Every field in your file, where it came from, and how it was checked. Suppliers in this market do not usually publish this. It is the part worth reading.
The free Companies House bulk register tells us which companies exist. Each company's own website tells us how to reach it. There is no third source, no purchased database, no scraped social network, and nothing bought from another supplier.
Companies House publishes a complete snapshot of the United Kingdom register every month: the whole thing, not a list of changes. We rebuild from the latest one at least every 3 months, and the home page shows the date of the copy in use. We read that file and keep the fields the product uses:
| Field | Used for |
|---|---|
| Company number | Identifying the company |
| Registered name | Working out which website is theirs |
| Registered town and postcode | Answering "within 25 miles of here" |
| Industry code | Answering "used car dealers" or "letting agents" |
| Company type and status | Excluding dormant and non-trading companies |
Officer records are not ingested. The bulk file carries directors, and directors are people. We do not read that part into anything, so there is no way for a director's name to appear in a list built with Winnowed.
The register knows who exists. It does not know how to reach them. That comes from each company's own website, and finding the right website is the hard part.
Without it, a guessed domain that happens to resolve gets recorded as the company's website. Real examples from our own first run: johnshaw.co.uk is a parked domain for sale, not the haulier. A guess for one motor dealer redirected to an entirely different motor group. Both would have put a wrong website, and a stranger's email address, into somebody's file.
So a domain is only accepted if the company's name is recognisably in it. That refuses some real websites, and we would rather lose a lead than hand you a wrong one.
| Thrown away | Because |
|---|---|
| Named individuals' email addresses | An address that identifies a person is the one thing this tool will not put in a file. You get the company, website and telephone number instead, and the row says so |
| Addresses on parked or for-sale domains | The domain resolves; the company is not there |
| Template leftovers | info@website.com reads exactly like a working address and is not one. That reached a file once, before the check existed |
| Directory and marketplace domains | A trade directory ranks well for a company name and is never that company's own site |
| Addresses that are really image filenames | logo@2x.png matches the shape of an email address |
It identifies itself, obeys robots.txt, waits a second between requests to the same site, reads at most 5 pages, and backs off when asked. The full description, and 2 ways to stop it, are at winnowed.co.uk/bot.
How many companies produce a usable address depends on the trade. For contact by email or telephone, road freight came in at 19 in every 100, transport support at 23 and accountants at 29. Those figures were measured under our previous website check. The free count uses them for those trades and names the trade; for a trade we have not measured, it gives a range from other trades and says so. Its estimate will change as we measure under the current check. The companies that do not produce an address publish nothing, or nothing that can be confirmed as theirs.
You are not billed per contact, so a low rate in your trade costs you nothing beyond the subscription.
A supplier who will not tell you where their list is weak has either not measured it, or has and would rather you found out afterwards.
Kept, refused or withheld: each row in your file records which, and why, in words you can check yourself. A website that was kept carries what proves it is the company's, in its own column: the registered number, the full registered name or the registered postcode shown on the site. That is the part of the product that is hardest to copy, because it requires having actually done the checking.
Companies House tells us a company exists and where it is registered. It does not tell us the company's website. A website only counts when the site proves it belongs to the company: it prints the registered number, or the full registered name with Ltd or Limited, or the registered postcode. A group's shared website, printing a related company's number, counts only when it also prints the registered postcode, and a site with a trading address next to the registered office counts only when its domain also spells the whole registered name. A domain that merely looks like the name proves nothing, so it is not enough on its own.
If nothing on the site proves it, the row gets no website and no contact, rather than a guess. The row says the website was not proven to be theirs, names the possible website, and gives the reason, so you can open it and decide for yourself. A site that shows it belongs to somebody else, such as one printing another company's number, a business abroad, or one whose every address is far from the registered office, is refused, and the row says so.
3 in 500 companies were given a wrong website, and 3 of the 83 websites delivered were wrong, in our October 2026 test. The check we used before gave a wrong website to 38 of the same 500. 3 of 83 is about 1 in 28 of the websites delivered, and on a sample this size the true share could be anywhere from about 1.2 to 10.1 per cent. It is our own number, and it is not zero.
How the test was run. We drew 500 active private limited companies at random from the Companies House register, 50 from each of 10 trades we had not tested before: electricians, house builders, solicitors, estate agents, restaurants, hairdressers, information technology consultants, architects, vets and cleaners. Each company went through the previous check and this one, on the same search results and the same pages, using 2 search services. The 3 wrong websites came with Bright Data's search; with Brave's search it was 0 of 500. Every website was labelled as the company's own or somebody else's by an artificial intelligence model that did not know which check had chosen it, and 40 of the rows were then checked by hand.
The price is fewer contacts. In the same test a right website was delivered for about 16 per cent of companies, about 1 in 6, against 21 per cent under the previous check. Most of the difference is companies whose own site prints no registered number, full name or postcode.
The changes made after the test were tuned on these same rows, so these are the October 2026 test's figures for that sample, not a promise about your list. We update them after a body of new checks, not after a single run.
And we pay for the ones that are wrong. Find a row whose website is not that company's, tell us which within 30 days of the run, and we credit that company back to your allowance. Each row can be credited once, as allowance, never as cash. That is section 6 of the terms, not a goodwill gesture. A supplier who is confident enough to publish its own error rate should be confident enough to buy the errors back.
The home page says 40 per cent of an unverified list was dead or belonged to a different company. Here is exactly what that was.
We took 120 companies and built the list the ordinary way: guess the domain from the company name, accept whatever answers. Then we fetched every one of those addresses and compared the page that came back against the company it was filed under. 48 of the 120 either did not resolve at all or resolved to a business that was not the one named. AJC CARS LTD had been paired with coxinc.com, which is Cox Enterprises, an American media corporation. INA GROUP LTD had been paired with schaeffler.com, a German bearing manufacturer.
It is not an audit of any named competitor's product, and we do not claim it is. It is the method a list is built by when nobody checks, run on real companies, with every result kept. The comparison it supports is between checking and not checking, which is the only comparison we are entitled to make, because it is the only one we ran.
Winnowed