> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thedatacity.com/llms.txt
> Use this file to discover all available pages before exploring further.

# How We Match Third-Party Data

> How The Data City links external datasets, like investment, jobs, and grants data, to the right company on Companies House.

Companies House is our single source of truth for UK company data. Every other dataset we bring in, from [investment data](/our-data/third-party-data/investment-data) to [Lightcast jobs data](/our-data/third-party-data/lightcast-data) to [B Corp certifications](/our-data/third-party-data/b-corps), is written and stored by someone else, in their own format, without a Companies House number attached. Before we can show that data on a company's page, we first have to work out which Companies House company each record actually belongs to.

We call this process **matching**.

## Why matching is hard

Companies don't always describe themselves the same way twice. A single business might appear as "Acme Ltd", "Acme Limited", or "Acme (UK)" across different sources, with different addresses, or under a trading name rather than its registered name. Some sources don't include a company number at all.

So instead of a simple lookup, we compare each incoming record against Companies House using several pieces of evidence together, things like company name, website domain, registered address, and social media handles, and work out the most likely match.

## How we do it

We use a well-established statistical matching technique (built on an open-source tool called [Splink](https://moj-analytical-services.github.io/splink/)) to compare records at scale. In short:

1. **Narrow the field.** For each incoming record, we first shortlist a small set of plausible Companies House candidates, rather than comparing it against all 5+ million companies.
2. **Score the evidence.** We compare each candidate against the incoming record across multiple signals (name, website, address, and so on) and combine them into an overall confidence score.
3. **Pick the best match.** If a candidate's score clears our accuracy threshold, we accept it as a match. If nothing scores highly enough, the record is left unmatched rather than guessed at.
4. **Resolve corporate groups.** Many companies belong to a wider corporate group. Where that matters, for example deciding which entity an investment round or acquisition should be attributed to, we apply additional rules to roll matches up (or keep them distinct) in a way that reflects how the business actually operates, rather than just its group structure on paper.

Each data source has its own quirks, so the exact combination of signals we use varies a little from one to the next. You can find source-specific notes on the relevant [third-party data pages](/our-data/third-party-data/companies-house).

## Keeping matches fresh

Matching isn't a one-off exercise. As new companies are incorporated, existing companies change their details, and new records arrive from our data providers, we re-run matching regularly to keep everything up to date. For some sources, this happens daily.

## What this means for accuracy

No automated matching process is perfect. We aim for a high degree of accuracy, and we'd rather leave a record unmatched than force a low-confidence guess. That said, a small number of matches can still be inaccurate, for instance where a company has registered multiple similar entities, or where the source data itself is incomplete.