--- name: build-lookalike-audiences description: >- Builds Meta lookalike audiences by ranking seed quality - purchasers, high-LTV segments, and value-based lists rank far above plain engagement, which ranks above all-site-traffic - then sizing the 1%-10% country-scoped match band and gating on a ~1,000+ matched-seed floor before calling mads_create_custom_audience (lookalike subtype). Covers value-based lookalikes (the seed must carry real LTV or purchase-value variance, not a flat number), stacked non-overlapping bands (1%, 1-3%, 3-5%) for scaling past a saturated 1% tier, feeding a lookalike as an Advantage+ suggestion vs a hard ad-set target, and a refresh trigger tied to seed growth (roughly doubling) or a ~90-day ceiling. Reach for it once a qualifying seed exists and prospecting must scale beyond the warm pool. It ends at a created, sized lookalike - it does not build the seed (build-custom-audiences, upload-customer-list-audience) or exclude the seed/customers from delivery (set-exclusions-and-suppression). --- # Build Lookalike Audiences ## Purpose A lookalike is only as good as its seed, Meta's matching models people against the seed audience, so a broad, low-intent seed produces a broad, low-intent lookalike no matter how carefully the percentage is chosen. This skill forces seed selection through a quality ranking before any sizing decision, then walks the 1%-10% tradeoff, value-based lookalikes, and the refresh discipline that keeps a lookalike from drifting away from the account's actual best customers. ## When to run - Prospecting needs to scale past the current retargeting/warm-audience pool size. - A purchaser, high-LTV, or value-based source newly cleared the ~1,000-matched seed floor. - An existing lookalike's seed has grown materially since the lookalike was built. - An Advantage+ shopping or Advantage+ audience campaign needs an audience signal rather than hard targeting. ## When NOT to run - No qualifying seed exists yet → `build-custom-audiences` or `upload-customer-list-audience` first. - You need to exclude the seed or existing customers from the lookalike's own delivery → `set-exclusions-and-suppression`. - You are building the tiered retargeting structure rather than a prospecting audience → `design-retargeting-ladder`. ## Prerequisites - A qualifying seed: a custom audience (ideally purchasers or high-LTV customers) or a pixel purchase event, at or above ~1,000 matched users in the target country. - The business objective for the lookalike (prospecting volume vs. efficiency) to pick the size band. - Target country or countries decided, lookalikes are country-scoped. ## Procedure 1. **Select the seed by quality, not convenience.** Rank candidates: purchasers / high-LTV customer list / value-based pixel source, above engaged or AddToCart custom audiences, above a plain Purchase-event audience, above all-site-traffic. Do not seed from all-traffic when a purchaser-based source exists. 2. **Confirm seed size.** Call `mads_list_custom_audiences` (or `mads_list_pixels` for pixel event volume) and confirm the seed clears ~1,000+ matched users in the target country. Below that, stop and grow the seed first, do not build a lookalike on a sub-floor seed. 3. **Check value-data eligibility for a value-based lookalike.** Confirm the source carries a real LTV or purchase-value field with actual variance, a customer list with identical values per record, or a pixel with no value parameter, cannot support value-based lookalike modeling and should use the plain lookalike path instead. 4. **Choose the country scope.** One lookalike per target country (or supported region grouping); do not build a single lookalike spanning unrelated markets. 5. **Choose the size band.** Start at 1% for the closest match and smallest reach. When scaling, add 1-3% and 3-5% as separate stacked bands rather than jumping straight to a wide flat percentage. 6. **Create it.** Call `mads_create_custom_audience` with the lookalike subtype, seed audience ID, size ratio, and country. Preview the definition, get explicit user approval, then commit. 7. **Verify the result.** Call `mads_list_custom_audiences` and sanity-check the resulting size against the expected share of the country's population for that percentage. A materially undersized result signals a seed/country mismatch worth re-checking before launch. 8. **Decide hard-target vs. Advantage+ signal.** For standard ad sets, attach the lookalike as direct targeting. For Advantage+ shopping/audience campaigns, feed it as an audience suggestion, Advantage+ treats it as a starting point the algorithm can expand beyond, not a fence. 9. **Set the refresh trigger.** Rebuild the lookalike when the seed audience has grown roughly double its size at creation, or at ~90 days, whichever comes first. Note the trigger alongside the audience record so it is not forgotten. ## Decision rules - **Seed quality ranking:** purchasers / high-LTV / value-based customer list > engaged or AddToCart custom audience > pixel Purchase event > all-site-traffic. Never default to all-traffic because it is the largest available audience. - **Seed floor:** ~1,000+ matched in the target country is the minimum; more (2,000-5,000+) improves match quality further. Below floor, do not build, grow the seed instead. - **Sizing tradeoff:** 1% = smallest reach, closest match, best efficiency; 10% = broadest reach, weakest match. Default to 1% for efficiency-first prospecting; expand toward 3%/5% only once 1% is genuinely budget-constrained or saturated. - **Stacking bands:** build 1%, 1-3%, 3-5% as separate non-overlapping tiers (exclude the lower band from the higher one) instead of one flat 5% lookalike, so budget finds the best-matching tier first. - **Value-based gate:** only build a value-based lookalike from a source with genuine value variance; a flat-value source degrades it to a plain lookalike with false precision. - **Advantage+ usage:** signal/suggestion for Advantage+ campaigns; hard targeting for standard ad sets. - **Refresh cadence:** rebuild on ~doubling of the seed or ~90 days, whichever happens first. - **Done means:** the lookalike is built from a quality-ranked, floor-clearing seed, at a deliberately chosen size band and country, with its resulting size verified and a refresh trigger recorded. ## Common failure modes - **Seeding from all-website-traffic** because it is the biggest available audience, produces a broad, low-converting lookalike that looks like a targeting problem but is actually a seed-selection problem. - **Building at 10% for reach** without acknowledging the quality tradeoff, then attributing the resulting weak ROAS to the lookalike tactic itself rather than the size choice. - **Seed under the ~1,000 floor**, the platform cannot model similarity reliably; the lookalike under-delivers or plateaus almost immediately. - **Value-based lookalike from a flat-value source**, no better than a plain lookalike, but reported and trusted as if it were value-optimized. - **Never refreshing**, a lookalike seeded a year ago models last year's customer, not the account's current best buyer. - **Forgetting to exclude the seed and existing customers** from the lookalike's own delivery, so prospecting spend re-serves people who already converted. ## Related skills - Run before this: `build-custom-audiences`, `upload-customer-list-audience` (seed sources). - Run after this: `set-exclusions-and-suppression` (exclude the seed/customers from lookalike delivery). - Related: `design-retargeting-ladder` (the lookalike is a prospecting layer feeding the top of that ladder).