By Ramin Karimi · Updated 16 Nov 2026
The business of new customers on Meta
How a stranger becomes a customer, and whether it pays. What a new customer costs you, the most one may cost, and the 5 numbers that decide it, on a model business and on your own numbers.
The 5 numbersRead it allHow Ramin runs it
For owners in Canada and the US. Every made-up number says it’s a model.
Locked · The root · The business of new customers on Meta
You pay first. The customer decides.
Money goes out before the customer pays. The spend is certain, the return isn't.
Open the rootLocked · The root / Metamathics
Metamathics: math is mathing. Whether it pays.
The next one leaves more than he costs. You last until it's back.
OpenLocked · The root / Metamathics / The room
The room. What you can pay, before a dollar moves.
Spend too little and you can't tell if it works. Spend too much and you hit a ceiling: hands, cost or cash.
OpenDraft · The root / Metamathics / The road
The road. Every step from stranger to sale, each with its number.
You pay for views. Every step after them decides what a customer costs.
OpenDraft · The root / Metamathics / The read
The read. What's really happening, once the money moves.
Meta grades its own homework. Your books don't.
OpenDraft · The root / Strangers to Customers
Strangers to Customers. How it happens.
Nobody wants your product. A person in a moment wants something.
OpenDraft · The root / Strangers to Customers / The market
The market. Moments, not people.
Count the people in a moment, not the people who exist.
OpenDraft · The root / Strangers to Customers / Viscosity
Viscosity. Why 1 message doesn't get through.
The market is thick. 1 message alone doesn't get through. An arc does.
OpenDraft · The root / Strangers to Customers / Penetration
Penetration. Meta leads, the other touches help it through.
Meta leads. Everything else helps it get through.
OpenDraft · The root / Strangers to Customers / Remembered
Remembered. How to be the name he thinks of.
Be in the head before the moment comes.
OpenDraft · The root / Strangers to Customers / Profitable after
Profitable after. What he becomes after the sale.
How you win them decides what they're worth.
OpenThe surface
Clarity and breathing.
The 5 numbers, and today’s way to count each one: the owner’s word, the label beside it, then the formula. No ROAS.
- 1How manyNew customers in your books. Not Meta's count of purchases.The checkMeta's purchases ÷ your orders, every weekIf it jumps and your orders don't, Meta changed how it counts.
- 2What each costsWhat you paid for each new one, from your books.What a new customer costsnCACwhat you spent to get new customers ÷ new customers in your booksMonthly.Meta's own count of itnCPAMeta spend ÷ the new customers Meta reportsBeside your books, never alone.Meta's real costMeta spend ÷ the new customers the off test says Meta broughtThe off test: switch Meta off in 1 place, count what's missing.The next oneextra spend ÷ extra new customersOn big, dated steps.
- 3What each leavesWhat he leaves you, and the most he may cost.What a new one leaveswhat he pays you in 12 months − what serving him costs − refundsYour linewhat he leaves − the profit you want from himThe most a new customer may cost you.
- 4How fast it's backHow long your money is out.Back inthe weeks until what he leaves covers what he costThe holewhat you spend a week × the weeks until it's back
- 5How sureHow much of what you see is noise.The swingat about 50 a week, the cost per result moves 22% down to 39% up with nothing changedJudge a month, not a week.Ways to split the budgetad setsweekly budget ÷ (cost per new customer × 50)Meta's own check, as reported 21 Oct 2026.
Read it all
The whole web, in words.
Every node in the order of the dive: its first sentence, the math, a store, a service and an app, the latest, graded and dated, and the points. Open one to read it.
The root · LockedYou pay first. The customer decides.
The business of new customers on Meta. Everything else stands on it.
Money goes out before the customer pays. The spend is certain, the return isn't.
The job: decide how the money is spent so it comes back as profit.
The 5 numbers: How many, What each costs, What each leaves, How fast it's back, How sure. Each one, and how to count it.
The math
On a store, a service and an app
You pay for the ad, and often the stock, before the order.
You pay for the ad and the hours before the job is paid.
You pay for the signup months before his payments cover it.
The latest, graded and dated
Meta's average price per ad was up 14% on the year in Q3 2026. Paying first keeps getting dearer.Meta, Q3 2026 results (its own release, read 16 Nov 2026) · 21 Oct 2026
In the US and Canada, Meta's price per ad was up 23% in Q3 2026, on 8% more ads.Meta's Q3 2026 earnings slides, as reported by PPC Land and others (the slides not opened) · 21 Oct 2026
The points
- You pay first. The customer decides.
