Ramin Karimi, Campaign Director

← Ramin Karimi

The process and the price

Change 20 things at once and nobody can say what moved the number.

So everything below is one arc: one idea, carried all the way to the market, held to one number. What a new customer costs.

6 months
Today to A to B. A is the date the number becomes readable, written in your proposal.
1 number
What a new customer costs, read against a target set from your own numbers before anything runs.
The price
Starting at $8,000 USD a month. The ad money runs on your own card.
Month 1
The fit. Not a fit, clean stop.

I direct the arc. ORCAS builds it, or your studio does, to my brief.

02The roadThe shape of 6 months

Results start at the launch. The reading starts at A.

6 months, drawn as a road. Nothing on it is readable until A, then the reading comes in wide and settles.

TODAY COUNTED, BLENDED, NOT CAUSED THE NORMAL SWING THE TARGET A THE READING STARTS B · MONTH 6 TODAY THE NORMAL SWING THE TARGET A B · MONTH 6 TODAY: COUNTED, BLENDED, NOT CAUSED A: THE READING STARTS

Model · the shape, not your numbers

Today
The number you have. Counted, blended, not caused, and not on the cost scale. Real, and impossible to move on purpose.
Today to A
The build. Everything on this page gets studied, decided, made and wired inside this stretch. Nothing is readable in it, and that is in the proposal before any money moves.
Month 1
The fit. Not a fit, clean stop.
Point A
Everything live, the learning done, the gauge reading: the one readout that shows what a new customer costs against the line it has to clear. The eyes open, on the right lens. A is dated in your proposal.
Month 3
3 honest outcomes. The number. Or not yet, with the date. Or the world moved, and here is the number read against that.
A to B
The run. One change a month. Each cohort read on its own, never blended into the last one.
Point B
Month 6. The number read against the target line. A direction arrived at, not a point to hit.
The fuel
What it costs to find out. Authorised at A, runs to B. Paid to Meta on your own card, and it never passes through me.
03The studyMonth 1, the first 2 weeks

An idea built on half the facts is a guess with money behind it.

So I read 2 things before I build anything. What your business already says to the market, and what that already costs. Then what the market says back. Both are finished before one piece of creative exists.

WHAT YOUR BUSINESS SAYS NOW WHAT THE MARKET SAYS BACK BOTH READ BEFORE ONE IDEA EXISTS THE ARC WHAT YOU SAY NOW WHAT THEY SAY BACK BOTH READ BEFORE ONE IDEA EXISTS THE ARC
Everything I read4 reads

1 · The account, as it stands

  • Whose account it is, how old it is, and what it has been optimising on since the day it was opened.
  • The exclusions, first, on every account, no exceptions. An account can prospect straight into its own buyers.
  • The placements, against what the platform still supports this month. A retired placement never warns. Name it in a new or edited ad set and the request errors, or it is dropped without a word.
  • The special category declarations. Housing, employment, financial products and services, and social issues, elections or politics each have to be declared, or the ads get rejected. For the first 3, an ad set built through the API also has to state its audience setting, and a missing one fails on creation rather than warning.
  • Everything currently running, screenshotted before a single change is made.
  • What stays live. Existing winners keep running at their current spend on their existing event. Nothing is deleted, and no reading is claimed off them.

2 · Your numbers, checked against your words

  • The last 50 new customers, and what each one paid first.
  • What it cost to deliver each of those 50.
  • Refunds, reductions and no-shows across the last 3 months.
  • Repeat and lifetime, out of the order history rather than out of memory.
  • The profit you keep. Your number, in your words.
  • The sales cycle, from the last 10 closed: first contact to money in the bank.
  • Capacity. How many more next month before something breaks.
  • The 3 thinnest months of last year, and what cash looked like inside them.
  • Annual gross profit after delivery labour.
  • New against returning, last month, out of the server rather than the platform.
  • Branded search today. 16 months pulled out of Search Console at handover, before the window closes, with direct traffic beside it.

Never estimate what can be counted. A number nobody has becomes a model with a test date, labelled a model, and the first real reading replaces it. Never a guess dressed up as a reading.

