Methodology

Every recurring number on this site is computed from a frozen set of definitions, held in one file and version controlled. This page is that file, rendered. It is currently at version 4.

Definitions are frozen because a rate that moves when the definition moves is worse than no rate at all: a reader cannot tell the two apart. Changing one bumps the version, republishes every affected series, and adds a changelog line. A build check fails if a definition changes without that.

The Three Populations

Every report names the population it counts. They are not interchangeable, and most errors in startup statistics come from treating them as though they were.

All Latvian limited companies

REGISTER_BASELINE

Every SIA and AS in the UR bulk register. No industry filter, because none is possible. Unbiased with respect to survival, since a company that died in 2018 is still a row. Backfillable to 2015 and reproducible by anyone from one CC0 file.

Never used for: Any claim about startups specifically. This population contains holding companies, shops and haulage firms, and says so.

Organisations published on rigastartups.com

Directory companies witnessed nightly since 2026-09-06

OBSERVED

The snapshot store. The only population here whose deaths are observed rather than inferred, and therefore the only one that can ever carry an honest startup survival curve. Needs roughly two years of nights before it says anything.

Never used for: Anything before 2026-09-06. There is no history and none can be reconstructed.

Why There Is No Tech Filter

The obvious way to build a startup statistic is to filter the company register down to technology industry codes. That is not possible in Latvia, and saying why is more useful than pretending otherwise.

Latvia publishes no industry classification per company as open data. The register carries no industry field. The register’s area of activity file is free text copied from articles of association and covers only legal forms that predate the 2010 Commercial Law: it matched none of our 634 registered companies. The State Revenue Service publishes taxpayer ratings and VAT groups. The Central Statistical Bureau publishes the classification tree but keeps the assignments inside its statistical register, under statistical confidentiality.

A classification feed would not have fixed it either. Every signal that marks a company as a technology company is a signal that dies with the company: a website, a job board, a product, a listing in a directory. What the register keeps after a company is struck off is a name, a legal form, some dates and an address. We tested the one attribute that survives, running a deliberately generous keyword match over company names, and it found 24% of the startups we already know about, favouring the ones that named themselves after their industry.

So a technology filter fit for survival analysis is not merely missing, it cannot be built from public sources. The populations above split the work instead.

Why the Directory Cannot Measure Survival

The 634 organisations listed here were found through startup directories, member lists and editorial research. Every one of those sources lists companies that exist now. A company founded in 2016 that closed in 2018 was never in any of them, so it could never enter this list.

The effect is measurable. Grouped by the year they were founded, the companies in this directory survive at between 96% and 100% in every cohort back to 2015, flat. Every limited company in Latvia, over the same cohorts, survives at between 43% and 94%, decaying with age exactly as it should. A survival curve that does not fall as cohorts get older is not measuring survival, it is measuring how the list was built.

This is why the nightly observation store exists. Since 6 September 2026 every listed company and every open role is recorded each night, so from now on a closure is something we witnessed rather than something we inferred from a list that never contained the dead.

Definitions

Term	Definition	Note
Organisation	A record on the map with a UR registration number.	Excludes unverified submissions.
Formation	The first UR registration date of the entity.	Not the date the founders started working.
Cohort	All companies with a formation date in a given calendar year.
Active	Not marked terminated or liquidated in UR as of the snapshot date.	Separate from trading. A dormant company is active.
Terminated	UR status is terminated or liquidated.	A register fact, not an editorial claim.
Closed	Editorial judgment that the company no longer operates.	Only asserted with a named source. Never derived from UR alone.
Open role	A role present in the nightly ATS pull on the snapshot date.	Excludes LinkedIn only and hand built career pages.
Startup	Not defined as a computable filter, and deliberately so. Membership of the directory is an editorial decision, published per company and open to correction.	The reports plan proposed a NACE based definition. It cannot be built from public Latvian data, and no classifier applied to dead companies could be honest. Rather than dress a guess as a definition, reports name their population instead.

Rules and Thresholds

Baseline legal forms. SIA and AS. Sole traders and farms follow different economics and would swamp the cohorts.
First cohort. 2015. Earlier years are in the register but the address and capital data thin out behind them.
Minimum cell size. 20. A cut with fewer than that is suppressed rather than published as noise.
Job boards read. greenhouse, lever, workable, teamtailor, recruitee, ashby, personio, smartrecruiters, bamboohr. A role that exists only on LinkedIn or a hand built careers page is not counted, because it cannot be counted the same way twice.
Boards that can state pay. greenhouse, lever, teamtailor, recruitee, ashby. The rest have no salary field at all, so a role on one of them cannot state pay however willing the employer is. Any disclosure rate we publish uses these as its denominator, or it would report a missing feature as an employer’s refusal.

