Short answer: during an AI agent's overseas cold start, four core metrics each week are enough: outbound reply rate (are list and message right?), qualified meetings (is the need real?), trial-to-paid conversion (do product and price hold up?) and cost per paying customer (can the channel scale?). Add activation rate and model cost per customer, and you can tell whether your acquisition model works.
The four core metrics
| Metric | Formula | What it tells you | If poor, check first |
|---|---|---|---|
| Reply rate | Replies ÷ delivered outreach | Whether list and message are right | ICP fit of the list, inbox placement |
| Qualified meetings | Weekly meetings with ICP matches | Whether the need is real and the message lands | ICP, value proposition |
| Trial-to-paid | Paying customers ÷ trials started | Whether product and price hold up | Activation, pricing, trial follow-up |
| Cost per paying customer (CAC) | Total channel spend ÷ paying customers from it | Whether the channel can scale | Channel choice, conversion path |
Include people's time in acquisition cost, not just ad spend and tools.
Supporting metrics
- Activation rate: share of sign-ups who complete the key action (reach the "useful moment"). See Designing the conversion path.
- Positive reply rate: share of replies showing interest; more telling than total replies.
- Sales cycle: time from first touch to payment.
- Model cost per customer: a cost specific to AI agents that affects gross margin.
- CAC payback period: CAC ÷ (monthly revenue per customer × gross margin), the number of months to recover acquisition cost.
- Retention and net revenue retention: whether customers keep using and expand.
Split by channel and segment
Totals hide problems. An average overall reply rate may conceal one country or company type replying well and another poorly. Split every metric by:
- Channel: outbound, LinkedIn, communities, creators, partners, ads;
- Customer segment: industry, size, country;
- Message or landing page version.
Weekly review template
Spend half an hour each week answering:
- What are this week's four core metrics? Better or worse than last week?
- Which channel and which segment performed best?
- What did we change this week, and what happened?
- What problem or objection came up most in customer feedback?
- What is the one most important thing to change next week?
Common metric traps
- Vanity metrics: impressions, followers and site visits do not show whether acquisition works.
- Concluding from tiny samples: cold start data is thin; read it together with customer feedback.
- Changing several variables at once: you cannot tell which change worked.
- Benchmarking only against industry averages: definitions differ; comparing with yourself is more reliable.
FAQ
How often should we review during the cold start?
Weekly. Things move fast in the cold start; monthly reviews are too slow.
Do we need dedicated analytics tools?
A spreadsheet is enough early on. What matters is recording consistently with stable definitions. Add a CRM and analytics tools as channels and customers grow.
When has the cold start "worked"?
When one or two channels bring ICP-matching paying customers fairly reliably, acquisition cost is acceptable and customers stay. Then consider investing more and scaling those channels. See The complete guide to overseas user acquisition.