3 · The market, in a fixed order

  • Prices. Where the crowd sits, where the top sits, and the gap nobody is standing in.
  • The category’s entire advertising. The ad libraries are public. Every competitor’s live creative, read as one body, tells you what the whole category currently believes.
  • Buyers’ own words. Reviews, forums, comments, the support inbox. The words themselves, not the sentiment.
  • Your own files. The answer is often already in the building: an unused video, a dormant credit, a survey line nobody read.

4 · The rulebook

  • What the platform allows. Special categories, restricted verticals, and what survives in each. In health, Meta can block every lower funnel standard event, so the events get built neutral and no patient data ever leaves your server.
  • What the profession allows. Bar rules, licensing bodies, professional advertising codes, and the written consents a joint arc needs.
  • What privacy allows. The consent line on every form, and the messaging rules in every jurisdiction you contact people in.
  • One page, yours, written once. Every piece passes it before it is made, not after it is rejected.
04The arithmeticThe numbers that decide everything after

The most a new customer may cost.

The target comes out of your own gross profit, before anything runs. Dated, and never argued with afterwards. It sits on every version of your map for the whole 6 months.

WHAT A NEW CUSTOMER PAYS YOU, INSIDE THE WINDOW FIRST ORDER LATER ORDERS LESS REFUNDS AND REDUCTIONS LESS WHAT IT COSTS YOU TO DELIVER DELIVERY LESS THE PROFIT YOU KEEP PROFIT THE TARGET THE SAME TARGET, ON THE FIRST ORDER ALONE THE SHORTEST WINDOW IS THE FIRST ORDER WHAT A NEW CUSTOMER PAYS YOU, INSIDE THE WINDOW FIRST ORDER LATER ORDERS LESS REFUNDS AND REDUCTIONS LESS WHAT IT COSTS YOU TO DELIVER DELIVERY LESS THE PROFIT YOU KEEP PROFIT THE TARGET ON THE FIRST ORDER ALONE THE SHORTEST WINDOW IS THE FIRST ORDER

Model · the shape of the subtraction. You choose the window, the shortest is the first order.

Example numbers · type yours

The target

The most a new customer may cost you.

$
$
%
$
The target
$222
Gross profit per customer $312 · the profit you keep $90

The fuel

What it costs to find out. Paid to Meta, on your card.

$
The fuel, a month, on your card
$21,600
1 ad set measuring $10,800 · 2 ad sets hunting $21,600
Fuel across 6 months $129,600
Fee, starting at $8,000 USD a month Fee plus fuel, from $29,600 a month 6 months, all in, from $177,600 Fuel counted at full from day 1. Before A it is small.

Here is the fuel arithmetic, so you can run it yourself.

3 lines. No formula I keep to myself, and nothing in it you cannot check against your own account tonight.

Step 1
How many events a month
50 a week × 52 weeks ÷ 12 months = 216.7 used as 216
Step 2
What 1 ad set costs a month
216 × your cost per optimised event
Step 3
What the arc costs a month
That number × the ad sets running. 1 measures. 2 hunt.
Why 216, and the rest5 rows
Why 216Model
Meta says an ad set usually leaves learning after about 50 results in the week after its last significant edit. A guideline, never a gate. 50 a week, annualised and divided across the months, gives the multiplier.
The new number
The fuel is worked on the cost of the new event, not the one running now. Reported cost rises when the objective switches, because returning customers are the easiest people on earth to convert, and taking them out of the training label makes Meta’s figure look worse while the business gets better. I say that in writing before it happens, not after.
Measuring and hunting
1 ad set is what it costs to measure. 2 is what it costs to hunt. I quote both, always, so the number you authorise is a decision rather than a surprise.
Days to run
Fuel authorised divided by daily spend, on every deck from the launch. If the fuel runs out before the reading exists, the reading is invalid, and that is known on day 1 rather than discovered in month 4.
Affordability
If the numbers can’t carry the fee plus the fuel, I say no on the call, before anything is signed.

Your fuel decides how many ad sets it can feed.

