded status. Filter it by activity and region and you have a working frame.

Be precise about what that frame is. It establishes that a legal entity or establishment is registered as active. It does not prove present commercial activity, an operating team, capacity, interest, or any capability as a partner. A registered company can be dormant, and a company doing serious work can carry an activity code it chose years ago.

What the register does settle is the administrative layer, and it settles it quickly: which companies exist on paper, under which declared activity, in which size band.

Filing accounts is not the same as publishing them. Article L232-25 of the Code de commerce lets a micro-company declare that the accounts it files will not be made public at all, and lets a small company request that its profit and loss account is not made public. Companies belonging to a group cannot use the option.

The thresholds decide who that covers. Since 1 March 2024 a company counts as small if it does not exceed two of: 7.5 million euros of balance sheet total, 15 million euros of net turnover, and 50 employees; micro if it stays under two of 450,000 euros, 900,000 euros and ten employees.

Now set that against the kind of organisation a foreign software vendor often wants as a first French partner. A specialised integrator, a regional reseller or a consultancy with a practice in your field can sit comfortably inside the small band, in which case the balance sheet may be visible while turnover and margin are not.

The two figures you most wanted can therefore be the two you cannot get, and no amount of further searching produces them. That is not a gap in your research. It is a right the company has exercised.

Six questions, and what each method settles. The fourth column is the one worth reading twice: it is what survives both.

Open data invites a particular kind of overconfidence, because everything in it is true and none of it means quite what it appears to.

The APE or NAF code records the main activity a company declared when it registered. It is a classification, not a profile, and it does not narrow as a firm specialises. A single software-publishing code can cover a founder working alone and a large services group. Use it to build a frame; do not use it to conclude that two companies do similar work.

Where the profit and loss account has been withheld, what remains shows a position without a direction. A firm that is growing and a firm that is being wound down gently can present similar balance sheets for a year or two. If you cannot see the direction, do not infer one.

Logos, certification badges and case studies can stay on a website long after the relationship behind them ends. A vendor’s French partner page tells you a relationship was signed at some point, which is a weaker claim than it looks, and signature is the cheapest thing in the channel.

Certifications are the exception, because some French schemes publish their holders. If a firm claims HDS certification, the Agence du Numérique en Santé publishes the list of certified hosts. Where a register exists, check the register rather than the logo.

The case for asking people directly is not that conversations are reliable. It is that they reach questions nothing else reaches. It would be dishonest to stop the sentence there.

• Small samples stay small. A handful of conversations is a signal about a handful of organisations. Reported without its denominator, it reads as a finding about a market.
• You interview whoever replies. Non-response is not random, and in France you cannot easily test how unrepresentative your responders are, because the revenue figures that would let you compare them are exactly the ones a small company may withhold. Recording who did not answer is the partial defence that remains.
• Stated interest costs the speaker nothing. An interview measures what someone is willing to say about a hypothetical. That is genuinely useful, and it is not a commitment.
• A specific objection is harder to invent than warmth. That is a reason to weight objections more heavily than enthusiasm. It is a working rule for reading answers, not a measured effect.

The rule is short. Use public sources to eliminate. Use conversations to decide.

Public data is good at showing that a company is not what you hoped: wrong size band, wrong declared activity, no longer registered as active, part of a group that already owns a competing product. It is poor at showing that a company is what you hoped. Run the elimination first, because an hour of conversation spent on a firm a register lookup would have removed is an hour not spent on a real candidate.

Then write down the denominator. Whichever method produced your conclusion, it is worth something only if a reader can see how many organisations were in the frame, how many were approached and how many replied. Those three numbers sit on the page that states the conclusion in the sample report.

This is the page where our own work is easiest to explain, which is a reason to be careful rather than a reason to sell. We run the elimination stage from public sources because it is cheap and fast in France, then spend the time where nothing else reaches: asking named organisations what they think, and recording who did not answer. That method does not remove the fourth column of the table above, and it is not meant to.