- The spend is certain. The return isn't.
- The job isn't getting customers. It's spending so the money comes back as profit.
- 5 numbers decide it: how many, what each costs, what each leaves, how fast it's back, how sure.
In the method
Metamathics · LockedThe next one leaves more than he costs. You last until it's back.
Metamathics: math is mathing. Whether it pays.
Not the average customer. The next one, the one the next dollar buys.
Counted in your books, not in Meta's countAds Manager.
Its half of the root: “You pay first.”
The math
On a store, a service and an app
The next order the next dollar buys.
The next client, and how full you already are.
The next subscriber, who pays you back over months.
The latest, graded and dated
Meta's own bidding docs: costs can rise as the cheapest chances get used up, or as you raise the budget. The next one costs more than the last.Meta for Developers, Bid Strategy · read 16 Nov 2026
The points
- The next one leaves more than he costs. You last until it's back.
- The average lies. You're paying for the next customer, not the average one.
- A customer who pays you back can still sink you, if he pays back too late.
- Your books know who's new. Meta doesn't.
In the method
The room · LockedSpend too little and you can't tell if it works. Spend too much and you hit a ceiling: hands, cost or cash.
The room. What you can pay, before a dollar moves.
The sentence and its 4 pieces are locked. The formulas in them are drafts.
4 pieces, in order: what the next one leaves, the most he may cost, where the room starts and ends, and the hole.
They hold 4 of the 5 numbers, all known before a dollar moves. The 5th, how sure, is the read's.
The math
A model store
| He pays per order | $100 |
| Serving 1 order costs | $60 |
| Orders in 12 months | 1.5 |
| What a new one leaves | $60 |
| The profit you want from him | $20 |
| Your line | $40 |
| A new customer costs today | $30 |
| The floor in spend, 50 a week × $30 | $1,500 a week |
| Ways $1,500 a week can be split | 1 |
| The hole, paid within the week | about $1,500 |
Every number here is a model. Change them in Your numbers.
On a store, a service and an app
Cost hits first when the first order pays. Cash hits first when stock is bought before it sells.
Hands hit first. When you're full, the next one leaves less: overtime, a hire, a worse job.
Cash hits first. The wait for his payments is the ceiling.
The latest, graded and dated
Meta's help page: each way you split the budget leaves learning "usually after about 50 results in the week after" its last big change. Usually and about: a guide, never a gate.Meta Business Help Centre, the learning phase · read 1 Oct 2026
Meta for Business, peak season post: splits under about 50 results a week cost about 28% more per purchase. No method given.quoted by 3 trade outlets; the page not opened · 21 Oct 2026
The points
- Spend too little and you can't tell if it works. Spend too much and you hit a ceiling.
- 3 ceilings: hands, cost, cash. Your room ends at the lowest one.
- Set the most a new customer may cost before you spend a dollar. Then don't argue with it.
- Under about 50 a week, you're guessing. So is Meta.
- Split a small budget 5 ways and none of it learns.
- When you're full, the next customer leaves less.
- Same Meta spend. Wait 6 weeks for the money instead of 1, and the hole is 6 times deeper.
- A rare buy in a small town may not have 50 real buyers a week. Count them before you spend.
In the method
The road · DraftYou pay for views. Every step after them decides what a customer costs.
The road. Every step from stranger to sale, each with its number.
Meta sells you the view. Your ad, your page, your reply and your close turn views into customers.
Meta runs the auction for the view, and a better ad can win it against bigger bids. Every step after the view is yours.
The math
A model store
| Cost per 1,000 views | $20 |
| Click rate | 1.5% |
| Clicks that buy and are new | 3% |
| New customers per 1,000 views | 0.45 |
| Cost per new customer | about $44 |
| The page lifted from 3% to 4% | about $33 |
Every number here is a model.
On a store, a service and an app
View, click, page, cart, paid, new.
View, click, lead, reply, booked, closed.
View, install, trial, paid, still paying.
The latest, graded and dated
1 vendor's order-checked panel of stores, week of 1 Nov 2026: Meta's click was the cheapest of its series and the new customer still cost $104. The steps after the click moved, not the click.Wicked Reports, The Paid Traffic Truth, Issue 020 (a vendor) · 9 Nov 2026
Same ads, same week, same money. The slow reply cost far more per new customer than the fast one.a client's case, from before (its numbers are the client's and stay out) · from before
The points
- Meta sells you views. You turn them into customers.