Split past the last floor and each ad set usually stays in learning. The same rule at any size.

ONE FLOOR: ABOUT 50 A WEEK FOR 1 AD SET, SO 216 TIMES YOUR COST PER EVENT A MONTH A MONTH OF FUEL: $21,600, AT $50 AN EVENT THE MOST AD SETS IT CAN FEED: 2 A MONTH OF FUEL: $2,000,000, AT $100 AN EVENT THE MOST AD SETS IT CAN FEED: 92 EACH BAR IS 1 MONTH OF FUEL, CUT INTO FLOORS ONE FLOOR: ABOUT 50 A WEEK FOR 1 AD SET, SO 216 TIMES YOUR COST PER EVENT A MONTH FUEL: $21,600 A MONTH, AT $50 AD SETS IT CAN FEED: 2 FUEL: $2,000,000 A MONTH, AT $100 AD SETS IT CAN FEED: 92 EACH BAR IS 1 MONTH OF FUEL, CUT INTO FLOORS

Model · 216 events a month, per ad set

Then the offer gets the same arithmetic.

A business can have a perfect target, a perfect idea and a perfect machine, and still be selling a thing at a number that cannot work. So the offer is read before the idea, while the arithmetic is still on the table and nobody is attached to anything yet.

THE TARGET THE OFFER TODAY OVER THE LINE CONSTRUCTION 2 AT THE LINE, NO ROOM CONSTRUCTION 3 CLEARS IT WITH ROOM WHAT A NEW CUSTOMER MAY COST THE TARGET THE OFFER TODAY OVER THE LINE CONSTRUCTION 2 AT THE LINE, NO ROOM CONSTRUCTION 3 CLEARS IT WITH ROOM

Model · 3 constructions, one line to clear

Question 1
What is the thing a new customer buys first, and what does it cost them.
Question 2
What would have to be true for that first purchase to clear the target with room.
Question 3
What is the cheapest change to the offer that makes that true.
What you get
2 or 3 offer constructions, each with its own arithmetic against the target at the payback window. A different entry product, a different price point, a deposit or a split, a bundle that moves the first order, a risk reversal that moves the conversion rate. One recommended. The others killed in writing.
Your call
You see what the change does to the number. Then you decide. If the offer does not move, the arc runs against the offer that exists, and that goes on the map with a date.

A target nobody argues with is worth more than a forecast everybody likes.

05The arcMonth 1 into month 2

One idea. Many pieces.

3 situations come out of the market read, each with its own target, so the choice gets made with all 3 visible. 2 are killed in writing the same day, with the reason and the date. One is owned, named, and said the same way for longer than feels comfortable.

The volume
New work every week, built as genuinely distinct concepts with several executions each. Never a monthly batch. How much depends on the fuel and the market. It is set in the size-up, and it never changes the fee after.
One arc, many executions
Variety is good and the machine wants it, as long as all of it sits inside one arc. 12 pieces serving one arc is a healthy account. 12 serving 12 ideas is 12 accounts pretending to be one.
The hero
Never picked in advance. Run the variety, let the market pick, then put engagement behind what earned it, so the proof stacks on one piece rather than starting from zero on 1,000.
The hit rateModel
About 5 out of every 100 pieces are real winners. I say that in month 1 so month 3 is not a shock, and I plan the production against it rather than against hope.
100 PIECES 5 REAL WINNERS THE MARKET PICKS THEM. NOBODY DOES IN ADVANCE. 100 PIECES 5 REAL WINNERS THE MARKET PICKS THEM. NOBODY DOES IN ADVANCE.

Model · the hit rate I plan the production against

The rest of the arc6 rows
The mix
About 50% statics and moving stills, 50% video. I have not seen video win on its own.
The first 2 seconds
Branding built into the opening of the piece, never dropped on top as an overlay. A plain logo on the front gives people a reason to scroll.
Testing, honestly
Under one consolidated campaign the platform allocates the money deliberately rather than randomly, so differences between pieces are confounded by the allocator itself. Below a big budget, clean creative testing is unaffordable, not dead. What is real: which piece the machine chose to spend on, and why.
The pages
One per room the idea has, and more than one where they are worth testing against each other. The second act of the ad, not a landing page. Real original content, because a thin page that exists only to bounce people onward sits close to what the platform restricts.
The emails
3 of them, written into the arc. Never a welcome sequence.
The order
Talk it through, I state exactly what is needed, you supply the real assets, then it is made. Never the other way round. Nothing is built on an assumption about what you have.

