When your current ICP rubric is converting well but you've saturated the matched accounts, lookalike expansion uses ML to find accounts similar to your converters but outside your current criteria.
When to consider lookalikes
Your existing ICP is well-tuned and converting at target rates. You've worked through most of the matched accounts. You're hitting capacity on your current TAM and need to grow the addressable opportunity. Lookalikes expand TAM without diluting fit.
How lookalike modeling works
Train a model on your closed-won customers (the inputs are their firmographic, technographic, and behavioral attributes; the output is "closed-won"). The model identifies attribute patterns predictive of close. Apply the model to your full Golden DB; surface accounts that match the patterns but weren't in your existing ICP.
Setting up in turgo
ICP → Lookalikes → New Model → select source rubric (the rubric being expanded) → select closed-won data range (last 12 months recommended) → Train. The model produces a lookalike score for every account; accounts above threshold are added to your expanded universe.
Validating lookalikes
Don't trust the model blindly. Review the top 50 lookalike accounts manually — do they look like accounts you'd want to engage? If most do, the model is working. If many look wrong, the underlying training data has issues (small sample, bad labels, anti-pattern customers in won set).
Lookalikes as a tier
Treat lookalikes as a B-tier of pursuit, not A-tier. Your defined ICP gets first call; lookalikes get second call. As you accumulate close data from lookalike outreach, integrate confirmed patterns back into your ICP rubric.
