Is this for me?
This tab decides whether to build one at all, and what it costs to find out. If you read nothing else: put your six numbers in the card below and read the verdict. Some agencies get a no, and that is the page working.
You have bought leads before and watched the calendar stay empty. This page is the math that decides whether to pay again.
A video sales letter is a video that makes the whole case for a call. On this page it plays first, and under it sits one ask: an email, then three questions, then the calendar. The ad sends a person to the page, the video makes the case, the email is the first yes, the questions sort the person, the calendar books a fifteen minute phone call, and the call becomes a client or it does not.
It is for one job: booking calls that turn into clients you want. It will not get your name out, and it cannot pay for itself on a three hundred dollar project. The card below shows why.
Who it works for, and who should close the tab
It works for an agency, a consultant, or any firm that sells on a call: a first contract big enough to buy a lead, someone to take the call within a day, and a calendar with room on it. Each of these facts means no:
- Too few leads to feed the learning rate. Meta needs about 50 leads a week to learn. At the example's rates that is about 5 held calls a week on top of the work you already have. If nobody can answer about 7 new leads a day, this is not for you yet.
- Nobody to take a fifteen minute call within a day. A lead that waits a week is a lead that booked someone else.
- No calendar. If the call gets booked by email back and forth, the show rate dies.
- A client worth less than the floor the card prints. At your rates and your industry's click there is a client value under which no ad budget works. For the example agency it is about $12,308. The card below prints yours.
The math, before anything else
Six numbers in, five out, and one of three verdicts. Every output shows its arithmetic under it in your own numbers. The click and the conversion rate come from the industry row you pick on the numbers tab, so the card moves when you change it.
A lead here runs about $75. Your $300 a day buys about 28 leads a week, and the ad set needs 50 to learn.
Next: raise the daily budget to the floor, or run fewer days at that rate. Never run under it for longer.
The example agency reads below at its own numbers, and we kept it that way. A $15,000 client at a 15% book rate can pay $91 for a lead, and a lead runs about $75 across the country, so the math works. What does not work is $300 a day. It buys about 28 leads a week and Meta wants fifty. The card says raise the daily budget to about $536, or accept a slow read. Every other tab walks the example through the build, so you can see what a slow read looks like before it costs you a month.
The bar at the top of this page counts which numbers came from a study or a platform's own page, and which are arithmetic on those. We would rather you know which is which. No number here is pulled from one advertiser's ad account. An account is one offer in one market, and a rule built on it teaches you that account rather than the job.
The build order
Six steps, in the order that makes the ad, the page and the video read as one conversation. Each step names the tab that carries its detail and a time for one person, our call. The full spec, with a check per step and the example's own artifact beside each check, is the file this block is copied from.
- Lock the goal and the buyer. Run the card above with your numbers. Write down what a client is worth, the one buyer, the call and the one action. Tabs one and two. Half a day.
- Write the script from the twelve answers. Answer the questionnaire, have the script written to the beat map, read it aloud with a stopwatch. Tab four. One day.
- Wrap the page around the script. Headline, subheadline, the video, the ask that flips to the questions and the calendar, then the proof blocks, the questions people ask and the footer. Tabs three and five. Two days.
- Pull the ad angles from the page's pains. Six hooks, three bodies, each naming the buyer. Tab six. Half a day.
- Write the ads. Primary text, headline, the first three seconds, the card. Tab six. One day.
- Sign off the continuity. Read the ad, the page and the video as the owner who has been pitched by agencies before. Tab three. Half a day.
Then the launch checklist: the events, the optimization event, one ad set with its own budget and six ads, the budget from the card, the first test. It is filled in for the example agency in the spec, and the ads tab and the testing tab carry each line.
Where does a person go?
This tab decides the nine steps between the ad and the client, and which two of them carry the weight. If you read nothing else: the video plays first, Meta learns on the email given under it, you judge on the held call, and the show rate lives in the reminders.
| Stage | The one number | Where it comes from |
|---|---|---|
| Ad shown to clicked | 1% of people shown the ad click it, at the least | our call, the floor |
| Clicked to page loaded | the gap between the two is people who gave up before the page drew; no study publishes it | yours to read; Meta reports link clicks and landing page views as separate lines |
| Page loaded to watched | the host's play count over page views; no study measures it for a sales video | yours to read, on the hosting tab |
| Watched to email | 6.1% of visitors become a lead on the median professional services page; the study's lead is a form, so read it as the yardstick for an email | from a study; agencies are not a row of their own |
| Email to answered | nearly all of them, when every question is a tap and none is typed | our call; read your own drop between the two |
| Email to booked | 15% of email leads pick a time | our call |
| Booked to held | 65% of booked calls happen | our call |
| Held to the client | 25% of held calls become a client | our call |
The money chain at your numbers and your industry: an email lead $75, a booked call $500, a held call $769, a client $3,077, against a client worth $15,000. When the last two are the wrong way around, the first tab already told you.
The ask after the video, and why the email is the event
Nothing sits in front of the video. Under it sits the ask: one field for an email and one button that says what happens next, Enter your email to book a call. That email is the first yes, and it is the event Meta trains on. Meta's own fix for an ad set that cannot learn is, in its words, to "choose an optimization event that occurs more frequently." A booked call fires a few times a week at most budgets. An email fires every day. Train on the email, judge on the held call, and keep both events.
The order matters. An email asked before the video is a toll. An email asked after it is a decision by a person who has heard the whole case, and that is the one worth carrying to the calendar. Our call.
What the ask flips to
The block under the video has three states, and the person never leaves the page. First the email. Then three questions, every one a tap. Then the calendar: pick a time, then a name and a phone number, with the email already there. In GoHighLevel the email step is a form, the questions are a survey, and the calendar's own form appears, in GoHighLevel's words, "after a time slot is selected." Its Sticky Contact setting "uses cookies to store and retrieve data previously entered by a user," which is how the email reaches the calendar without being asked twice. It works on one browser and one domain, so all three live on the page's domain.
The booking layer, in plain words
GoHighLevel is the booking and customer record layer in this playbook. It holds the form, the survey, the calendar, the reminders and the list of who booked and who showed. A few of its words, translated. A pipeline is a list of people at a stage: booked, showed, closed. An opportunity is one person on that list. A workflow is an automatic step, like a text that fires an hour before the call. The pixel is Meta's tracker in the visitor's browser, and a server-side event is the same fact sent from GoHighLevel's server.
The two links that matter
Email to booked. The email is the shallow event. It fires often enough for Meta to train on, and a human can be written to. The booked call is the deep event. It is the one you count, and it fires too rarely to train on.
Booked to held. This is where the show rate lives. The lever is the confirmation at booking, a reminder the day before and a reminder an hour before. Any text in that sequence waits on the phone carriers approving the number that sends it. Until they do, every text you promised is an email.
The wiring, and where it breaks quietly
Every one of these fails without an error, and you find out a month later.