- Cost per customer = what you pay for 1,000 views ÷ how many of them become new customers.
- Lift any 1 step 10% and the customer gets about 9% cheaper.
- Meta runs the auction. A better ad can win the view for less. Every step after the view is yours.
- The leak is after the lead.
- Same ads, same week, same money. The slow reply cost far more per new customer.
- A cheap click that never buys is the most expensive thing you can buy.
In the method
The read · DraftMeta grades its own homework. Your books don't.
The read. What's really happening, once the money moves.
Meta's countAds Manager doesn't know who's new. It doesn't know what would have happened anyway.
Your books know who's new. The off testa holdout knows what Meta brought.
The math
A model store, 1 month, $6,000 on Meta
| Meta's count: 230 purchases | $26 each |
| Your books: 150 new customers | $40 each |
| The off test: 120 of them Meta really brought | $50 each |
| Your line | $40 |
Every number here is a model. Meta's count looks great. The truth is over the line.
On a store, a service and an app
Your orders know who's new. Check every order against past orders.
Your books know who's new. The closed customer can land weeks after the click, outside Meta's 7-day window.
The paid account knows who's new. What he's worth shows over months, so read the count early and the worth late.
The latest, graded and dated
Meta changed how it counts 3 times in 2026: the 7-day and 28-day view windows removed (12 Jan), a click counted only when it's a link click (March), and sales matched on the customer's device, on by default (12 Nov).Meta's developer blog and help pages, in the meta-codex; dates from trade press · 2026
Meta's own help page on new customers: judge on what a new customer costs by your own tools, not on the campaign's cost per resultCPA alone.Meta Business Help Centre, Customer lifecycle strategy · read 1 Oct 2026
839 Meta lift tests, Nov 2019 to Mar 2020: counting the last clicklast-click over 7 days overstated what the ads really brought by 33% on average.Gordon, Moakler and Zettelmeyer, PIE, a working paper revised Apr 2026 · 2026
640 off tests, 2024 to mid 2025, brands averaging $14M a year on Meta: on 7-day click, Meta under-reported what it brought by 15% on average. The error runs both ways.Haus (a vendor) · Aug 2025
1 vendor's order-checked panel of stores: what a new customer cost on Meta, about $85 to $91 over the summer, $104 in the first week of November.Wicked Reports, The Paid Traffic Truth (a vendor) · Jul to Nov 2026
The points
- Meta grades its own homework. Your books don't.
- Meta's count, your books and the off test give 3 different costs for the same customer. Only the last 2 count.
- Meta changed how it counts 3 times this year.
- If Meta's purchases jump and your orders don't, Meta changed its counting, not your sales.
- At 50 results a week, the cost per result swings 22% down to 39% up with nothing changed.
- Judge a month, not a week.
- ROAS counts revenue. You keep what's left.
- Meta can't tell a new customer from an old one. Your books can.
- Even Meta's own help page says: judge new customers with your own tools.
- Meta's error runs both ways. Only the off test tells you which way yours runs.
- The off test needs no software. The only cost is the sales you skip in that slice.
In the method
Strangers to Customers · DraftNobody wants your product. A person in a moment wants something.
Strangers to Customers. How it happens.
When the moment comes he does 1 of 4 things: calls the name already in his head, asks someone (more and more an AI), goes looking, or freezes.
He doesn't pick the best. He picks the first name that's good enough.
Its half of the root: “The customer decides.”
The math
On a store, a service and an app
The moment is a need, a gift, a season, a thing that broke.
The moment is an event in his life. Something happened.
The moment is a new job to do, or the old tool failing.
The latest, graded and dated
Meta, 19 Oct 2026: a new ad starts from what Meta already knows about similar ones. "Every new ad starts at zero" is not true.Meta Engineering, GenRecs Tokenization · 19 Oct 2026
The points
- Nobody wants your product. A person in a moment wants something.
- He doesn't pick the best. He picks the first name that's good enough.
- When the moment comes: he calls a name he knows, asks someone, goes looking, or freezes.
- More and more, the first one he asks is an AI.
In the method
The market · DraftCount the people in a moment, not the people who exist.
The market. Moments, not people.
Something happened first. The moment opens a window, and windows close.
Most of the market isn't in a moment right now. It's marinating.
