Pulsar · Metrics playbook · No. 03 · The pixel-data → lookalike quality problem. Real seed numbers from the test pixel, 30-day window.
Garbage In, Lookalike Out
Your pixel is teaching Meta who to chase. Feed it junk and you pay to find more junk.
Companion to No. 01 — Efficiency & Fatigue and No. 02 — Quantity vs. Quality.
Contents
01 Meta learns from your pixel · 02 Three ways a seed goes bad · 03 The real seed problem · 04 The doom loop · 05 Match quality: the silent poison · 06 What to seed on instead · 07 Where Pulsar fits
The thing nobody explains: Meta LEARNS from your pixel
Meta doesn’t just REPORT your pixel data — it OPTIMIZES from it. Everything the pixel records becomes training material for who your money chases next.
So the quality of everyone Meta finds is capped by the quality of the handful you fed it. There is no algorithm that upgrades a bad seed — scale only multiplies whatever pattern was in it.
Meta doesn’t just count your customers — it clones them. Bad clones cost you every single day.
The three ways a seed goes bad
A lookalike fails in three distinct ways — and each one is invisible in the ads dashboard until the invoice arrives.
Too few
Thin seed
quantity failure
A lookalike wants roughly 100–500+ quality seed profiles to read a pattern. Give it a handful and Meta is guessing — it pads the audience with noise, because there isn’t enough signal to extrapolate from.
Wrong action
Wrong event
targeting failure
Seed on window-shoppers — View content / Add to cart — instead of real buyers (Purchase / high-value) and Meta learns to find more window-shoppers. Your CTR looks fine; your CPA quietly climbs.
Poisoned data
Dirty seed
integrity failure
Bots, mis-fired events, duplicates, and unmatched (low match-quality) conversions all corrupt the pattern — and the damage SCALES, because you’re paying to find “more people like these.”
All three are seed problems, not bidding problems. No budget change, bid strategy or creative refresh fixes an audience that was extrapolated from the wrong people.
Your best seed this month is smaller than you think
Distinct visitors on the test pixel, 30-day window. Every step up in intent purity costs you an order of magnitude in seed size.
Pick a seed event — watch size and purity trade off
Counts are real (distinct visitors, test pixel, 30 days). The intent-purity meter is a general principle — an illustration of signal quality, not a measured score for this pixel.
Seed size
10
distinct people, 30 days
Intent puritypure buyer signal
The seed picker is interactive in the HTML version. The five candidates: Just visited 3,675 (pure quantity, ~zero intent) · Viewed a product 1,529 (mild intent) · Added to cart 93 (real intent) · Started checkout 37 (strong intent) · Purchased 10 (the only true customer seed).
The doom loop — why it COMPOUNDS
This is the core danger: a bad seed doesn’t just waste one campaign — it feeds itself. Each cycle of the loop below makes the next lookalike a little worse than the last.
The core danger — every other mistake in this series costs you money once. This one compounds: today’s junk converters become tomorrow’s seed, and tomorrow’s seed buys the day after’s junk.
Match quality: the silent poison
Even genuine buyers don’t help the seed if Meta can’t MATCH them to a person. An unmatched conversion can’t join your seed — it simply doesn’t exist to the lookalike.
A purchase Meta can’t match is a customer your lookalike never learns from.
Full treatment coming — a future lecture in this series covers Event Match Quality and CAPI (server-side sending) in depth. Here, just remember: clean seed also means matched seed.
What to seed on instead — the playbook
Five steps, in order. Each one exists because of a failure mode from section 02.
Seed on your BEST outcome
Purchase, ideally high-value buyers — not View content or Add to cart. The lookalike copies whatever you point it at; point it at the people you actually want more of.
Get enough CLEAN volume
If purchases are thin, widen the window or step up to the nearest high-intent event (e.g. Checkout) — rather than dropping to raw traffic. Trade a little purity for size; never trade all of it.
EXCLUDE the junk
Bots, internal/test events, duplicates — anything that isn’t a real outside human distorts the pattern the lookalike learns.
VERIFY match quality
Before you build the audience — a seed full of unmatched conversions is a seed full of ghosts (section 05).
REFRESH the seed
As better data accumulates, rebuild. Don’t let a stale or bad seed compound — that’s the doom loop from section 04 running on a timer.
Seed strength = size AND purity
A general-principle model, not a Meta formula: seed strength is capped by the weaker of the two. Move the sliders or hit a preset — both real options from this pixel score weak.
Seed strength2 / 100
This meter is interactive in the HTML version. Both presets score weak: 10 purchases = max purity but far below the hundreds a lookalike wants; 1,529 viewers = plenty of volume but 1,519 of them never bought. Strength needs BOTH size and purity.
Feed Meta your best 100 customers and it finds more like them. Feed it 10 noisy ones and it finds more noise — with your money.
Where Pulsar fits
Pulsar exists to make the seed clean before Meta ever sees it.
Events deduped, bots filtered out, counted by real PEOPLE (personas) not sessions, sent server-side for high match quality — and audiences you can build from a QUALIFIED funnel stage instead of raw traffic. So the seed you hand Meta is one it can actually learn something good from.
Cross-reference — see No. 02 — Quantity vs. Quality for the funnel these qualified stages come from, and No. 01 — Efficiency & Fatigue for how a bad audience shows up later as rising CPA.