- The email goes to Meta from the server as the Lead event, once. The form under the video writes the contact and one workflow sends the event, with the form's own browser event off.
- The booked call goes to Meta from the server as the Schedule event. The CRM sends it with the visitor's name, email and phone. Watch the match quality score in Events Manager, not the coverage figure beside it. A phone number without its country code matches nobody, and coverage still reads full while those events land unattached.
- The CRM's server events carry no event id. A browser copy and a server copy of the same event cannot be matched and dropped, so send each fact once, from one place.
- The questions go on a survey, and the email is never asked twice. Sticky Contact is on for the form, the survey and the calendar, so the email typed under the video is already in the calendar's form. With a custom form on the calendar, that switch lives on the form. Map every survey answer to a custom field or the answers never reach the record.
- A partial calendar update through the API resets every field it does not send. Change one setting and the open hours and the slot length can come back at whatever the default is. Read the whole calendar, change the one thing, write it all back, then check the fields you never touched.
- Book the right calendar. A page can embed a calendar that contradicts every line above it, a video meeting where the copy promised a phone call, and nothing warns you. Embed the phone call calendar with its confirmation and reminders switched on, then book yourself through the live page and read the email that arrives.
- Three pipelines give the three rates. One for booked, one for showed against no-show, one for closed. The rates go stale the moment nobody moves people through the stages. A show rate read off a pipeline nobody has touched in months is a record of when somebody last did data entry. Move the people, or the number lies.
What does the page look like?
This tab decides every block on the page, top to bottom, on a phone, and what each one is for. If you read nothing else: the video first, one ask under it that flips to the questions and the calendar, then the blocks a cold visitor needs before they give a stranger fifteen minutes.
Build it at phone width first
Build the page at 390 pixels wide and never look at it on a laptop until every block reads there. A block that needs a second column has not earned its place. Text, buttons and the video run the full width. Then open it on a laptop once; only the empty space on either side should change.
The top: the headline, the subheadline, the video
The headline repeats the promise from the ad's headline, in more words. A person who clicked the ad should see the same idea. Ten words or fewer. A statement, never a question. The example agency's: Fifteen minutes on the phone tells you why the ads stopped booking. The subheadline says who it is for and what happens next: For owners of service companies who pay for ads and get a report back. A strategist, not a closer, and a plan whether or not you hire us. The wording follows Alex Hormozi's value equation: name the outcome, make it believable, make it fast, make it easy.
The video sits under the subheadline, as wide as the phone, with a big play button on a still of the person talking. Nothing sits between the visitor and that button. It does not autoplay with sound, because every browser blocks that. Captions are on, because most people watch with the sound off.
The ask under the video, and what it flips to
Right under the video, one field and one button: Enter your email to book a call. Not "to watch," because the video already played. Not "to get started," because the button should say what it does. Under the field, one small line says what the email is for and links to the privacy policy. The ask stays in view from the first second, our call. A cold visitor who scrolls past the video should still find the door. Hiding it until the offer beat is a test for the fifth rung, not the default.
When the email lands, the block flips, in place, to three questions. Every one is a tap. Nothing is typed, because every typed field costs people. When the last tap lands, the block flips to the calendar: pick a time, then a name and a mobile number, with the email already filled in. Three states, one record, and the person never leaves the page.
| Order | Question, every answer a tap | Why it is there |
|---|---|---|
| One | What you spent on ads last month (pick one): nothing, under two thousand, two to ten thousand, more than ten | the call is a different call at each; "nothing" is a real answer and it books too |
| Two | The one thing you want fixed (pick one): no leads, the wrong leads, leads that never book, no idea what is working | the strategist knows what to open the account on |
| Three | Who decides on hiring help (pick one): me, me and a partner, someone else | a call with the decider holds; one without is a second call |
Three questions and no more: the budget, the problem, the decider. Each earns its place by changing what the strategist opens the call with, or by sending a person who is not a fit to a page that says so kindly. The name and the mobile number are the only things anyone types, and they come last, in the calendar. Ask for a text yes there if you plan to text; the law needs it before the first message goes out. The calendar books a fifteen minute phone call, not a thirty minute video call, and it carries the confirmation and the reminders. A person who gave the email and never picked a time is still a lead, and the reminders work on them too.
Under the ask: what a cold visitor needs before they book
Everything under the calendar is for the person who watched, did not book, and kept scrolling. They are looking for a reason to trust a stranger with fifteen minutes. Each block below gives them one, and the page ends at the same ask.
| Block | What it is for | The rule |
|---|---|---|
| The proof band | three facts a stranger can check: years in business, clients served, one named result | numbers with a date and a name; no logo you were not given permission to show |
| What happens on the call | the fear of a sales call, answered in three lines | who is on it, how long it runs, what they leave with; say "no pitch unless it fits" only if that is true |
| Who it is for, who it is not for | the reader sorts themselves; the wrong ones leave and the right ones lean in | two short lists in the buyer's words; the not-for list is the one that raises the show rate |
| The proof | two or three clients in their own words, each with a name, a company and one number | the honest opinion of a real client; when a result is unusual, say what a client can generally expect |
| The person on the call | a face, a name, two lines on why they are qualified | the person who will actually be on the call, not the founder if the founder will not be |
| The questions people ask | the six objections that stop a booking, answered plainly | real questions from real calls, in the visitor's words, never a pitch in disguise |
| The ask, again | the same button, scrolling back up to the calendar | the same words as the first ask; no countdown and no second offer |
| The legal footer | the lines that keep the page and the ads running | the business name and address, the privacy policy, the terms, the results line, the line that the site is not part of Facebook |
Three of those blocks follow a rule that is not ours. The proof block follows the Federal Trade Commission's endorsement guides, revised in 2023. "An endorsement must reflect the honest opinion of the endorser." When a client's result is exceptional, the page "must make clear to the audience what the generally expected results" are. A "results not typical" line does not do that job. The questions block follows the Nielsen Norman Group: "If you're listening to your users, you know what their questions are." Ours came from the calls. Is the call free. How long is it. Who is on it. What if we are not a fit. Do I need to prepare. What does it cost after.
The footer follows the law and the platform. California requires a commercial site that collects personal information from its residents to "conspicuously post its privacy policy." The policy names "the categories of personally identifiable information that the operator collects" and who it may be shared with. Meta's advertising standards say "the products and services promoted in an ad must match those promoted on the landing page." Its review "may include an ad's associated landing page." Every marketing email you send after the first carries "your valid physical postal address" and a way out, honored within ten business days. That is the Federal Trade Commission's rule. The line that the site is not part of Facebook is practice, not law.
