Here's a scenario that breaks naive plagiarism checkers: you're checking a blog post you wrote, and the top "match" is... your own blog, which republished it last week. Or you're checking a paper that quotes a source you've already cited properly, and the checker keeps waving its arms about the quote you handled correctly.
The tool is technically right and practically useless. You need a way to say "yes, I know, that one's fine."
Rules, not just a mute button
Some checkers let you dismiss a match after the fact. That's fine, but it's reactive — you do it every single time. What you actually want is a rule: this source, always, don't count it. Or the inverse: only check against these sources and nothing else.
So Silvertext has source rules, and they work the way anyone who's edited a robots.txt file will find familiar.
Wildcards, because domains are messy
You can exclude or include by exact hostname, by root domain, by URL prefix — or by pattern, with wildcards. *.example.com/essays/* excludes an entire section of a site across all its subdomains. sci-hub.* catches every TLD a site hops between. *.pdf$ (the trailing $ anchors the end) targets a file type.
If you add any include rules, only matching sources count and everything else is ignored — useful when you're checking against a specific corpus. Exclude rules subtract from whatever's left. It's the same mental model as robots.txt, which is to say: boring, predictable, and exactly what you want from something that decides what counts as evidence.
Deterministic, not a suggestion
The important part: these rules are enforced in code, before the report is built — not passed to a model as a polite request. When you exclude a domain, sources from that domain are dropped, full stop, and the report tells you how many were removed by which rule. A rule you set is a guarantee, not a hint. That's the whole point of giving you the control in the first place.