How to Tell Real Viewers from Chat Bots by Their Behavior in Chat

Why chat is more informative than counters

Views and subscriptions are the cheapest to inflate: counters only need presence. Chat requires a reaction to the broadcast, so behavior in chat provides more signals. But a single sign proves nothing — it's the combination and the dynamics across several streams that work.

Templated messages: where suspicion begins

The crude level is repeated text from dozens of accounts: "hello everyone," "great stream," greetings with no connection to what's happening. Such replies don't respond to other viewers and don't develop the topic.

The threshold for assessment is not an absolute number but a share of the overall flow. If over a broadcast more than roughly a fifth of messages come from a short set of repeated phrases and they come from different accounts — that's a reason to look into it. A few isolated "hellos" from newcomers mean nothing: real people who drop in for the first time write that too.

Generation from template libraries looks more varied, and here it's not the text that gives it away but its relationship to the broadcast. The message appears in a vacuum: no one asked a question before it, and no one replies after. A live chat is chains where replies hook onto each other.

Synchronicity and rhythm

People write unevenly: a burst at an interesting moment, silence during a drawn-out segment. Bots operate on a schedule — in a steady flow or in batches.

• A steady flow of replies when nothing is happening in the broadcast.
• Messages from several accounts seconds apart.
• Reactions to events with a constant delay.
• Equally enthusiastic responses to everything, with no dips and no criticism.

Reaction to context and moderation

A live viewer notices a change of topic, a joke, a mistake, a reply to someone. Their rare messages are tied to what was in the broadcast a minute ago. A bot responds to keywords: mention a word in your speech — and if templated replies with it from unfamiliar accounts immediately appear in chat, take a closer look.

Moderation also provides a signal: real people react to a moderator's actions, bots don't. A message was deleted, yet the "viewer" carries on as if nothing happened — that's telling.

Be careful with the "ask an open question" test. It filters out primitive bots, but bots built on language models pass it: they'll answer coherently and on topic. Treat it as a low-level filter, not proof.

What the profile shows

Behavior in chat should be looked at together with the account: a recent registration, an empty profile, no avatar, subscriptions to hundreds of channels with no activity. On its own it's not proof — a silent real viewer looks similar. The profile refines the picture but doesn't replace chat analysis.

An example of an inflation pattern

A typical picture: after buying promotion, accounts appear on the channel that join at the start of the broadcast and leave at the end, post one phrase from a short set at intervals of a few minutes, don't react to a change of topic or to a moderator's actions, and don't show up at all on the next stream or show up with the same set of phrases. Individually each signal is weak; together they form a consistent picture.

How to collect data

You can't keep track of that much by eye over a single broadcast. Export chat history: on YouTube this is available via the API and third-party chat archives, on Twitch — via IRC clients and chat logging services, on VK Play and Kick — depending on the platform and the bots used. Then calculate in a spreadsheet: the share of repeats, the number of messages per account, the intervals between replies, return rates between streams.

There are few ready-made detectors you can trust without verification: chat analytics services and anti-fraud modules from some platforms and moderator bots give their own estimates, but their criteria are closed, and they err in both directions. Use them as a second opinion, not as a verdict.

Look not at a single broadcast but at a slice over several weeks: who comes back, whose messages change from stream to stream, who remembers past topics. Regular participants with a history are the most reliable sign of a live audience. And remember the weak spot of any method: a real viewer can also be templated, and a quality bot can imitate context.