The continuity table
Each hook on the ads tab lands on its own headline, so the page says what the ad said. Nothing else changes.
| The hook, in the ad | The headline it lands on |
|---|---|
| If you run a service company and your ads report reads better than your phone, watch this before the next invoice. | Your ads report reads better than your phone. Fifteen minutes on the phone tells you why. |
| Your agency's report says the ads are working. Your calendar says they are not. One of them is wrong. | The report says it works. The calendar says it does not. Fifteen minutes settles it. |
| Three things a new agency finds in the first hour inside an ad account the last one ran. | What a new agency finds in your ad account in the first hour, in a fifteen minute call. |
| You do not need a new agency to find out why the ads stopped booking. Fifteen minutes on the phone will tell you. | Fifteen minutes on the phone tells you why the ads stopped booking. |
| The ad budget that books nothing costs more than the agency fee does. Here is how to read yours. | The ad budget that books nothing costs more than the fee. Fifteen minutes reads yours. |
| If your phone goes quiet every slow season, your ads are talking to the wrong people. | Your ads are talking to the wrong people. Fifteen minutes tells you who to talk to instead. |
What goes in the video?
This tab decides how long the video is, what it covers in what order, and how to write it. If you read nothing else: four to eight minutes, six beats in a fixed order, one person talking to one person, and a script written from twelve answers.
How long
Wistia's own data, from 13 million videos, puts the band for a video that has to convert at 5 to 30 minutes, and says most viewers drop off around the halfway point. Vidyard's report, from 940,000 videos, has 65% of viewers finishing a video under a minute and 20% finishing one over twenty minutes. Our call for a call-booking video: 4 to 8 minutes. The example runs six.
What must be covered, in this order
- The hook. Name the buyer and the pain in the first sentence. Nothing comes before it, not even a greeting.
- The problem. What is going wrong right now, in their words. The thing they already suspect.
- The cost. What it costs them in dollars if they do nothing, and what it costs them personally. The nights, the payroll, the tab they keep checking.
- The proof. Why they should believe you: the years, the clients, one story of a company like theirs, and the number that story produced.
- The offer. One plain sentence. What the call is, how long, who is on it, and what they leave with. No price on the video.
- The ask. Say exactly what to do next, once. Then stop talking.
The tone
One person talking to one person, at a table rather than on a stage. Use plain words and keep the sentences short. Leave the music off under the voice and keep the captions on. Say who it is for and what you do out loud in the first ten seconds. If a line would make an owner roll their eyes, cut it.
Two things the camera should show. The person who will be on the call, so the face on the page is the voice on the phone. And nothing else: no logo wall, no office tour, no screen full of dashboards. Everything else in the frame competes with the only thing that has to land.
How to write it
Do not write the script first. Answer the twelve questions first, then have the script written from the answers, then read it out loud with a stopwatch. Cut every sentence over fifteen words. Cut every claim you could not back on the phone. The prompt below asks the questions, waits, writes to the beat map, and checks its own work. Then record the video in one sitting, in the same shirt and the same room, so the six ad hooks you shoot afterward match the video they land on.
The questionnaire prompt
VSL script questionnaire, version two, September 9, 2026 You are writing a call-booking video sales letter for a marketing agency, a consultant, or a firm that sells on a call. Before you write a single word of script, ask me the twelve questions below, one at a time, and wait for each answer. Do not guess an answer. Do not skip one. One. What is the company, what do you sell, to whom, and how long have you done it? Two. Who is the one person this video talks to? Describe their business and their week. Three. What is the problem they already suspect they have, in their own words? Four. What does it cost them in dollars if they do nothing for a year? Five. What does it cost them personally: the nights, the payroll, the tab they keep checking? Six. Why should they believe you? Years, clients, one story of a company like theirs. Seven. What is the one number that story produced? Eight. What is the call? How long, who is on it, and what do they leave with? Nine. What is the one thing you will never do on that call? Ten. What is the single action you want them to take at the end? Eleven. How long should the video run, in minutes? Four to eight. Twelve. What words does this company never use in front of a client? When you have all twelve answers, write the script to this beat map, and print the beat name in capitals before each paragraph: HOOK, THE PROBLEM, THE COST, THE PROOF, THE OFFER, THE ASK. The hook is the first sentence and names the buyer and the pain. The cost comes before the proof. The offer is one sentence. The ask is said once, then the script ends. Write for one person talking to one person at a table. Sentences under fifteen words. No music cue, no logo, no greeting, no price. Then run this check and print it: the promise is inside the first eighteen seconds; there is exactly one ask; the length lands inside the band I gave you at a hundred and thirty words a minute; no price appears; none of the banned words appear. If any line fails, fix the script and print the check again. Written from the SwarmSystem Call Booking VSL Playbook, swarmsystem.ai/playbook/vsl.
The example agency's twelve answers
The example agency, fictional, labeled example One. [your agency name], a marketing agency that runs ads for service companies. Eight years, twelve clients on retainer. Two. The owner of a service company doing one to three million a year. Runs the crew and answers the phone. Has paid two agencies and got a report back from each. Three. "The ads are spending, the report says it is working, and the phone is not ringing." Four. A retainer plus an ad budget that books nothing runs past fifty thousand dollars a year, and the slow season it was meant to fix comes anyway. Five. The nights reading a report they do not believe, the payroll covered from a quiet month, the tab they keep checking. Six. Eight years, twelve retainers, and the plumbing company in Fort Worth whose ads spent the same and booked three times the calls after one fifteen minute call changed what the ad said. Seven. Three times the booked calls on the same spend. Eight. A fifteen minute phone call with the strategist who would run the account. They leave with the one thing wrong with their ads, in plain words, and a plan whether or not they hire us. Nine. Nobody sells anything on the call. Ten. Book the fifteen minute call. Eleven. Six minutes. Twelve. Guaranteed, free, scale, ROI, funnel.
The script the prompt wrote for the example agency
The example agency's script, six minutes, September 9, 2026 HOOK If you run a service company and your ads report reads better than your phone, this is for you. I run an agency. I have run ads for companies like yours for eight years, and I am going to tell you why the phone went quiet while the report stayed green. THE PROBLEM You already suspect it. The report comes every month. Reach is up. Clicks are up. The word impressions is in there somewhere. And the phone rang less this month than last. Most ad accounts I open are talking to the wrong people. The ad says the right words to a crowd that was never going to call. From the report, you cannot tell. From inside the account, you can, in about an hour. THE COST Here is what doing nothing costs. A retainer plus an ad budget that books nothing runs past fifty thousand dollars a year. The slow season it was meant to fix comes anyway. That is the dollar number. The other cost is the nights. Reading a report you do not believe. Covering payroll from a quiet month. Checking the tab one more time before bed. THE PROOF Eight years. Twelve companies on retainer today. We answer the phone, and we send calls, not reports. Last spring a plumbing company in Fort Worth called us because their spend had not changed and their calls had halved. Fifteen minutes on the phone. We opened the account together. The ad was talking to homeowners two counties away. We changed what the ad said and who it said it to. Same budget. Three times the booked calls the next month. That call did that. THE OFFER So here is what I am offering. A fifteen minute phone call with the strategist who would run your account, not a closer, where you tell me what you are seeing and I tell you what it means. You leave the call knowing the one thing wrong with your ads, in plain words, and a plan whether or not you hire us. Nobody sells you anything on that call. THE ASK Under this video there is a calendar. Answer its three questions and pick a time. That is it. Book the fifteen minute call.
