Experian feeds Veridion's 642-million-company graph into its credit models

Somewhere in a lending team, an analyst will open a file on a two-person plumbing firm that had no file last quarter, and will have to say where the trading status came from and when anyone last checked it.

The pitch and the reference customer describe different products. Stefan Gergely, Veridion's head of growth, said in the announcement: "Information updated quarterly or annually is no longer intelligence. It's history." ^[5] Refresh speed matters for a company already in your file. Experian's director of data strategy and innovation, Jon Roughley, said the signals let the bureau capture "risk that traditional credit data simply couldn't see" ^[7] and score businesses that had been invisible to its clients ^[8].

Veridion described more than 80 million weekly-updated company profiles when it raised $6m in 2023 under the name Soleadify ^[4]; the boilerplate now says 642 million companies across 249 countries with 461 attributes each ^[3]. That is 562 million additions in three years, about 15 million a month ^[22]. The count has moved faster than any definition of it. The company says it resolves each company into legal registrations, operating locations and corporate hierarchy ^[9], and those are three separately countable things. TNW reported it could not tell whether 642 million counts firms, entities, sites or records ^[10].

A graph built from websites, registries, filings, product catalogues and social profiles meets the same business more than once, and deduplication is the hard engineering problem in this category ^[11]. That gap lands on the buyer's side. Whoever runs the credit model inherits those merge decisions. Commonly cited estimates put the world's businesses in the low hundreds of millions, with formally registered companies well below that, and TNW said it could not trace those estimates to a single authoritative count ^[15].

Veridion says every data point traces back to its source ^[12]. Multiply the boilerplate out and that is a provenance commitment across roughly 296 billion data points ^[23]. Provenance and verification are two different things: the source of a fact is one question, its truth another ^[13]. TNW put the trade plainly, saying registries and filings are slow because they are authoritative and web signals are fast because they are not ^[14].

The entities missing from conventional credit files are the smallest ones, and a large share of those are sole traders ^[16], who are natural persons in law. The EU AI Act lists systems evaluating the creditworthiness of natural persons among its high-risk categories ^[17]. That is where the compliance question sits, at the thin end of the market. Brussels agreed this year to thin the Act out, moving standalone high-risk obligations to 2 December 2027 ^[18]. The Commission can inspect models, restrict market access and fine 3% of global turnover ^[19], and the data broker industry already spends heavily lobbying against restrictions on what it collects and sells ^[20]. The release itself opens on the Strait of Hormuz, described as disruption and a single geopolitical event ^[21].

The test that works runs per field. For every attribute the model acts on, that means a named source, the date it was last checked, and a rule for what happens when a web signal contradicts a registry. The subjects the vendor adds then split into two columns: the ones a register can find, and the ones inferred from a web presence. If the lift comes from the second column, the buyer has bought coverage of natural persons, and the score arrives with the compliance question attached.