Every sales team has a handful of accounts that went unusually well. Short cycle, low friction, still a customer two years later. The obvious instinct is to find more companies like them, and the obvious method, filtering on the same industry code and size band, usually returns a list that feels nothing like the original.
That is because the thing your best customers have in common is rarely the thing the register records about them.
Why activity codes miss
A NACE-BEL code is chosen once, at registration, and updated when someone remembers to. It describes what a company registered as, not what it does now or how it behaves.
Two companies can share a code and have nothing else in common: one is a twelve-person consultancy, the other a subsidiary of a French group with four sites. Meanwhile the account that most resembles your best customer might be filed under a completely different code because it started life doing something else.
Belgium added a wrinkle in January 2025 by moving to NACE-BEL 2025. Existing codes were converted automatically from the 2008 version, and where one old code mapped to several new ones, only a single successor was written, so code-based segments need checking against both classifications. More on that in building a B2B ICP on filters that actually exist.
Codes are a useful first cut. They are a poor definition of similarity.
Four ways that work better
1. Describe the shape, not the label
Start by writing down what your best accounts actually have in common, then check which parts are filterable. Usually it is some combination of:
Size, from something filed: headcount rather than revenue in Belgium, since turnover is an optional line on roughly 96% of filings
Workforce composition: a hundred-person company that is ninety blue-collar and one that is ninety white-collar are different businesses
Structure: independent, part of a Belgian group, or a subsidiary of a foreign parent
Footprint: one site or fourteen, which is what establishment units record
Trajectory: flat headcount and a third added in three years are different companies
That list is deliberately unglamorous. It is also checkable, which the phrase "companies like ours" is not.
2. Use what the company says about itself
The fastest signal that two companies resemble each other is that they describe themselves in similar language. Websites say what a business does now; activity codes say what it registered as years ago.
Semantic search across site content reaches segments no code list contains: "companies doing AI in logistics" is not a NACE category and never will be. This is the method behind lookalike search in most tools, including ours: point at a company you already understand and let the language of its site find the ones that read like it.
3. Segment on how the workforce is organised
If you sell HR software, payroll, insurance, workwear or training, the sharpest similarity signal in Belgium is not the activity code. It is the joint committee, paritair comité, which sets sector-level collective bargaining terms for the employer's workforce.
Two companies under the same joint committee face the same pay scales, the same working-time rules and the same sector funds. For a whole class of products that predicts fit far better than what the company sells. No other register we work with exposes it as a filterable code, which is why segments built on it tend to be uncrowded.
4. Use timing to rank, not to select
Similarity tells you who to approach. It does not tell you when. Headcount growth, open vacancies and recent appointments published in the Official Gazette all carry a date, and dated events are what should decide sequence within a list, not membership of it.
A worked order
Take five to ten accounts that went well, not your largest ones. Volume of revenue and quality of fit are different things.
Write one paragraph true of all of them and false of most other companies. If you cannot, they are not a cluster and no tool will make them one.
Translate each clause into a field: size band, region, structure, workforce type. Then mark the ones that have no field. Those are the parts a rep has to establish in conversation.
Run it, then run it backwards. Apply the filter to your existing closed-won accounts. If it does not return most of them, you have described an aspiration rather than your business.
Layer timing on the survivors and work them in order of recency.
Bizzy is built so steps three and four are one query rather than four exports: filed headcount and its multi-year trend, the ownership tree with percentages and countries, establishment counts, the joint committee code, the blue- and white-collar split, radius search around a city, open vacancies and detected technology all filter over the same Belgian population. Lookalike search is the shortcut when the profile is easier to point at than to write: give it the account you already understand and it finds the companies whose online content resembles it.
What to be careful about
Your best customers may share a cause you cannot filter on. Sometimes the common factor is that all five were introduced by the same partner, or all bought during one regulatory deadline. That is real, and no similarity search will reproduce it. Notice it before you build a segment on the wrong variable.
A lookalike list is not a qualified list. It is a hypothesis about fit. Ownership, budget authority and timing still have to be established, and two of those are published.
Check the segment is big enough to be worth building. Narrow far enough and you have thirty accounts, which is a fortnight of work and then nothing. TAM, SAM and SOM with real numbers works the arithmetic through.
Frequently asked questions
How do I find companies similar to my best customers?
Describe what those accounts share in filterable terms (size from filed headcount, structure, workforce composition, footprint, trajectory), then use semantic search over company website content to catch the ones an activity code would miss.
Why don't NACE codes find similar companies?
Because a code is registered once and rarely updated, so it describes what a company signed up as rather than what it does now. Two companies sharing a code often have nothing else in common.
What is a lookalike company search?
It takes a company you already know and returns others that resemble it, usually by analysing the language of their websites rather than matching classification codes.
What is the best similarity filter in Belgium specifically?
For anything sold to employers (HR, payroll, insurance, workwear, training), the joint committee code, because it determines the terms the workforce sits under and no standard classification captures it.
How many accounts should I start from?
Five to ten that genuinely went well. Fewer and you are describing one account; more and the common factor usually dissolves.
Written by Arthur Cremers at bizzy., which builds European B2B company data from official registers.
Photo of the Klein Rusland garden district in Zelzate by Kris Vandevorst, CC BY 4.0, via Wikimedia Commons