The math
A model on an official rate
| An event that happens to | 700 in 100,000 people a year |
| In an area of 250,000 people | about 34 a week |
| Meta's guide for 1 split of the budget | about 50 a week |
Use your own event and 1 official source. The whole market can sit under Meta's floor.
On a store, a service and an app
Moments come in seasons and gifts, and when a thing breaks.
Moments are events. For a rare buy, count them before you spend.
Moments are switches: a new tool to try, the old one failing.
The points
- Count the people in a moment, not the people who exist.
- A moment opens a window. Windows close.
- Most of your market isn't buying today. It's marinating.
- Something happened first. Find out what.
- A made moment is a real event. Never fake urgency.
- If you sell a rare buy, the market can be your first ceiling.
In the method
Viscosity · DraftThe market is thick. 1 message alone doesn't get through. An arc does.
Viscosity. Why 1 message doesn't get through.
Scattered messages hit different parts of the market and stop there.
Line them up into 1 campaign arc and they get through: the ad, the page, the reply and the follow-up, run as 1 thing.
On a store, a service and an app
1 promise, said the same way, on every ad, page and email.
The ad, the reply and the call say the same thing.
The ad, the store page and the first day in the app say the same thing.
The latest, graded and dated
Meta's own advice: fewer ads in each split of the budgetad set, with different kinds of ad inside each. Variety inside 1 arc, not 12 arcs pretending to be 1.Meta Business Help Centre · read 1 Oct 2026
The points
- The market is thick. A single ad doesn't get through.
- Scattered messages hit different parts of the market and die there.
- Line them up into 1 arc and they get through.
- Say 1 thing, the same way, long enough to be kept.
- Variety is for Meta. The message is for the market.
- Be boring on purpose. The message stays.
- 12 ads serving 1 arc is a healthy account. 12 ads serving 12 ideas is 12 accounts pretending to be 1.
In the method
Penetration · DraftMeta leads. Everything else helps it get through.
Penetration. Meta leads, the other touches help it through.
The other touches: agentic, digital, personal, emotional, human, physical, commerce.
On the feed, what's in the ad helps decide who sees it. The ad helps pick its buyer.
The math
On a store, a service and an app
Meta finds him. The reviews, the email and the box keep him.
Meta finds him. The people he asks and the reply close him.
Meta finds him. The store page and the first day keep him.
The latest, graded and dated
Meta reads the ad itself to help decide who sees it: what's in the ad is a ranking input, not decoration.Meta's own research, in the meta-codex · 2026
Meta's next Andromeda, 28 Oct 2026: retrieval rebuilt, about 300 times less compute. Still retrieval only: it picks the few thousand ads that get a chance.Meta Engineering · 28 Oct 2026
The points
- Meta leads. The rest helps it get through.
- On the feed, the ad helps pick its buyer.
- A business that lives on 1 platform is renting its customers.
- The name in the head is the only ground you own.
- Your share of the moments: your new customers ÷ the moments in your area.
In the method
Remembered · DraftBe in the head before the moment comes.
Remembered. How to be the name he thinks of.
His head, the heads of the people he'll ask, and what an AI reads about you.
File it under his moment, not under your name.
On a store, a service and an app
Be the name that comes up when the thing breaks.
Be the name his friend says. For a rare buy, the people he'll ask come first.
Be the name that comes up when the old tool fails.
The latest, graded and dated
Studies, from before: most forgetting happens in the first day; to hold a thing for a year, about 3 weeks between 2 meetings worked best; people rate what they've heard before as more true.THE-MASTER 2.1: Ebbinghaus 1885, replicated 2015; Cepeda and others 2008; Dechêne and others 2010, 51 studies · from before
The points
- Be in the head before the moment comes.
- File it under his moment, not under your name.
- Most forgetting happens in the first day.
- Heard before feels true.
- He'll ask someone. Be in their head too.
- More and more, the first one he asks is an AI. Be in what it reads about you.
In the method
Profitable after · DraftHow you win them decides what they're worth.
Profitable after. What he becomes after the sale.
A deep first deal can buy cheap customers who leave. Test it, never assume it.
The good ones come back and send others.
The math
On a store, a service and an app
After is the 2nd order.
After is the referral.
After is the month he stays.
The points
- How you win them decides what they're worth.
- A deep first discount can buy cheap customers who leave.
- The good ones come back and send others.
- Ask every new customer who sent them. It's the only way to count it.
- Some new customers would have come anyway. Some were sent by the ones you paid for. Meta's count shows neither.