Where does the video live?
This tab decides where to host the video, what each host tells you, and how to read the drop-off curve. If you read nothing else: pick the host that gives you a retention graph and fires events, send four watch-depth events to your pixel, and do not change the video before two hundred plays.
One thing first. No independent study ranks these hosts for this job. So this tab compares them on what each one publishes about its own product, and the last column links the page each row was read from. Read them as claims a company makes about itself, checked against its own documentation.
The hosts, on their own published capabilities
| Host | What it reports natively | What it is good for here | Where we read it |
|---|---|---|---|
| Wistia | a heatmap per viewer and an engagement graph across all viewers, rewatches marked | the drop-off curve at its clearest; the graph is the product | Wistia's own help page on heatmaps |
| Vimeo | a retention graph, plays, finishes, and a completion rate that counts a finish inside the last two seconds; the full engagement report is gated to the paid plans | a clean player with no other videos around it | Vimeo's own analytics page, which says "a sharp drop at any point indicates that many viewers stopped watching there" |
| YouTube | an audience retention report with the dips and spikes marked, once the video has at least a hundred views and runs a minute or more | the report is good; the player pushes other videos, which is our call against it for a sales page | YouTube's own help page |
| GoHighLevel hosted video | a video up to four gigabytes, and a workflow that fires when a viewer reaches a completion percentage you set; turn the progress bar off so the tracking is honest | the watch-depth event lands in the same system as the form and the calendar | GoHighLevel's own help page |
| Your own file | nothing, until you send events yourself | total control, no curve unless you build one | our call |
Our call: host where the drop-off curve is easiest to read, and fire the watch-depth events to the pixel from wherever the video sits. The curve is the whole reason to care where it lives. Two things decide it for most owners. If the calendar and the form already live in GoHighLevel, its hosted player keeps the watch-depth event in the same place as the booking, and one workflow can text the people who watched the offer and did not book. If you want the clearest picture of where people leave, Wistia's per-viewer heatmap is the one built for that job, and you send its events to the pixel yourself.
How to read the curve
The curve is the share of viewers still watching at each second. Every video slopes down; that is normal. What you read is where it bends. A cliff in the first ten seconds means the hook did not name them, or the promise was not worth the time. Fix the first sentence. A slide through the middle means they stayed for the promise and the video stopped keeping it. Cut the middle, or move the proof up. A cliff at the offer means they were with you until you asked. The ask is wrong, or too big, or too early. The plays that reach the end are the number that matters, because those are the people who saw the ask. The number on the axis matters less than the bend. Fifty viewers who all leave at the same second are telling you one thing. Fifty who drift away one at a time are telling you something else, and the fix is different for each.
How many plays before you change anything
YouTube will not highlight moments on its report until a video has at least 100 views. Our floor is 200 plays before the curve is allowed to change the video. Under that, a dozen bored people can move the line. Read the curve on the same day each week, with the same window, and change one beat at a time. Write the play count down every time you read it. A curve read against last week's count is a trend, and a curve read once is a guess.
The setup list
SETUP, in order One. Send a play event to the Meta pixel and to analytics when the video starts. The play button is the first yes. Two. Send four watch-depth events, at a quarter, a half, three quarters and the end. Same names in the pixel and in analytics. Three. Put tracking parameters on every ad link: the source, the campaign, and the ad name, so a play can be tied to the ad that bought it. Four. Know the difference between a platform play and a page play. The host counts a play when the video runs. Meta counts a video view when its own ad plays for three seconds. They will never match, and they are not supposed to. Five. When three counters disagree, trust the host's play count. It is the only one that counts the video on your page actually running. Our call. Six. If the video sits in GoHighLevel, turn the progress bar off and set one workflow at seventy five percent watched, so the people who saw the offer get a text or a call. Seven. Write down the date and the play count before every change to the video, so the next curve has something to be compared to. Eight. The ask. The email given under the video sends the Lead event once, from the server. The play event fires before it, so read plays against emails to see how many watched and walked.
How do I run the ads?
This tab decides the campaign shape, whose budget it is, the creative and the test order. If you read nothing else: one campaign, one ad set with its own budget, three to six hooks inside it on day one, and a daily budget that buys fifty email leads a week.
The setup, spelled out
- One campaign on the leads objective. Every extra campaign splits the leads Meta learns from.
- One ad set, with its own budget. The ad set is the box holding the audience, the budget and the ads. Meta learns per box, so the money sits on the box.
- One optimization event, the one that fires often enough to train. Pick the email given under the video; a booked call fires too rarely. Judge the business on held calls anyway.
- Advantage+ placements on. Meta's own engineering post says Advantage+ automates "dynamic placement across Meta surfaces." We let it.
- No interest stacking. Layered interests shrink the audience and starve the learning. The creative does the targeting now.
- Three to six ads inside the one ad set, all on day one. Meta's own list of edits that restart learning includes "Adding a new ad to your ad set," so a hook added later restarts the week for the whole set.
Whose budget: the ad set's or the campaign's
A campaign budget, which Meta calls Advantage+ campaign budget, is one pot at the top. Meta moves it between the ad sets as it sees fit. An ad set budget is a pot per box. Each ad set spends its own and nothing moves. Meta's own page says the campaign budget is for when "you value all ad sets the same," and the ad set budget is for when "you want to control the amount spent on each ad set."
Our call: ad set budgets for everything here, never a campaign budget during a test. The ladder on the testing tab is the reason. Every rung changes one thing and holds the money still, and a campaign budget cannot: it chases yesterday's best box and starves the rest before their week is up. Move to a campaign budget only when several proven ad sets run at once and any of them may take the money.
Meta also offers ad set budget sharing, which lets up to 20% of a set's daily budget drift to a sibling. Leave it off while testing.
Why the creative is the targeting now
Meta picks ads with a system it calls Andromeda, described on its engineering blog in December 2024 and still the newest account of how this works. In its own words, the first step is "selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates" for each person. Our reading: the ad decides who sees it by what it says. So name the buyer, on the screen and in the first line.
What to make
| Piece | The rule | Grade | Where |
|---|---|---|---|
| The ad video | 30 to 90 seconds. Name the buyer in the first three seconds. One promise, one ask. | our call | Wistia's data has videos under a minute holding a 52% engagement rate |
| Statics and images | A flat color card. Big condensed white type that fills the frame. A small line under it that names the buyer and the place. No logo, no photo, no stock scene. | our call | big type on a flat ground is legible at thumbnail size in a scrolling feed and a photo scene is not; no study publishes a static against video split for lead campaigns with a method behind it |
| Primary text | 50 to 150 characters. The first line is the hook. Name the buyer, name the pain, stop. | from the platform | Meta's ad spec page |
| Headline | 27 characters. It is the promise the page's headline repeats. | from the platform | Meta's ad spec page |
| The frame | 4:5, tall. Captions on. Sound on but never needed. | from the platform | Meta's ad spec page |
The example agency's six hooks and three bodies
Six hooks over one body is the launch. A hook is the first line and the first three seconds. A body is everything after it. Written for an agency that runs ads for service companies, talking to their owners.
HOOKS One. If you run a service company and your ads report reads better than your phone, watch this before the next invoice. Two. Your agency's report says the ads are working. Your calendar says they are not. One of them is wrong. Three. Three things a new agency finds in the first hour inside an ad account the last one ran. Four. You do not need a new agency to find out why the ads stopped booking. Fifteen minutes on the phone will tell you. Five. The ad budget that books nothing costs more than the agency fee does. Here is how to read yours. Six. If your phone goes quiet every slow season, your ads are talking to the wrong people. BODY ONE, the proof [your agency name] has run ads for service companies for eight years. Twelve of them are on retainer today. The call is with the strategist who would run your account, not a closer. Fifteen minutes, your account on the screen, then a plan whether or not you hire us. BODY TWO, the cost A retainer plus an ad budget that books nothing runs past fifty thousand dollars a year, and the slow season it was meant to fix comes anyway. Both send a warning first: the cost per lead climbs and the calls do not. A fifteen minute call with a strategist reads the warning. Book one and leave with a plan, or get told your ads are fine. BODY THREE, the burn You have paid an agency that sent a report instead of a phone call. This is the other way around. Fifteen minutes on the phone first, the strategist on the line, the one thing wrong with your ads in plain words before anyone asks you to sign. Pick a time that suits you.
Hooks are ads in one ad set, never one hook per ad set
The arithmetic settles this one. Meta's learning line is about 50 leads per ad set per week. Put eight hooks in eight ad sets and each has to reach fifty on its own, so you buy eight learning phases at once and finish the month with most of them unreadable. Put the same eight in one ad set and the set needs fifty, once. Separate ad sets cut the pile into pieces too small to read.
The budget, and the floor Meta sets
Meta's learning line is about 50 leads per ad set in the week after your last big change. It is a rate, not a total, so the daily budget is fifty times what a lead costs, over seven days. That is the whole calculation, and it is the one that decides whether you can afford to run at all.
Your $300 a day buys about 28 leads a week here. At that pace it reads 1 ad a day, so run three anyway and expect a slow read.
Never edit the budget while it runs. Meta's own page says a jump from a hundred a day to a thousand can restart learning. If the money runs short, cut days, never the daily rate.
A daily budget is really a weekly one. Meta's own words: a daily budget sets "a daily spending target over the course of a calendar week," Sunday to Saturday, and across that week it will not spend more than 7 times that number. The day itself is elastic. Meta is rolling wider flexibility out to some accounts, where a good day can run 75% over the daily number. So read the week, never the day, and never chase a Tuesday.
The shallow and the deep event. Meta trains on the email because it fires often. You judge on the held call because it becomes money. Point the pixel at the deep one and the campaign never trains. Judge on the shallow one and you scale ads that make emails, not clients.
Refresh every two to three weeks. Meta's own status says when: "Creative limited" means the ad costs more per result than your past ads, and "Creative fatigue" twice as much. When one person has seen the ad about 3 times and results dip, the ad is tired. Make a truly different ad and keep the old one running while it learns.
What numbers am I aiming at?
This tab decides what a good number looks like at each step, from the click to the client, and which numbers you can trust. If you read nothing else: the cost you aim at is the cost of an email, the booked call and the client are built down from it, and a click rate is only a floor.
Every number, with its grade word beside it
Two kinds of numbers live here now, and the third kind was removed. From a study means a platform's own page, or a dated report that states how much it counted. Every target, floor and reason comes from one of those, published since January 2025. Our call means arithmetic on a study's number, or a starting default you replace, with the reason in the mark. Nothing here is read out of one advertiser's ad account. Where the research came back empty we say so instead of borrowing a number.
| Number | Value | Grade | Where |
|---|---|---|---|
| Link click rate, the floor | your industry's median, not a flat number. Business to business on Facebook is 0.8% | from a study | hold a hook to the median of the row you picked below; a flat floor punishes the industries that click least |
| Link click rate, business to business advertisers on Facebook | 0.8%, all lead campaigns 2.6% | from two studies | Metadata, calendar year 2025; WordStream by LocaliQ, September 15, 2025 |
| Cost per click | business to business $1.95; all lead campaigns $1.92; per industry in the picker above | from two studies | pick your row above and this changes with it; no study publishes a click cost by geography, so nothing here does either |
| Cost per thousand times shown | $16 for business to business | from a study | the business to business study. Meta says only that an audience that is "too specific" may cost more, with no figure attached, and no study measures a click cost by geography at all |
| Clicks that load the page | no number | none found | Meta reports link clicks and landing page views separately, so the gap is yours to read; nobody publishes a benchmark for it |
| Three second plays that reach fifteen seconds | no number | none found | Meta calls the fifteen second mark a ThruPlay and reports both; no study publishes the share for a sales video |
| Clicks that give an email, the yardstick | 2.6%, all lead campaigns 7.7% | from two studies | the study's lead is a form; nobody measures an email asked after a video, so the study is the yardstick until your page clears fifty emails a week |
| Cost per email, the number you aim at | $75 at the study's click; the study's own form lead $145; all lead campaigns $28 | our call, from two studies | the click over the click to lead rate; the business to business study; the WordStream study |
| Any page, visitors who convert | 6.1% median for professional services pages, all industries 6.6% | from a study | Unbounce, the 2024 edition; agencies are not a row of their own |
| Where viewers drop | around the halfway point of the video | from a study | Wistia, July 1, 2025, thirteen million videos |
| Emails that book a call, and what a booked call costs | the closest study measures a warmer starting point and reads 62% median; the example agency runs 15%, so a booked call runs $500 | from a study, then our call | the study counts a qualified demo request reaching a calendar, which is warmer than an email typed under a cold video; expect a lower number and measure your own |
| Booked calls that happen, and what a held call costs | no trustworthy public number exists; the example agency plans on 65%, so a held call runs $769 | nobody publishes it | read the block under this table before you trust any show rate, including your own |
| Held calls that close, and what a client costs | the published numbers count a narrower population; the example agency plans on 25%, so a client costs $3,077 against a client worth $15,000 | our call | published win rates count an opportunity, created only after qualification, so a chunk of held calls never enter their denominator and every published figure runs high |
| How often one person sees the ad before it tires | measured median 2.5 for business to business; recall peaks near 2 a week | from two studies | Databox's business to business accounts; Brainlabs on Meta brand lift studies, UK market. Nobody publishes the point at which an ad is spent, only where the middle of the market sits |
| Emails an ad set needs each week to learn | 50 | from the platform | Meta's own help page |
| Time on page, bounce rate | no number | none found | no study measures these for a video page; read the drop-off curve on the hosting tab instead |
The show rate nobody publishes
We went looking for a show rate and found nothing we would stand behind. The companies holding this data are the scheduling tools. They publish how many leads reach a calendar, then stop. The best of them counts over a million form fills, and none carries the funnel one stage further to who turned up.
The one real dataset we found runs 6,428 meetings, and it is one week from December 2024. Too thin to plan on, and still worth reading for what its two numbers do side by side.
It reports a 6.5% no-show rate and a 76.1% completion rate. Those do not add up, and the gap is the lesson. About a fifth of booked calls were moved or called off rather than missed. So the share that actually happened was seventy six in a hundred, not ninety three.
A no-show rate and a show rate are different numbers. Put yours next to somebody else's and you will read yourself as healthier than you are. Count it one way: calls that happened, over calls booked, for the same set of bookings. A call moved twice and never held is not a win.
Two more things we could not source, said plainly rather than filled in. Nobody has measured what reminders do to a sales call; the studies people quote are clinic appointments booked weeks ahead, a different person with a different reason to turn up. And nobody publishes what share of held calls close, because published win rates count an opportunity, which a salesperson creates only after deciding the call was worth something.
Pick your industry
Two studies measure the ad layer by industry with a stated sample. One is a business to business benchmark, and that is the agency row. The other lists lead campaigns by consumer industry. The module runs the industry you pick from the click down to the client: the click and the email rate come from the study, the cost of an email is built from them, and everything under it from the rates you set on the first tab.
The click rate rule
We hold every hook to a 1% link click rate. Under that, the hook did not earn a look. That is the whole job a click rate does. It is a floor and a check on the hook. It never ranks a creative and it never decides what to scale.
Here is why. What a lead costs is two numbers: what a click costs, over the share of clicks that give an email. The hook wins the click. The page and the offer win the email. Rank the industries in the picker above by click rate, then rank them again by the share of clicks that convert, and the two orders barely resemble each other. Physicians hold the fourth best click rate in that study and its second worst conversion rate, 4.5%.
Now the part most pages leave out. Rank those same industries by click rate against cost per lead and they do line up somewhat, the clickier rows tending to be cheaper. That is a fact about markets, not about ads. Dentists click badly and pay dearly because dentistry is an expensive auction, and that tells you nothing about which of your six hooks brings the cheaper lead in your account. Nobody has published that. So the rule rests on the arithmetic: a click rate with no conversion rate beside it is half a number.
A click rate tells you the hook earned attention. It never tells you the creative made money. So it never decides what to scale.
Cost per email ranks the ads. Cost per held call decides whether to spend more, because the held call is the one that becomes a client. When a hook clears the floor and still carries a high cost per email, keep it out of the winner's seat and use it to learn who it pulled in. That is the only thing a high click rate is good for.
The sources, and which ones count
45 of the 48 sources below carry a dataset, or are a platform's, a regulator's or the law's own page. A named source is not always a usable dataset. Three of them are people or a research group whose method we borrow and whose numbers we do not print.
| Source | Class | What this page takes from it |
|---|---|---|
| Meta Business Help Center, About learning limited | the platform's own page | fifty events per ad set per week; the five causes; change the optimization event to one that fires more often |
| Meta Business Help Center, About the learning phase | the platform's own page | wait to edit; avoid high ad volumes; use realistic budgets |
| Meta Business Help Center, Significant edits and learning phase | the platform's own page | the edits that restart learning, including adding an ad |
| Meta Business Help Center, About campaign budgets and ad set budgets | the platform's own page | when the campaign budget fits and when the ad set budget fits; budget sharing at a fifth |
| Meta Business Help Center, About daily budgets | the platform's own page | a daily budget is a weekly target; three quarters over on a good day, seven times the daily number as the weekly ceiling |
| Meta Business Help Center, Best practices for A/B tests | the platform's own page | seven days at the least, thirty at most, one variable |
| Meta Business Help Center, Creative fatigue recommendations | the platform's own page | Creative limited under twice past cost, Creative fatigue at twice |
| Meta Business Help Center, Link clicks and landing page views | the platform's own page | what a landing page view is |
| Meta Business Help Center, About ThruPlay | the platform's own page | fifteen seconds; a full play is most of the video |
| Meta Ads Guide, video ad specs for the Facebook feed | the platform's own page | primary text and headline length; the frame |
| Meta Transparency Center, the Advertising Standards | the platform's own page | the ad must match the landing page; the review may include the landing page |
| Meta Engineering, Andromeda, December 2, 2024 | the platform's own page | what the retrieval engine does; what Advantage+ automates |
| Metadata, business to business advertising benchmarks, calendar year 2025 | a dated study, one hundred and fifty three business to business advertisers, fifty eight million dollars of spend, two hundred and eleven thousand leads; the Facebook rows come from seventy one advertisers | the agency and business to business row in the industry table; the business to business click rate, click cost, click to lead and cost per lead |
| WordStream by LocaliQ, Facebook Ads Benchmarks, updated September 15, 2025 | a dated study, seven hundred and twenty six lead campaigns, medians | the consumer rows in the industry table |
| Wistia, State of Video data, July 1, 2025 | the vendor's own data, thirteen million videos | five to thirty minutes to convert; the halfway drop |
| Vidyard, Video in Business Benchmark Report, data to December 15, 2024 | a dated study, nine hundred and forty thousand videos | who finishes a short video and who finishes a long one |
| Unbounce, Conversion Benchmark Report, the 2024 edition | a dated study, forty one thousand pages | the median page conversion, all industries and the professional services row |
| YouTube Help, key moments for audience retention | the platform's own page | a hundred views before the report reads; dips and declines |
| Vimeo Help, analytics questions | the platform's own page | the retention graph; what a sharp drop means; plan gating |
| GoHighLevel Help, hosted video and viewer progress | the platform's own page | a workflow fires at a completion percentage; turn off the progress bar |
| GoHighLevel Help, Sticky Contact in forms, surveys and calendars | the platform's own page | cookies carry the email from the ask to the calendar; one browser, one domain; the switch lives on a custom form |
| GoHighLevel Help, Adding custom forms to calendars | the platform's own page | the form appears after a time slot is selected; answers save to the contact when the fields are mapped |
| Federal Trade Commission, the Endorsement Guides, what people are asking | the regulator's own page | the honest opinion rule; say what a client can generally expect |
| Federal Trade Commission, the CAN-SPAM Act compliance guide | the regulator's own page | the postal address and the way out in every marketing email; ten business days |
| California Business and Professions Code, section twenty two thousand five hundred seventy five | the law's own text | the privacy policy, posted conspicuously, and what it must name |
| Evan Miller, sample size calculator and sequential A/B testing | a published method with the math shown | the sample size arithmetic on the testing tab |
| Meta Business Help Center, About performance goals | the platform's own page | the budget rule the seventh rung runs on: at least ten times the average cost of the goal, daily |
| Meta Business Help Center, About performance goals for lead ads | the platform's own page; the comparison is Meta grading its own feature, with no sample published | what volume optimization targets, in Meta's words; the quality goal's claimed gain; the server feed now required for it |
| Meta for Developers, Conversion Leads integration | the platform's own page | the only published gate for a deeper event: two hundred leads a month, a daily upload, a stage between one and forty percent |
| Google Ads Help, your guide to Smart Bidding | the platform's own page | the machine sets the bid, you set the goal |
| Meta for Business, Advantage+ audience | the platform's own page | what the machine may overrule, and the two things it may not |
| RevenueHero, inbound benchmark, restated February 2026, over one million form fills | a dated report with its count stated; the vendor's own customers, so read it high | the closest published number to an email becoming a booked call |
| RevenueHero, no-shows benchmark, December 13, 2024 | a dated report with its count stated; six thousand four hundred meetings, but only one week of data, and older than the rest of this page | the only real show rate figures we found, and the gap between them that makes the point |
| Databox, Facebook ads frequency, updated February 2026 | a dated report with its count stated | the measured middle of the market on how often one person sees an ad |
| Brainlabs Digital, forty six Meta brand lift studies | a published analysis with its method shown; United Kingdom market | where recall peaks against how often the ad runs |
| Gartner, May 11, 2026, four hundred and two marketing chiefs | a dated survey with its count stated | how much of this job software runs today, and the forecast |
| The CMO Survey, Duke University with the American Marketing Association and Deloitte | a dated survey with its count stated | a second, independent read on the same share |
| Klaviyo, consumer research, fielded December 2025, eight thousand buyers | a dated survey with its count stated | what visible machine-written marketing does to trust |
| Federal Trade Commission, crackdown on deceptive claims about software, September 2024 | the regulator's own page | there is no exemption for software, and the case punished for a claim nobody tested |
| European Commission, the artificial intelligence act | the law's own summary | what the rules actually aim at, and what they leave alone |
| TheAgentCompany, a benchmark of agents on real workplace tasks, revision three, September 2025 | a benchmark with its environment and scoring published | how much of a digital employee's work the best agent finishes on its own |
| Towards a Science of AI Agent Reliability, Princeton, February 2026 | a paper with its metrics and models stated | that rising capability has not brought matching reliability |
| GoHighLevel marketplace documentation, the LeadConnector server | the platform's own page | what a CRM opens to software: the operation count, the areas, and the scopes that fence it |
| GoHighLevel marketplace documentation, the workflows reference | the platform's own page | the boundary: one read only call and nothing that writes a workflow |
| Anthropic, Introducing the Model Context Protocol, November 25, 2024 | the publisher's own page | the open standard the tools hang off, and its date |
| Nielsen Norman Group, FAQs Still Deliver Great Value, Susan Farrell, December 21, 2014 | method, cited by name; guidance, no dataset published | the questions block: real questions, in the visitor's words |
| Russell Brunson, Dotcom Secrets, the Perfect Webinar script | method, cited by name; their claim, no dataset published | the beat order the video borrows |
| Alex Hormozi, the value equation, from his book on offers | method, cited by name; their claim, no dataset published | the offer test the page applies |
Left out on purpose: every blog that repeats a number without the study behind it, every undated page, and every number from a forum. Numbers measured as of September 10, 2026.
What do I test next?
This tab decides what you change, in what order, and when you have seen enough to move. If you read nothing else: one change at a time, judged on cost per email, one full week each, and never a classic split test under fifty emails a week.
The ladder
Start at the top and climb. Each rung changes one thing and holds everything else still. The rung that matches your creative stage on the ads tab is lit, in the drawing and in the table. Every rung runs at least 7 days, and no ad is judged under 2,000 times shown or $600 spent.
The table under the drawing carries the rules behind each rung. Its layer column says whose week it is. The ad rungs belong to whoever writes the ads, the page rungs to whoever built the page, and the booking rungs to whoever owns the calendar. One owner per rung, or the same change gets made twice and read once.
| Rung | Layer | You change | You hold still | Judged on | Enough data | Move on when | Iterate when |
|---|---|---|---|---|---|---|---|
| Hook | ad | the first three seconds and the first line | the body, the page, the budget, the audience | cost per email; hold rate as a check; click rate only as the floor | one full week out of learning, with the ad set clearing fifty emails | one ad's cost per email sits clearly below the set average and holds three days | every ad sits in a narrow band |
| Body and offer | ad | everything after the hook | the winning hook, the page, the budget | cost per email, same rules | one full week out of learning | one body clearly below the set average, holding three days | every body sits in a narrow band |
| Headline and subheadline | page | the two lines above the video | the ads, the video, the budget | cost per email over a fixed window | one learning cycle at an unchanged budget | clearly lower and holding | inside the noise band |
| Video opening | page | the first thirty seconds | everything else | the drop-off curve at thirty seconds, plus cost per email | the view floor on the hosting tab | the curve flattens earlier and cost per email does not rise | the curve does not move; the promise is wrong, not the delivery |
| The ask and the questions | booking | the ask's words, the three questions, the calendar; last, whether the ask hides until the offer beat | the ads, the page, the video | emails over visitors, then bookings over emails | not testable at this volume | fix by inspection, recordings, and what people say on the calls | same |
| Show rate | booking | the confirmation and the reminders | everything else | held calls over booked calls | not testable at this volume | fix with confirmation and reminders | same |
| The event you optimize on | ad | the event Meta trains on | the creative, the page, the audience | cost per held call over a month, never cost per email, because the point of moving is to buy a different person | the deeper event, on its own, clears Meta's budget rule for a month first | cost per held call falls and stays down through a second month | go back. The clock reset when you changed it, and a rarer event restarted it against worse odds |
The seventh rung: pointing Meta at a deeper event
Someone always asks why you train Meta on an email when you want a held call. Meta publishes the arithmetic that answers it.
Meta's own rule: "In general, your daily budget should be at least 10 times the average cost of your performance goal."
Run it on a booked call before you touch the setting. At the example's rates a booked call costs $500, so training Meta on it needs $5,000 a day. That works out near 70 events a week, above the fifty Meta wants. Most people reading this are nowhere near it, which is why the email stays the event.
Meta's own advice runs the same way. When signal is thin its help page says to pick an event that happens more often, and gives moving from purchases to add to cart as the example. It also lists a rare optimization event as a reason an ad set goes learning limited. There is no matching instruction anywhere to go deeper.
The one published exception is a lead ads feature with strict gates: at least 200 leads a month, a daily upload from your own records, and a stage that converts between 1% and 40%. That band is the part nobody quotes. A stage almost nobody reaches is out, and so is one almost everybody reaches.
Does it pay? Meta says its quality goal cut cost per quality lead by 21% on its own forms. Read that as what it is. Meta grading its own feature, with no sample and no method published. We found no independent study that measures this, so the published record does not settle it.
Two traps. Changing the event resets the learning clock, which then runs against a rarer event. And since April 2026 the quality goal needs a live server feed from your records. Both dates have passed.
Why you cannot split test at this volume
A classic split test sends half the people to one version and half to another and waits for the difference to be sure. Under about 50 emails a week per ad set, do not run one. That is Meta's learning line, and under it the ad set itself is not stable enough to read. Here is the arithmetic in your numbers.
At your budget and your industry you get about 28 emails a week. A headline test that could tell a 20% lift apart needs about 394 emails per side. That is about 29 weeks. The baseline is the study's 2.6% click to lead rate, which means about 15,123 visitors. Move the daily budget above the $536 floor and the weeks fall, but they never fall to one.
The method is Evan Miller's sample size calculator, which prints the same number for the same inputs. Nobody running an agency waits that long. So test the way the ladder says: one change, one clock, judged on cost per email, and read the result on the same day of the week each time. Evan Miller's sequential method is the honest version of this when you insist on a formal test: it lets you stop early on a clear winner instead of waiting out the full count.
The clock, and the two things that break it
Every rung runs on the same clock: a full week from the last big change, read on the same weekday. Two things reset that clock without telling you. The first is any edit to the ad set: a new ad, a new audience, a budget jump. Meta's own list of edits that restart learning is on the ads tab. The second is you, checking daily and moving money because Tuesday looked bad. A test you touched mid-week is not a test. Write the read date down before you start and do not open the account until it comes.
When nothing moves
Numbers measured as of September 10, 2026. Next read December 8, 2026. The numbers here are dated, and when they change, so does this page.
Can a machine run this?
This tab decides which parts of everything above software can run on its own, and which stay yours. If you read nothing else: let it work inside limits a person set, and never let it move the limits.
Where the platforms already drew the line
Both platforms hand their own software a great deal and hold two things back, and both publish which is which. Google's bidding sets "precise bids for each and every auction," while you pick the goal it bids toward. Meta's broad audience "treats your inputs like interests or age as suggestions rather than hard rules," then names the exceptions: location and minimum age, nowhere else.
Read that twice. It is the whole answer in small. The machine got the auction, you kept the goal. It may overrule your guess about who should see the ad. It may not put the ad in front of a nineteen year old in a state you do not serve.
The split
Every rung on the testing tab is a rule with a number in it, which is what makes this answerable at all. A rule with a number is arithmetic, and arithmetic is what software is good at. Judgment is not, and nor is anything you would defend to a client or a regulator later.
So ask two questions of any decision on this page. Is it counting, or is it a call? And if it goes wrong, can you put it back? Counting that you can undo goes to the machine. Everything else keeps a person on it.
| Decision | Who | Why it lands there |
|---|---|---|
| Stepping a budget up | software | arithmetic against gates written down before the data arrived: the week ran out, the set left learning, cost per lead held, the ceiling not touched. Reversible, and a bad read steps down rather than off |
| Pausing an ad under its floor | software | it has earned a fair read on times shown and money spent, and turning one ad off is cheap to undo |
| Calling a rung's gate | software | did the clock run out, did the set clear the event rate, does the winner sit outside the noise band. Counting, not judgment |
| The offer, and any claim on the page | a person | the trade regulator's line is that there is no exemption for software, and the case it made an example of was punished for shipping a claim nobody had tested |
| Money leaving past a ceiling | a person | the ceiling is the fence, and it has to sit where the thing it holds cannot reach it. A stop button is worth nothing if whatever you are stopping can rewrite the rule that fires it |
| Who gets on the phone, and their details | a person | this is where the rules actually bite. Europe's new law leaves ad buying alone and aims at scoring people, and at designs built to push them |
| Marking who turned up | a person | nothing in the stack knows whether a call happened. The rate that decides every number on the last two tabs starts as somebody typing it in, and software reading a stale pipeline reports a show rate that is really a date |
What the wiring actually looks like
Four pieces, none exotic. The first is the ad platform's own interface for software. Meta publishes one, and everything the table hands to software runs through it: read what an ad set spent and produced, change a budget, pause an ad, launch a new one. Since November 2024 there is also an open standard for handing an interface like that to a model, so the model calls named tools rather than writing raw requests.
The second is a memory it reads before deciding and writes to after. That is the whole difference between a script and something worth calling an agent. Store every decision and what came of it, and the next run starts from what the last one learned.
The third is the runbook itself, written down once as a procedure instead of retyped at it every morning. Every rung on the testing tab is already a rule with a number, a clock and a gate.
The fourth is where the truth lives. The ad platform knows clicks and emails. It does not know who turned up. Booked, held and closed sit in your CRM, and those three rates decide every number on the last two tabs. GoHighLevel now publishes a server on that same standard, with more than 550 read and write operations across 38 areas, every one of them fenced by the scopes you approve. So the agent reads the held call there and writes it back to the ad platform as a server event.
The part it cannot touch, and why that is fine
Now look through that catalog for the automation builder. It is not there. The whole workflow section of their published reference is one read only call listing the workflows you already have. Software can drop a contact into one, and a workflow can call out to your system. Nothing lets a machine build or edit the sequence itself.
That sequence talks to somebody who just handed you their email. It is the last thing you would want rewritten at three in the morning by something chasing a number. Read the limit as a design, not a gap.
How much of this job is machine run today
Less than the noise suggests. Asked in October 2025, marketing leaders put the share of their work run by software at about 16%, and expected 36% two years on. A different survey put it near 24%. Close enough to believe.
So most of this is still done by people, and even the optimistic forecast keeps it that way for years. Anyone selling a hands off version is describing a room nobody stands in yet.
The benchmarks say why. Put agents inside a simulated company, hand them the work of a digital employee, and the best finished 30% of it alone. Another team measured reliability rather than accuracy and found recent capability gains "have only yielded small improvements in reliability." Right most of the time is not right every time, and the gap is money.
Letting it show costs something too. Of buyers shown marketing they could tell was machine written, 32% said it lowered their trust and 7% said it raised it. A person signs what goes out.
The fence
So if you build any of this, build the fence first. A dry run is the default, and writing anything needs a switch a person flips. Missing data means stop, never guess. A bad result steps the money down rather than turning the ad set off, because off throws away learning you paid for. Every decision gets written down, changed or not.
The last rule is the one people skip. The fence sits outside the thing it holds, somewhere that thing cannot reach. Otherwise you have not built a limit, you have built a suggestion.
Numbers measured as of September 10, 2026. Next read December 8, 2026. The numbers here are dated, and when they change, so does this page.