The variety is for the machine. The message is for the market.

06The machineWhat happens between a message and a person

The race is over before the screen finishes drawing.

Everything below is Meta’s own published architecture, with Meta’s own names and dates on it. Every decision on this page is made against it. If you pay for an arc, you should see what it is aimed at.

ELIGIBLE ADS TENS OF MILLIONS RETRIEVAL A FEW THOUSAND RANKING SCORED AGAINST THE PERSON THE AUCTION BID, ACTION RATE, QUALITY 1 AD, 1 PERSON THE WHOLE RACE IS OVER BEFORE THE SCREEN HAS FINISHED DRAWING ELIGIBLE ADS TENS OF MILLIONS RETRIEVAL A FEW THOUSAND RANKING SCORED AGAINST THE PERSON THE AUCTION BID, ACTION RATE, QUALITY 1 AD, 1 PERSON

Meta’s published pipeline · counts only where Meta published them

3 places you touch it. That is all of them.

1
The creative
The machine reads it. Meta’s researchers, January 2026, on the content features read out of images and text: “not a marginal add-on but a prerequisite for effective scaling.” Their models only get smarter when they can read the work. Thin work is not a taste problem. It is a prediction problem.
2
The bid
One of the 3 components of total value, and the only one you set directly. So structure and pacing are decisions, not settings. An ad set that will not spend is a bid problem, not a mystery.
3
The conversion event
What the machine is trained to go and find. It cannot read your campaign arc. It reads this. When the 2 disagree, the event wins. Which is the entire reason the next act exists.
The 2 clocks and the 8 partsMeta’s own words

It runs on 2 clocks.

Almost nobody draws this part, and it decides what you can and cannot influence. Most of the thinking about a person finishes long before your ad is in a race for them.

Offline
Before the request exists
GEM trains on everything and teaches the smaller models. Sequence learning reads a person’s long history. LLaTTE turns that history into a cached vector, 2,048 values wide, and files it. The heavy part is finished before anyone opens the app.
Online
The request arrives
Retrieval narrows the pool. Ranking scores what survives against the cached vector and the person’s last few moments. The auction picks a winner. Your creative, your bid and your event enter here.

The parts, in Meta’s own words.

Andromeda2 December 2024
Retrieval. It narrows, in Meta’s words, “tens of millions of ad candidates into a few thousand relevant ad candidates.” The only stage of the pipeline with published candidate counts on both ends. It runs on NVIDIA Grace Hopper Superchips and Meta’s own accelerator. Meta’s published figures for it: 8% better ads quality on selected segments, 6% better recall, and 10,000 times the model capacity.
GEM10 November 2025
The teacher. It trains on everything the system sees and passes what it learns down to hundreds of smaller models that actually face the request, through knowledge distillation, representation learning and parameter sharing. Meta calls those models “student models” and GEM “the central brain.” Published: 5% more ad conversions on Instagram, 3% on Facebook Feed, and, in August 2026, training compute scaled 4 times in 12 months.
Sequence learningNovember 2024, new architecture published August 2026
It reads a person’s behaviour as an ordered history rather than a bag of attributes. Meta calls it a paradigm shift for personalized ads recommendations, and in August 2026 a core component of GEM. Published in August 2026, together with Meta’s other model changes: 6% more conversions on Instagram, 3% on Facebook, 3.5% more ad clicks.
LLaTTEJanuary 2026
The split that makes all of that affordable. Running a transformer over thousands of events at auction time is impossible, so an upstream model reads the history offline and publishes a cached vector of the person, 2,048 values wide, off roughly 400 events per source. It costs more than 45 times the online model’s sequence work, and it is already finished before the request arrives. At request time a lightweight counterpart reads only the freshest events, and that part is about 30% of the online ranking model’s work. The architecture handles histories of 500 to 5,000 events, across trillions of requests a day.
Lattice11 May 2023
One high capacity prediction architecture in place of hundreds of smaller independent models, consolidating what used to be separate systems per objective and per surface. Meta’s description: “a new model architecture that learns to predict an ad’s performance across a variety of datasets and optimization goals.” Meta published about 8% better ad quality with it.
Adaptive ranking31 March 2026
Serving infrastructure. It replaces one size fits all inference with intelligent request routing, so a high intent request gets the expensive model and a cheap one does not. It decides how much computation a request is worth, which is a different question from which ad suits which person.
The auctionJanuary 2023
3 components decide it, in Meta’s own sentence: “There are three key components of total value: the advertiser bid, the estimated action rate, and ad quality.” The bid is multiplied by the estimated action rate, ad quality is taken into account, and the highest total value is displayed. You set the bid. Your creative and your event move the other 2.
PacingMarketing API
Standard pacing enters your ad into “every relevant auction” and scales your bid across the day to spend smoothly. Which changes the whole diagnosis: an ad set that will not spend is not being held back from opportunities, it is entering them and losing them. The question is never why Meta will not show the ad. It is why the effective bid is uncompetitive.

