12 September 2026 · 5 min read

How to tell a bot from a customer in your comments

Every rule that hides spam will one day hide somebody real. What decides whether that matters is which signal you built the rule on, and most of the obvious ones are wrong.

Automatic moderation is a trade: you accept some false positives to avoid reading a hundred pieces of junk. The trade is only worth it if the false positives are rare and recoverable. Here is what we found actually separates the two, measured against real accounts.

The signal that never misfires

Latin words written in lookalike characters. A comment like "ѕєη∂ мє тнιѕ ρσѕт" is not a language. It is English disguised in Cyrillic, Greek and mathematical symbols to get past a filter.

The signals that look convincing and are not

Two of them cost us real customers before we measured them.

• Repeated identical text. Real people repeat themselves, including twice on the same post. One follower left the same run of emoji on three different posts; two others asked the identical question twice. At two repeats you hide customers.
• A comment that is only an @mention. Tagging a friend is somebody showing your post to a person who has not seen it, which is the most valuable free thing a follower does. It has no words in it, so a classifier with nowhere to file it will call it spam.

A comment that is only a hashtag is the same act with a different mechanism, and it is not spam either.

Count proportions, not strikes

A plain counter cannot tell an account that came to spam from a follower who has been commenting for months and was misread three times. We had one bot flagged twice out of the three comments it ever left, and one real follower flagged once out of five. A threshold of three would have caught neither correctly: the bot was never going to reach it, and the follower was one flag away from it.

What separates them is the ratio. Somebody whose flags outnumber everything else they have ever written to you did not slip. Somebody with a history has earned the benefit of the doubt by writing it, and however many times a classifier trips over them, it is the classifier that is wrong.

And leave the complaints alone

A complaint is a customer. Any moderation that quietly hides criticism is not protecting your brand, it is removing the one comment you could have answered. Hide spam. Keep the rest, including the parts you would rather not read.