How to Filter Out Irrelevant Telegram Leads: Stop Words and Filters

How to remove noise from Telegram leads and lists: stop words in monitoring, keyword wildcards, account filters and qualifying leads in conversation.

You set up monitoring for the word "renovation", and 140 alerts arrived on day one. One hundred twenty of them came from contractors themselves: "full renovation, great prices". Instead of requests you got a feed of other people's ads.

Irrelevant leads are filtered at three levels: keywords, stop words and account filters. Akula Marketing supports all three, so your sales team only hears from people who are actually looking for something.

Why there is so much noise

In the example above, 120 of 140 alerts were ads. Sellers post in topic chats every day: contractors, agencies and freelancers describe their offers with the same words buyers use to describe a request. A keyword without context catches everyone who mentions the topic, so filters should start from buyer intent, not from the topic.

Level 1. Keywords with intent

Each keyword is matched as a substring, * matches any characters and (option1 | option2) sets alternatives. Instead of the topic, write the buyer's action:

  • "renovation" → "looking for (contractor* | builder*)";
  • "apartment" → "(want to rent | looking for) apartment*";
  • "ads" → "need a media buyer", "recommend a media buyer".

Ready-made sets for different niches are in the example groups: import them and extend.

Level 2. Stop words

Stop words follow the same rules as keywords: substring, * and options in brackets. If a message matches at least one stop word, it is skipped. A base set for most niches:

  1. Service offers: "I offer", "we do", "affordable", "discount*".
  2. Sales listings: "selling", "for sale", "in stock".
  3. Hiring: "vacanc*", "we are hiring", "join our team".
  4. Ads and links: "advertis*", "subscribe", "follow the link".

Put stop words into a separate group and attach it to all configurations at once.

Level 3. Account filters when scraping

When you build a member list, the noise is different: bots, empty and long-inactive accounts. Before export, filter by profile photo, online status in recent days, keywords in the bio and the text of the last message. AI enrichment adds gender, location and interests, so you can segment the list.

More on building lists in the guide to scraping Telegram group members.

Set up monitoring with stop words →

Qualification: what comes after the filter

A filter answers "is this a request or an ad". Qualification answers "is it worth the time". Ask 2–3 questions in your first message: budget, timeline, location. Mark people who do not reply or do not fit right away, so you never come back to them.

When conversations pile up, hand qualification to the AI outreach agent: it holds the conversation, scores each person against your criteria, tags them "Qualified", "Maybe" or "Not a fit" and passes the ready ones to your manager.

How to check that the filter works

  • After 2–3 days, count what share of alerts turned out to be real requests.
  • See which words triggered the noise and add them to stop words.
  • If alerts barely come in, the filter is too narrow: drop some conditions or add synonyms.

The full monitoring setup is in the guide to Telegram chat monitoring.

FAQ

How many stop words can I add?

Start with the base set above and add the words that trigger noise. The more precise your keyword phrases, the fewer stop words you need.

Do filters work in private chats?

Yes. A private chat added through a connected session is checked with the same keywords and stop words as a public one.

Can I filter messages in different languages?

Yes. Add keywords and stop words in each language separately: the service looks for an exact substring match, not a translation.

Is it legal to collect and filter chat messages?

The service works with publicly available messages and chats your account belongs to. Handle the contacts you collect according to your country's rules on personal data and marketing.

Hand qualification to an AI agent →