How this list is built

Every UK lens begins with Companies House — the statutory register every UK company must file with. We sweep the industry classification (SIC) codes for a sector and take the full set of active registered companies, not a sample, not a paid lead list. Starting from the registry is what makes “every” honest: a manufacturer that never buys an advert is discovered exactly the same way as one that does.

A registry row and a website are two different facts, so we never just guess. Each company is matched to its web presence through an evidence-checked binding: the candidate site is fetched politely, and the binding must pass a named, versioned rule (a “warrant”) that weighs the company’s legal name, address, and registration details as they appear on the site itself. High-confidence bindings go live; ambiguous ones go to a human review queue rather than onto the page.

Capabilities — processes, materials, certifications, machine brands — are extracted from the manufacturer’s own site, and every extracted fact keeps the paragraph it came from. That’s why values across cmgraph open a popover showing the quoted source text and a link to the page it appeared on. Each fact also carries a confidence band, and facts we couldn’t verify are listed as such on every entity page (“24 of 44 facets not yet verified”) — absence of evidence is shown, never papered over.

A small number of facts are derived rather than extracted — for example, an aerospace quality certification implies part-marking capability. Derived values are labelled with the rule that produced them, so they are never mistaken for something a manufacturer said.

Statements like “only ISO 13485-certified CNC shop in the South West” are computed against the entire lens cohort, stamped with the scope they were computed over and the date of the scan, and decomposed into their atomic inputs — each of which can be inspected and challenged individually. When the cohort changes, claims are recomputed; superseded versions are retained for audit.

Every claim carries a dispute affordance — no account required. A plausible dispute suspends the claim publicly while it is checked (suspension is visible, never silent), and each claim’s dispute history is kept on a public audit page. Verified business owners can additionally propose corrections through the owner portal; owner edits are reviewed before they go live and are labelled as owner-provided.

A manufacturer can pay to present its own verified strengths. It cannot pay to hide a competitor’s: the systems that rank, compare, and compute differentiation claims take no payment-related input at all, so there is no mechanism — not a policy, an architecture — by which money can remove or bury a competitor’s evidence.

• No LinkedIn scraping. People-data stays out of the graph; we describe companies, not employees.
• No captcha bypass. If a site asks not to be read by machines, we respect that and find public sources instead.
• No officer or shareholder ingestion. We use Companies House for company identity, not for personal data about directors or owners.
• No pay-to-rank, no lead selling. The four things cmgraph sells (visibility of a company’s own verified facts, API access, data feeds, and analyst tooling) all leave the comparison itself untouched.

Everything above runs on a small number of sources, each used for what it is authoritative about:

• Companies House. Company identity — legal names, registration numbers, registered addresses — from the statutory register, used under the Open Government Licence.
• Manufacturers’ own websites. The primary evidence source: every capability fact is extracted from what a company publishes about itself, and links back to the page it came from.
• postcodes.io. Open postcode geocoding, behind every map pin and regional scope.
• Public directories and member lists — used only to discover that companies exist, never as evidence for capability claims. Credited as they are used; currently this includes the ADS Group member directory. Where a directory’s robots.txt doesn’t permit our named crawler, we ask the operator for permission before reading anything.