Sources

Uzņēmumu reģistrs open data. The full company register, republished daily under CC0. Current dump 27 September 2026, covering registrations through 25 September 2026.
Register capital data. Registered and paid up capital, CC0. Published only for companies that still exist, which is why capital figures are suppressed for older cohorts.
Annual reports. Employee counts and profit and loss figures as filed with the State Revenue Service, CC0. Personnel costs appear only in the minority of accounts prepared by nature. They lag by about a year, and they stop when a company becomes regulated and reports to Latvijas Banka instead.
State Land Service address register. Coordinates for every Latvian address, which is how a registered address becomes a neighbourhood.
Eurostat migration statistics. First permits by reason and citizenship, Blue Card decisions by occupation and rejected Blue Card applications, one definition across every member state. Used wherever Latvia is compared with another country.
PMLP residence permit statistics. First-time permits by reason and citizenship, published by the migration authority each half-year. Used only for what Eurostat does not publish, and never mixed with it in one figure.
Uzņēmumu reģistrs members. The shareholders of every live SIA, with the date each acquired their shares. Used to identify founding shareholders. Names and partial identity codes are read only to recognise the same person across companies, and are never stored or published.
Liquidation and insolvency records. Open liquidation processes and insolvency proceedings from the Register, CC0. Used to tell companies winding down from companies that simply stopped filing.
PMLP personal names. How many women and men in the population register hold each given name. Used only to infer sex from a founder’s or owner’s given name, and never joined to a person.
Eurostat structural business statistics. Net turnover for Latvia’s business economy. Used only as a check on the size of register totals, never as a comparison.
Archived register files. The Register’s capital, liquidation and suspension files as published in April and May 2025, and its shareholder file from July 2024, as kept by the Internet Archive. Compared with today’s files because the Register keeps no history of them: capital changes, and the stated reason a company was being wound up. Shareholder names are compared and never stored.
Company names and name history. Current names from the register and earlier names from the Register’s name history, CC0. Read with LibreOffice’s Latvian and English spelling dictionaries and word counts from film subtitles (FrequencyWords, CC BY-SA 4.0), to tell a Latvian name from an English one. Checked against names labelled by hand.
Reorganisations and suspensions. Mergers and divisions, and suspensions of economic activity ordered by the tax authority, from the Register, CC0. Used to tell a company that merged from one that failed, and a company being removed from one that is merely quiet.
Announced funding rounds. Equity rounds into Latvian startups, collected by hand from Labs of Latvia, Invest in Latvia, EU-Startups and the announcements themselves, each kept with its link and converted to euros at the European Central Bank reference rate. Used only for the funding round tracker, and checked against filed accounts rather than taken on trust.
Eurostat Labour Force Survey. Employers and the self-employed by sex. Shown as EU context beside register figures, never compared with them.
Registrų centras open data. Lithuania’s register of legal entities, registered and deregistered, downloaded by hand because the site refuses scripted requests. Used only to compare Lithuania with Latvia.
World Bank Entrepreneurship Database. New limited liability companies per year, drawn from each country’s register. Used for Estonia, whose open register lists only companies that still exist, and as a check on the other two.
Eurostat business demography. Enterprise births and three-year survival on one definition across member states, covering every legal form.
Company job boards. Read nightly from each company’s own public feed, the same one their careers page uses.

All of those are open data, and so is everything computed from them here. Every figure on this site is downloadable as CSV, with the source, the dump date and the definitions version in the file rather than on the page that linked it, so a number keeps its provenance after it leaves us. The data page lists what is published, what is available under licence, and which parts of the register you can simply fetch yourself without asking.

What Has Changed

Every version of these definitions, newest first, and what moved when it changed.

Version 4
9 September 2026
Added BambooHR to the machine readable boards. It is the ninth board we can read and the first found by crawling company careers pages rather than by guessing an account name. It carries no salary field, so it is deliberately absent from the boards that can state pay, and adding it lowers the share of roles on a pay capable board without any employer having changed what they disclose.
Version 3
9 September 2026
Split the directory into the companies being measured and the support organisations around them. Reports over the directory now count startups and scaleups only, where they previously counted every published record: 22 of 617 were funds, universities, accelerators, service providers or hubs. Retained earnings, Startup Law take-up and the directory section of foreign ownership all move slightly as a result, and the change is being made before co-working spaces are added rather than after.
Version 2
7 September 2026
Added a separate minimum cell size for published rates, set at 200. The existing minimum of 20 is adequate for suppressing a count and far too small for a percentage: municipality survival varied by 31.7 points below n=200 and by 3.4 points above n=1,000, which is cell size rather than geography. No previously published number changes, because no published report used the old threshold for a rate.
Version 1
7 September 2026
First frozen set. Replaces the proposed NACE based tech set, which cannot be built from Latvian public data, with three named populations.

Corrections

If a number here is wrong, we want to know and we will say so in public. Corrections are noted on the affected report rather than applied silently, because a statistic that changes without explanation is indistinguishable from one that was never right. Write to vinayak@sageo.ai.