Why email filters miss what is inside the attachment
This is not a complaint about Gmail. It is a description of what envelope-level matching can and cannot decide — including the version of it we shipped and had to remove.
Why can't email filters sort your attachments?
Because a filter reads the envelope, not the contents: it knows the sender, the subject line and whether a file is attached, but it never opens the PDF. It can tell you a document arrived; it cannot tell you whether it is an invoice, a delivery note or a signed contract.
The envelope is not the document
Every mail filter you have used matches on metadata: sender, recipient, subject, size, whether an attachment exists, sometimes its filename. That is enough to move mail into folders and it is why filters have barely changed in twenty years — for routing conversations, the envelope is the right unit.
Filing documents is a different job. The thing you want to route on lives inside a PDF the filter never opens, so any accuracy you get is borrowed from a correlation — that this sender only ever sends one kind of document — and it lasts exactly as long as the correlation does.
Why a near-miss is worse than a visible failure
A filter that fails loudly is cheap: you notice, you fix the filter. A filter that files the wrong document into the right-looking folder is expensive, because everything looks fine. You find out at the point where you needed the folder to be correct.
That asymmetry is the whole argument for reading the document. Not that it is never wrong, but that when it is unsure it says so and leaves the file alone, rather than confidently matching a substring.
The same intent, matched two ways
On the left, the rule name that caused the misfiling. On the right, the same intent expressed as something a reader can act on — which is all a rule is now.
“electricity invoices”
As a keyword this is a trap: any document tagged “Invoice” satisfies it, because "electricity invoices".includes("invoice") is true.
“Electricity bills for the shop, from any energy supplier. Not gas, not water, not the ones for my flat.”
A sentence a reader can apply to a document. Nothing in it depends on a substring appearing anywhere.
What the misfiling actually looked like
This is the file that ended up in the wrong place, and it is worth looking at because nothing about it is malformed:
Invoice_Make-com_2026-02-11.pdf
It was correctly identified as an invoice, correctly named, and written to Expenses Utilities Electricity — a folder for electricity bills. Every step worked. The step that did not exist was anyone asking whether a Make.com invoice is an electricity bill.
The filename above is a real mistake, not an illustration. Findest's first rule engine compared strings, so a rule named “electricity invoices” matched every document the classifier had tagged “Invoice” — and a Make.com software invoice was filed into the electricity folder. It was not an edge case; it was what the algorithm did by design. Migrations 0018 and 0019 removed keyword matching entirely.
Questions about filters and classification
Can I still filter by sender if I want to?
Not as a rule condition — those fields were deleted, deliberately, because they were the mechanism that produced the bug above. You can restrict a rule to one mailbox, which covers the common case of keeping work and personal mail apart.
How does it decide between two rules that both fit?
All of your rules are sent to the model in a single request, so it compares them against each other rather than judging each one alone. Asked one at a time, a model says yes to the first rule that looks close; shown the whole set, it can discriminate.
What happens when nothing matches?
The document is left unfiled and the History shows the reason in plain language. It is not archived to a fallback folder — a wrong folder is worse than no folder, because you stop trusting the ones that are right.
Related questions
Every invoice filed, without opening the mail
The category with the most senders, the least consistent subject lines, and the highest cost when one goes missing. Read it →
When a Zap is the wrong shape for filing attachments
Zapier is a better product than Findest at almost everything. This page is about the one job where being narrow wins. Read it →
Save Gmail attachments to Google Drive, automatically
Gmail routes envelopes. It has never been able to route what is inside the attachment — which is the part you actually want filed. Read it →
Write a rule the way you would explain it.
No keyword fields, no condition builder. One sentence per rule, and two rules on the free plan.