Where every address comes from

The claim this product rests on is narrow and checkable: every discovered address keeps the page it was published on. Not a data supplier, not an aggregate of several, not “our proprietary database” — a URL you can open. This page is how that happens and what it does not cover.

There is no index-size figure anywhere on this page or on any other, deliberately. Index size is the number this category competes on and it is the number least connected to whether any individual answer is right.

1. 1

   Resolve the origin

   A domain is resolved to the site that actually serves it, so a redirect chain or a parked domain does not produce a crawl of the wrong thing.

2. 2

   Read robots.txt first

   Before any page is fetched. A disallowed path is not fetched, and the crawler identifies itself by name so a site owner can see it in their logs and write a rule for it. The rules it binds itself to are published at /bot.

3. 3

   Plan the pages

   Shallow and targeted rather than exhaustive. Candidates come from the site’s own links and sitemap, then from well-known paths — /team, /about, /contact, /impressum, /kontakt — then from anchor text. Blind path probing is capped at eight: a crawler that guesses a thousand URLs on someone else’s server is a load problem, not a strategy.

4. 4

   Extract

   Four methods, strongest first: mailto hrefs and Cloudflare data-cfemail read off the parsed DOM rather than by regular expression, then textual obfuscation (bob [at] acme [dot] com), then plain text. Asset filenames, documentation placeholders and error-tracking keys are excluded, because each of those produces something that looks exactly like an address and is not one.

5. 5

   Attribute a name

   Also strongest first, and the ordering is reported rather than hidden: a schema.org Person in JSON-LD or microdata, then proximity in the DOM, then the local part — and the local part only where the separator is unambiguous. jsmith is not trusted. No competitor distinguishes "the publisher told us in markup" from "we split the local part and guessed".

6. 6

   Score, with the factors attached

   Confidence is 0–100 with a band, and every score carries the factors that produced it: a code, a label, whether the effect was positive, negative or a ceiling, and its weight. A score is a disclosure control, not a summary — an opaque number is what lets a product ship confident nonsense.

7. 7

   Index with the citation

   Every stored address carries where it came from: the URL, the domain, which of the four extraction methods found it, the date it was first extracted, the date it was last seen there, and whether it is still on that page today. That last flag is the difference between a contact you can explain and one you cannot.

8. 8

   Infer the address pattern

   From addresses whose name is known, the domain’s format is inferred — {first}.{last} and the rest — with confidence from agreement multiplied by sample size, and an ambiguous sample splitting its vote. Transliteration runs both ways, so muller and mueller are the same person. The pattern finder shows the samples it used rather than only the answer.

Erasure that prevents re-collection

What we do not hold, and will not

• —Personal consumer addresses. The index holds business addresses companies publish about themselves. Somebody’s private Gmail is a different activity with different ethics and it is not one we are in.
• —Purchased or scraped lists. Nothing in the index arrived from a data broker, and the seeded company set is not drawn from one either.
• —Data from platforms whose terms forbid collecting it. Exporting contacts you are already connected to is a feature those platforms offer. Scraping profiles you are not is not, and we would not publish a method we would not use.
• —Spam-trap identification. We have none and claim none. A trap address is indistinguishable from a real mailbox at the protocol level, and a product claiming certainty there is selling a guess with a label on it.
• —Inbox placement and blocklist monitoring. Different products. The free tools read your DNS records and look up a host on request; they do not watch anything for you.