The heavy thinking about who a person is has already finished before any advertiser competes for them. Your creative, your bid and your event are what enter. Every other lever in the account is a way of arranging those 3.

07The wiringMonth 1, day 2

Optimise on a plain purchase and the machine finds the people who already bought.

They are the highest probability buyers on earth. They know you, their card is saved, their history is visible. Reported cost falls, and much of what it counts is customers you already had. That is the retargeting trap running inside broad prospecting, driven by the training label rather than by where the money went.

THE BROWSER YOUR SERVER THE PLACE THAT KNOWS WHO IS NEW PIXEL CONVERSIONS API ONE EVENT ID COUNTED ONCE BACK INTO THE MACHINE IT LEARNS ON THIS OUT TO THE GAUGE YOU READ THIS THE BROWSER YOUR SERVER PIXEL CONVERSIONS API THE PLACE THAT KNOWS WHO IS NEW ONE EVENT ID COUNTED ONCE BACK INTO THE MACHINE IT LEARNS ON THIS OUT TO THE GAUGE YOU READ THIS

One event, sent twice, counted once

The first-time-buyer event
The centre of the whole thing. Fired from your server, after the order has been checked against your own customer history. The browser has no idea whether somebody has bought before. Only your database does. An existing customer can never produce a positive label on that event, so the model learns to deprioritise them from data rather than from a blocklist you have to maintain.
Why not an exclusion
Every audience control the platform ships answers who. None of them change what the model is trained to predict. Exclusions also keep disappearing from the product. A training label is not a control that can be taken away.
The volume lineModel
Below roughly $13,000 to $22,000 a month in fuel, for a new customer costing $60 to $100, there is not enough new customer volume for the machine to learn on that event, so it gets measured on rather than optimised on, and the value goes on the standard event instead. Measuring works at any volume. Optimising wants about 50 a week. That range is 216 times your own cost per new customer. Run it on your number, not mine.
The time it takes
Done once, at the start, on your stack’s own calendar. On a standard store it is about a day of development work. It starts on day 2 because it depends on your stack rather than on my thinking, and starting early buys 3 to 4 weeks of readable arc for nothing.
THE EVENT IT LEARNS ON: THE DEEPEST ONE WITH ABOUT 50 A WEEK ABOUT 50 A WEEK PAGE VIEW VIEW CONTENT ADD TO CART CHECKOUT PURCHASE FIRST-TIME BUYER THE ONE IT LEARNS ON 1 10 100 1,000 10,000 THE EVENT IT LEARNS ON: THE DEEPEST WITH ABOUT 50 A WEEK ABOUT 50 PAGE VIEW VIEW CONTENT ADD TO CART CHECKOUT PURCHASE FIRST-TIME BUYER THE ONE IT LEARNS ON 1 10 100 1,000 10,000

