Lookalike & Similar Audiences

9 min read

Learn

Quick Reading

Estimated 9 mins

Prereq

Intermediate

Basic ML concepts helpful

Interactive

Static Playbook

Static guide & reference tables

The core idea: expand reach from a known-good seed

A Customer Match upload gives you a precise, high-confidence audience of people you already know converted or engaged — but it's inherently capped at the size of your existing customer list. Lookalike (Meta's historical term, now folded into Advantage+ audience features) and Similar Audiences (Google's equivalent, deprecated for some surfaces but conceptually persistent in Google's modeling) solve the scale problem: the platform analyzes the behavioral and demographic patterns shared by your seed audience, then finds new users elsewhere in its user base who share those same patterns.

Seed quality determines everything downstream

A lookalike model is only as good as the seed it learns from. A seed of 50,000 recent high-value purchasers produces a meaningfully different, usually better-performing lookalike than a seed of 50,000 email newsletter subscribers who never purchased — the model finds people who resemble whoever is in the seed, not people who resemble your ideal customer in the abstract. The most common mistake is seeding from whatever list happens to be largest or easiest to export, rather than the list that most precisely represents the actual outcome you want more of.

Seed audience options and what they actually optimize toward

Seed audience	What the lookalike will find	Best use case
All website visitors	Broad, low-precision resemblance to generic traffic	Rarely a good seed on its own
Purchasers in the last 90 days	Users resembling recent converters	Acquisition campaigns targeting likely buyers
Top 10% LTV customers	Users resembling your most valuable segment, not just any buyer	High-value acquisition, often smaller but better-performing reach
Cart abandoners	Users resembling people who showed intent but didn't convert	Different targeting intent — closer to retargeting-adjacent prospecting

Similarity percentage/size is a precision-vs-reach dial

Meta's lookalike sizing (1%-10% of a country's population, roughly) and Google's audience similarity tuning both represent the same tradeoff: a tighter, smaller similarity setting (closer to the seed's exact profile) yields higher precision but less reach; a looser, larger setting trades precision for scale. There's no universally correct setting — it depends on whether the campaign goal is efficient acquisition at a known CPA target or maximum reach within a broader acceptable cost range.

Platform modeling has moved toward automation, reducing manual lookalike control

Both Google and Meta have shifted substantial audience modeling into automated systems (Google's Performance Max audience signals feeding automated targeting, Meta's Advantage+ Audience) that use seed-like inputs as *signals* to an algorithm rather than exposing an explicit, manually-sized lookalike audience the way older interfaces did. This reduces granular manual control but often improves performance, since the algorithm can incorporate real-time conversion feedback that a static, manually-refreshed lookalike list cannot.

seed_selection_query.sql

typescript

-- Example: building a high-value seed audience from the warehouse before upload,
-- rather than seeding from an undifferentiated "all customers" export.
SELECT user_pseudo_id, email_hash
FROM customer_ltv_summary
WHERE lifetime_value_percentile >= 90
  AND last_purchase_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 180 DAY)
  AND marketing_consent = TRUE

Lookalike audiences inherit any bias in the seed

If a seed audience is skewed (e.g., overwhelmingly one demographic because of how it was historically acquired), the resulting lookalike audience amplifies that skew rather than correcting it. This is a real optimization risk, not just a fairness concern — a biased seed can systematically exclude segments of your actual addressable market from ever being reached by prospecting campaigns.

What's Next?

Lookalike and similar-audience modeling is one expression of a much broader shift toward platforms absorbing targeting and bidding decisions into automated, signal-driven systems — the same underlying trend reshaping how bidding strategies, personalization, and even creative optimization now work.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →