A contact form that forwards everything is a contact form that stops being read.
Not all at once. It goes in stages: first you skim, then you check it every few days, then you open it when you remember. By the time somebody says out loud that nobody is reading the enquiry inbox, that has been true for months, and nobody can say what came in during them.
The mail causing it is not sophisticated. It is search-engine agency pitches, backlink offers, directory listings, and the same three paragraphs arriving from a rotating set of throwaway domains. Individually each is ignorable. Arriving daily, they are the reason a real enquiry goes unanswered.
What was arriving
The template repeat was the clearest signal. The same body, lightly reworded, from different senders across the same fortnight — enough of them that a list of the sending domains was worth keeping, and every one on it was an outfit selling the same service.
Everything else was ordinary: messages that were entirely a link, messages of four words, messages with a phone number and an instruction to move the conversation onto a messaging app.
None of it needed a model to recognize. That matters, because the temptation at this point is to reach for something clever, and clever is hard to explain to the person whose enquiry gets held by mistake.
What was built
A score, not a verdict.
Each message accumulates points across around seventeen signals — how many links it holds, whether it carries a phone number, whether it pushes toward a messaging app, the ratio of capitals and non-letters, how short it is, whether the sending domain is one already seen doing this, whether it reads as English at all. Cross a threshold and the message is held. Below it, the message forwards as normal.
Two of the signals are worth describing because they do most of the work.
The repeat check. Each message is normalized — links, addresses, digits and spacing stripped out — and hashed. If that hash has been seen in the last seven days, the score goes up. It costs one cache write and it catches the template-from-nine-domains pattern that a keyword list never would, because the wording changes and the skeleton does not.
The language check. A message that is not recognizably English on an English-language form is unusual enough to add weight, though never enough to hold a message on its own.
In front of all of it sits a bot check on the form itself, loaded lazily so it costs nothing until somebody actually starts filling the form in.
Why nothing is deleted
This is the part that made it acceptable to run.
A held message is not dropped and is not bounced. It goes into a quarantine with its score and the list of reasons that produced the score, and it stays there for sixty days. The person who submitted it sees the normal thank-you page, because telling a suspected spammer that they were caught only teaches them what to change.
An administrator has a review queue. Two actions: release or dismiss. Releasing a message forwards it to the real recipients and replays it into the normal submission stream with its original timestamp, so the record afterwards looks like the message was never held at all.
That last detail is the one worth copying. A filter you cannot audit is a filter you have to trust completely, and nobody should trust a spam filter completely. Being able to see everything it caught, and undo any of it without losing the timestamp, is what makes an aggressive threshold safe to set.
The team now reads an inbox that is mostly real. That is the entire outcome, and it was worth more than the engineering in it suggests.
What we would do differently
Two things, both uncomfortable.
The honeypot never worked. A honeypot is a hidden field that a human never fills in and an automated submitter fills in every time — free, reliable evidence. One was added to the form. The scorer checks for a honeypot too. They use different field names, so the scorer has been looking for a field the form does not create, and the honeypot has contributed exactly nothing since the day it shipped.
The reason it survived a year is the part worth learning from. There is a test covering that submission path, and it passes. It passes because it replaces the scorer with a stand-in that returns a fixed score, so it exercises the plumbing around the check and never the check itself. A green test on a mocked dependency proves the wiring, not the behavior, and it is very easy to read it as proving both.
The scores are not reproducible. The language check reaches for a system dictionary and falls back through two alternatives when it is missing. That dictionary exists on a developer machine and does not exist in the container the site runs in. The same message can therefore score differently in the two places, which makes the threshold impossible to tune with confidence — you are never sure whether you are looking at a scoring change or an environment.
Neither is dangerous. Both are the kind of thing that stays invisible precisely because the system’s overall behavior is good enough that nobody goes looking.
If your enquiry inbox has stopped being read
Ask two questions before anybody proposes a filter.
What happens to a message the filter gets wrong? If the answer is that it is deleted, or silently bounced, the filter has to be timid to be safe, and a timid filter does not solve the problem you have. If a wrong decision is visible and reversible, the filter can be set where it actually works.
Can somebody see what it caught last month? If not, you have replaced an inbox you did not read with a filter you cannot check, which is worse, because at least the unread inbox still contained the message.
If it would be easier to have somebody look at what is actually arriving, write and tell us what your form is receiving. The reply is free and a person writes it.