Model · one account, one week

The rest of the wiring12 rows
One pixel
Never 2. One schema and one naming convention across every page, so the data can be read back later rather than decoded.
The Conversions API
The same events sent again from your server, so they survive ad blockers, browser restrictions and everything else that has eaten browser side measurement since 2021. On every page, not just checkout.
Deduplication
The browser event and the server event carry the same event id and the same event name, so one sale counts once. Without it every sale counts twice, and every number you decide off is wrong in the same direction.
Match quality
Every identifier you legitimately hold goes with the event, hashed where Meta requires it: email, phone, name, city, state, postcode, country, your own customer id, the click id, the browser id. Low match quality means Meta cannot connect a sale back to the person who saw the ad, so the reading under-reports and you decide off a broken figure.
Value on the event
First order gross profit, or predicted lifetime value, sent with it. Then the machine can bid on what a customer is worth rather than on how likely they are to buy, which is as close as the objective gets to your P and L.
The honest cost
2 differently named events do not deduplicate against each other, so reported purchases will not equal orders. That is said before the build, not discovered after it.
Lead and case work
The raw enquiry, the qualified lead fired out of the CRM carrying the original id so it stitches back to the ad that brought it, and the closed customer sent with its value the day it closes. When that is weeks after the click, Meta will not credit it to the ad, so it gets read on your server. Optimise on the deepest event that has volume. Measure on the deepest event regardless.
Attribution, 2026
The default is 7 day click-through, 1 day engage-through, 1 day view-through. The 7 day and 28 day view windows were removed on 12 January 2026. Click-through was redefined in March 2026 to link clicks only, with everything else moved into engage-through, which on its own made reported conversions drop with nothing about the work having changed. I pre-empt that in writing rather than explain it after somebody panics.
Incremental attribution
It looks like a reporting setting and it is an optimisation control. It changes who Meta goes and finds, and choosing it locks your attribution windows. It never gets switched mid-arc to make a chart look better.
The naming
One convention on every ad, every link and every destination, so the reports read the same names your CRM uses and nobody has to reconcile 2 vocabularies at month 3.
The memory line
Search Console and analytics connected, 16 months of branded search pulled at handover before the window closes, direct traffic beside it. Meta keeps reach and hourly detail for 13 months and frequency for 6, so it gets exported monthly. Read September 2026.
Messaging registration
Before a single automated text can go to a US number, the messaging has to be registered with the carriers. Anywhere from a few days to 3 weeks. It goes in the wiring week rather than being discovered the day the desk goes live.

A perfectly wired account optimising on the wrong event is a machine seeing in 4K and still hunting people who already bought from you.

Ramin Karimi at work
08The podWho does it

You are hiring judgment, not account managers.

Everything that is not a decision gets assembled behind me by people who receive briefs rather than conversations. Who makes the pieces is set in the proposal: ORCAS, or your own studio to my brief.

Me
Every judgment. The size-up. The read. The target. The offer constructions. The roads and the 2 kills. The idea. The look and the voice. Both gate conversations. The re-brief at month 6. Inside the pod, if a decision can be argued about, it is mine.
The desk
Automation first on everything that comes back, every hour. On leads: the reply inside the minute, the qualifying questions, the booking link, the entry in your CRM. On orders: the question answered before the cart is abandoned, the note after it ships, the return handled before it becomes a review. A person follows the automation and never leads it. Every message carries your business name and an opt out, opt outs are honoured the same minute, and contact hours follow the customer’s own time zone.
Closers
The desk hands a qualified lead to a named person by a named route inside a set time, and the log is visible to you. The response time sits in the readout next to the cost per new customer, from day 1.
The rest of the pod6 rows
The assembly
Everything that is not judgment and is not my face. Your file, the stage decks, the stage emails, the map, the drift register, the market digs handed to me with the choice still open, the research and its grading, the pages. Any surface that needs you to maintain it was designed wrong.
Editor and designer
Per batch, on the weekly creative cadence. They receive briefs, not conversations, so the work arrives shaped rather than interpreted.
Development and analysis
The pixel, the Conversions API, the first-time-buyer event, the map’s wiring. Done once, at the start, on your stack’s calendar.
The pages talk back
Anybody who reads a page and asks something gets an answer the same hour, in the same voice as the piece that sent them. A page that takes a question and does nothing with it costs you the sale twice.
The training leads
The first 100 to 150 leads are training for a brand new event. That is said up front and counted down on the map, never quietly absorbed into a number.
The rule
If it is not judgment or my face, it is not my job. The first week I spend more than 5 hours on hands work, that job gets a person.
09The runMonths 2 to 6

