Outreach conversion
from 2% to 47%: how the jump happens

When you see an ad creative claiming "47% conversion on Telegram outreach" — it's most likely either out of context (response rate on a specific segment) or marketing stretch. The real picture is more nuanced: 47% is the upper bound, achievable only with the right stack in the right niche.

This breakdown isn't about "magic" 47%. It's about which specific mechanisms lift conversion from base 2% to "decent" 20-30%, and how 47% is reached under optimal conditions.

What "2% conversion" actually means — which metric

First source of confusion: everyone says "conversion" but means different things. Cold outreach typically tracks 4 metrics:

In "2% → 47%" we're talking about response rate. It's the most honest metric for evaluating outreach quality — it doesn't count whether the person became a customer (offer plays a role there), but counts whether your approach got a reply.

Component 1: Spintax randomization (+2× deliverability)

Spintax (spinning syntax) — a technique where variants inside a template are chosen randomly for each send. Example:

Why it works: Telegram groups identical messages by text hash. If 100 accounts send identical text — SpamBot sees a cluster and applies group-wide sanction (usually FLOOD_WAIT + potential PEER_FLOOD after 2-3 hours). Spintax breaks this grouping at the sender level.

Empirical numbers from our 2.4M sends:

Component 2: LLM personalization (+2-3× opens and replies)

Spintax randomizes phrasing, but the content stays the same. The next level — LLM rewrites the message for recipient context.

What "recipient context" means:

• Channel where the lead was found (crypto, SMM, e-com, edtech — different vocabulary)
• Language and region (by lang_code and bio)
• Last channel post (current topic → mention in message = relevance)
• Recipient's name (if present, substituted grammatically correctly)
• Channel tone (formal / informal → different stylistics)

Basic LLM prompt template (simplified):

Important nuance: the LLM doesn't generate every message from scratch — too expensive ($0.002 × 10K = $20 per batch, and randomness too high). In practice, a hybrid works: 20-30 LLM variants pre-generated, combined with spintax on the fly.

Effect on metrics:

Yes, 47% is not a starting number — it's the ceiling with full optimization in a favorable niche. Most clients see 20-35%, which is still ×6-10 over baseline.

Component 3: Human-like cadence (+1.5-2× reply rate)

Third layer — not text, but how the text arrives to the recipient. Mass senders usually ignore this aspect because it's not obvious in the UI.

What counts as cadence:

1. Typing indicator: show "typing..." for 1-4 seconds before sending. Small thing, but recipient sees it.
2. Delays between one account's messages: not burst, but distribution. Norm — log-normal with mean ~90 seconds and variance 60-300 seconds.
3. Different times of day: not all 500 messages at 14:00 Monday. Spread across 3-4 hourly windows with breaks.
4. Reaction to "read": if recipient read and didn't reply — don't hammer with "?" after 10 minutes. If unsubscribed — don't add to follow-up.

Why cadence affects response rate (not just delivery): the recipient subconsciously feels the rhythm. "Typing..." before a message, a 2-minute pause after reading before your reply — all of this creates the sense of a live conversation. In controlled tests on our data, cadence adds +40-60% to response rate vs "instant" sending.

Why the effects are multiplicative, not additive

Key observation from the data: components don't add up. They're multiplying coefficients in the funnel.

2.0 × 2.5 × 1.7 × 1.5 = ~12.75×. From baseline 2% that gives 25.5%. From 3% — 38%. From 3.7% — 47%.

Why multiplicative: each component operates at its own funnel layer:

• Spintax — increases share of delivered messages
• LLM — increases share opened out of delivered
• Cadence — increases share replied out of opened
• Niche fit — shifts baseline response rate among engaged users

If any component is zero — the whole funnel drops. 100% deliverability × 0% opens = 0% response. That's why "added spintax, conversion stayed the same" is a common complaint: one component without the others does little.

How it's assembled in TG:ON

The whole funnel is configured in the UI without code:

1. Template with spintax — editor with live preview, shows a specific combination example.
2. LLM rewriting — connect your API key (OpenAI, Anthropic, DeepSeek, Groq, Gemini) or use the built-in TG:ON pool. Prompt is configured for niche and tone.
3. Cadence profiles — "conservative" (medium delay 120s), "aggressive" (60s), "overnight" (450s). Choose or customize manually.
4. Content context — automatically picked up from the channel where the lead was found (if scraped through TG:ON).

All telemetry per message — delivery, opens, replies, block/spam-report — goes to the dashboard, you can A/B test templates.

Where this isn't enough

Honest: 47% doesn't work everywhere and always. Conditions where conversion will be lower:

• Cold niche or "wide" offer — if you write "someone about something", even a brilliant LLM won't save you. Niche down.
• B2B with a long decision cycle — there might be a response, but not to the offer, more like "tell me more in a Zoom". That's success too, but not "customer today".
• Offer doesn't match the channel — you scraped a crypto trading channel, offering SaaS for e-com. No magic there.
• Recipients aren't in your session — a user from a channel might have been inactive for a month. last_seen filter is important.

The TG:ON stack solves the problem of technical delivery and personalization. It doesn't solve: bad offer, bad targeting, bad continuation after first reply.