Before a message reaches the AI - and before it costs you a credit - it goes through free filters. In v2 there are three of them: keywords, stop words, and minimum length.

Order of checks

1. Stop words - if there is at least one stop word → the message is discarded right away, for free.
2. Minimum words in a message - messages that are too short ("+1", "thanks", "up") are not sent for analysis and cost nothing. Configured separately for each channel on the "Where to look" step.
3. Keywords - by default the AI looks only at messages that contain your words; the rest are skipped for free.
4. AI check - whatever passes the filters is read by the AI: 1 credit per message, regardless of the verdict. Without AI (an empty micro-prompt) keyword matches become "raw" signals at 0 credits.

"What to check for leads" - two modes

The toggle on the "What to look for" step:

• "Only messages with my words" (default) - the AI looks only at messages that contain your keywords; the rest are skipped at no cost. The wizard names the downside honestly: if someone describes their task in different words, we will not see them. How to broaden your words depends on the channel - see below.
• "Every message" - the AI reads every message in the selected chats, even without your words. It finds more, but each message read costs 1 credit, even if it is not a lead.

💡 Tip: start with "Only messages with my words". Turn on "Every message" in narrow, topic-specific chats where almost every message is potentially yours.

Keywords: the rules are DIFFERENT for different channels

Telegram: words are matched inside the message text - case-insensitively and as part of other words ("repair" also matches "repairs"). Intent phrases work well here too: "any recommendations", "who does", colloquial variants.

Threads, Reddit, LinkedIn, X (Twitter), Bluesky, Hacker News: the word goes into the platform's own SEARCH - and a phrase almost never finds anything there. Measured on live campaigns (10.08.2026): five phrases returned 61 posts and ZERO leads; five single nouns returned 50 posts and three leads. The rule: one unambiguous word per key, with inflected forms as separate keys ("repair", "repairs", "repairing") rather than as a phrase. The wizard hints at this right on the keywords step depending on the channels you selected, and the weekly review in the bot names the specific dead word.

The wizard also suggests recommended words - look through the list and remove what you do not need. Several words separated by spaces are ONE phrase, not a list: separate them with Enter or a comma.

Stop words

The same matching mechanism, but in reverse: one hit and the message is discarded for free. Short roots also catch derived words - be careful with them.

What to add: irrelevant context ("looking for a job", "rent" - if you are about sales), competitor names, the typical noise of your chats.

Where the "scenario" went

In v1 there was a "seekers / sellers / both" preset. In v2 its role is played by the analysis micro-prompt - one or two sentences about who counts as a lead ("looking for a contractor for a renovation", "offering media buying services"). This is more precise than a preset: you tell the AI exactly what you meant. The article "The analysis AI prompt and the Improve button" is right next to this one in the same section.