Live, and then deliberately boring.

The launch runs 1 or 2 ad sets, broad, on the platform’s own default structure. Highest volume, no cap, through the learning window and the build window. Support channels feed Meta rather than splitting from it. And before the first number exists, I tell you how big a weekly move has to be before it means anything at all.

The week
One line, every Monday, drawn with the normal swing under it. Never a bare number.
Decides nothing
The month
One change. One reason. One date. Never twice in a month.
The valve
The cohort
The loop closes when a reading actually exists, not when the calendar says it should.
The turn
50 NEW CUSTOMERS A WEEK MINUS 22% TO PLUS 39% THE TARGET 500 A WEEK MINUS 8% TO PLUS 10% 5,000 A WEEK 3% EITHER WAY INSIDE THE BAND, NOTHING HAPPENS 50 A WEEK MINUS 22% TO PLUS 39% THE TARGET 500 A WEEK MINUS 8% TO PLUS 10% 5,000 A WEEK 3% EITHER WAY INSIDE THE BAND, NOTHING HAPPENS

Model · the least it swings, from your volume, before anything runs. A seasonal business swings wider.

The normal swingModel
Computed from your volume before anything runs, and asymmetric, because cost misbehaves upward more than downward. At 50 new customers a week it is roughly minus 22% to plus 39%. At 500, minus 8% to plus 10%. At 5,000, 3% either way. Inside the swing nothing happens, and doing nothing is the correct action.
Counted, not caused
The number is counted where the sale actually happens: on your own server, and wherever else you sell. Only a randomised test says what was caused: a holdout, a lift test or a geo test, agreed at the start or not at all. It is never introduced at month 2 to rescue a chart.
Why that line is hardProven
Attribution says what happened next to the ads. A randomised test says what happened because of them, and those are 2 different questions. A published study on Facebook’s own data checked the usual ways of estimating lift against randomised tests, across 15 experiments and roughly 500 million user observations. They mostly overstated it. A later one, across 663 experiments, found the same. So the number on your map is counted. Caused is a word I only use behind a test.
The map
Every month: the target line, the actual line, the memory line, days to run, capacity used, every change dated, and the original target on every single version. If I disappeared, the map is the handover. That is the test of whether it was ever good enough.
The mapMonth 4 · model
COST PER NEW CUSTOMER THE TARGET MEMORY 01 02 03 EVERY CHANGE, ONE A MONTH, DATED READ, MONTH 4 TODAY A B COST PER NEW CUSTOMER THE TARGET MEMORY 01 02 03 TODAY B A EVERY CHANGE, ONE A MONTH, DATED
Days to run71
Capacity used62%
Changes, all dated3
Original targetUnchanged

That is the artefact. One page, every month, rendered from your own file, with the original target on every version of it.

ESTIMATED WITHOUT A TEST MEASURED BY A RANDOMISED TEST OFF BY 3X, IN 50% OF THE STUDIES ESTIMATED WITHOUT A TEST MEASURED BY A RANDOMISED TEST OFF BY 3X, IN 50% OF THE STUDIES

Proven · 15 experiments, about 500 million observations, randomised tests as ground truth. 663 experiments later, the same direction.

The rest of the run6 rows
The random walk
Adjusting toward wherever the last result landed is a random walk, and it ruins good accounts with activity. It never gets run here.
The next dollar
Up in steps of about 20%, held through the learning window, and the new cohort read on its own rather than blended into the old one. The average still looks fine while the new dollars lose money. Per unit is what decides.
Where it stops
When the marginal number goes over the target, or capacity is full, or frequency starts climbing with the spend. Frequency climbing with spend is saturation at this idea, and the answer to that is the next arc, not more fuel.
Capacity
Set in week 1, in your words: cases you can carry, patients you can see, orders you can ship. Scaling never passes that line. Demand you cannot deliver is a complaint machine and a refund line.
The drift register
Every dated platform change that touches your account, with a screenshot. When a number moves for a reason that is not mine, this is how you know.
A valid reading
3 conditions, all required: the learning window held, the sales cycle passed, 12 or more conversions in the cohort. An invalid reading is reported as not yet. Never as a pass.
10The answersWhat comes back

Clicks can look excellent while the business gets nothing.

Everything comes back at 3 speeds. 2 of them decide anything.

Attention
Same day
Clicks, views, cost per thousand, frequency, likes. Mine to watch, and useful for explaining a result. It never decides one.
Memory
Weeks to months
Branded search and direct traffic. How the market treats the name after the work has been running. Slow, and it falls when the work stops.
Money
The sales cycle plus the learning
What a new customer costs, on a mature cohort, counted where the sale happens. The slowest one, and the only one that pays rent.

Attention is not a third answer, it is the explanation of the other 2. It moves the same day and it can be bought. So it gets watched and it never decides. What decides is what the market pays you, and what the market remembers you for.

How to get new customers. How to impact the market.

One reading each, on every map, from the launch to month 6.

The number
Read against the target line, counted where the sale happens: your own server, and wherever else you sell.
THE LAUNCH BRANDED SEARCH AND DIRECT
Market impact
Branded search and direct traffic against time, on their own line, on every map. How the market treats the name after the work has been running.

I hold the arc to the number, and I read the market beside it.

11The priceThe money

Starting at

$8,000 USD a month.

Set in the size-up, by how much the arc has to make. It never moves after.

  • The ad money runs on your own card. It never passes through me.
  • Month 1 is the fit. Not a fit, clean stop.
  • Never priced per result. A number computed after the fact can be argued into any shape, and paying per result pays for customers who were coming anyway.

Inside it

  • The size-up, the account read, the read of your numbers, the target, the offer constructions.
  • The market digs, the 3 roads, the 2 kills, the idea, the rulebook.
  • The wiring, the event, the gauge and the map.
  • The arc: every piece ORCAS makes, every page, the emails, the launch and the run.
  • The desk, the log, the readings, the monthly decks and the re-brief at 6.
  • My judgment and the whole pod. What the arc has to make is set in the size-up, and it never changes the fee after.

Outside it

  • The fuel, on your own number, paid to the platform on your own card. Quoted as 2 figures before anything is signed: what measuring costs and what hunting costs.
  • Taxes.
  • Nothing else. No setup fee, no percentage of the fuel, no cut of anything.
What lands, and when7 lines
  • A page the day you pay, with everything I need from you listed on it, and the money said once.
  • One email per stage, numbered, so your inbox becomes the record of the work rather than a thread you have to manage.
  • A deck per stage: what was done, what it means, what is next. Rendered from your own file every time, never written from scratch, so it cannot drift from the numbers.
  • A group thread for the quick decisions and for the Monday line.
  • The map, in every deck from the launch on.
  • The desk’s log, visible to you, with the response times in it.
  • The written pieces behind each stage, so you learn what is being done while it is being done.
12The way inAfter you reach out

What happens after you reach out.

01
The message
A DM or an email. I read it and decide whether there is a call. No calendar link and no form, because the first filter is me.
02
The call
About an hour. Every question I need, recorded, so nothing gets remembered wrong later.
03
The rest by email
Then I wait for your answers. The numbers matter more than the speed.
04
The size-up
The fee. The fuel worked on your own number, as 2 figures. The road with A and B dated. One line of what I would do.
05
The proposal
Then I wait again, because this is a 6 month decision and nobody should make it on a call.

None of the reading happens before money. No access, no assets, no idea. The size-up is light on purpose.