Meta and Facebook ads agency for B2B SaaS, startups and ecommerce
Cell Athe hook that’s live · you’re reading it
Meta automated the targeting. Results now come from the two levers left: the signal you feed the system, and how many genuinely different concepts you give it to test. We run both, end to end.
It starts free: we read your account and write you three new ad concepts in five business days.
Cell Bthe board it came from · next hooks queued
“The price is the reason it works.”
Founder to camera · beat control on spend-weighted CAC, budget shifted in
“Your spreadsheet is doing this job badly.”
Static · 3 variants · in learning, judged at 7 days, not on day 2
“Watch the same task, twice.”
Screen demo · 45s · hook rate fine, holds died at 3s. Logged, next angle queued
A senior replies within one business day. Mutual NDA before access, and you keep every file.
Winners scale. Losers die at day 7. Your account runs this test every cycle, 8 to 12 concepts at a time.
Free, in 5 business days · Read-only access · Mutual NDA · You own the Business Manager and every file
What’s in the retainer
Meta ads run as a learning loop, and these are its four stations, itemised. A senior owns and signs the account, a specialist team produces the creative, and you keep the Business Manager, the dataset, and every file.
What you tell Meta happened is what it optimises toward. We make that signal true.
A consolidated structure that gets out of the algorithm's way instead of fighting it.
The volume of genuinely different ideas the system needs, produced on a schedule.
Reporting against your books, so in-platform ROAS never gets the last word.
Month to month, scoped from what the Learning Teardown finds as a custom quote. No long contract, and nothing we build is locked to us.
What actually happened
The winning ad from March is dead. CPMs drifted up. Blended CAC crept past the line in some month nobody can point to. The usual explanations are iOS, the algorithm, or the market. None of them are checkable, which is the problem.
Here is what is usually true instead: the account is spending fine. It is learning badly.
How Meta actually works now
Targeting is no longer a lever. Detailed targeting was consolidated in 2025 and exclusions were removed, so the only hard constraints you still control are location and minimum age. Everything else you enter is a suggestion. Meta now matches individual creatives to individual people, which means two things decide your results: the quality of the outcome data you send back, and how many genuinely different ideas you give it to test.
The loop, station by station
What you tell Meta actually happened.
What Meta concludes about who is worth showing you to.
What Meta has available to show them.
What comes back, and whether you can trust the number.
When the loop is clean it compounds. When any station is broken it decays quietly, and the dashboard keeps reporting green.
Decaying
return arc broken
Event match quality
EMQ 4.2
Optimisation event
optimising to the wrong event
Distinct concepts live
3 concepts live
Reported against your books
reported 2.1x / actual unknown
A clean loop compounds. A broken one decays quietly, because the reporting stays green while the return arc is open.
Illustrative composite of the pattern we open most often. Not a real client, and no client numbers appear anywhere on this page. Each readout is a check the Learning Teardown actually runs.
Two platforms, two different leaks
Google wastes your money on the wrong searches. Meta wastes it learning from bad information.
Station one
Meta optimises against the outcomes you report. If half of them never arrive, or arrive twice, the system optimises hard toward a fiction. This is the most common and most expensive problem we find, and it is invisible in Ads Manager.
Purchase event
About four buyers in ten are matchable at this score.
Pixel against CAPI
Under 70 percent the same event is double counted or lost.
What the account bids toward
Optimising two steps before the money changes hands.
Reported against real
The window closes 55 days before the median sale.
184 matched, 128 with nothing behind them
Drawn on the same scale
The gap is 128 conversions, 41 percent of everything Meta reported. Until those two counts agree, the ROAS in Ads Manager is not a number you can take to your accountant, and the model has been learning from the difference.
The four checks, and the threshold each one is scored against
Published in advance
Target 7.0+. At EMQ 4 the model is effectively seeing about 40 percent of your buyers. Usually fixed by sending phone, external ID and the click identifiers, not just email.
Target 70 to 100 percent. When event_id and event_name disagree across pixel and CAPI, dedup fails silently and you either double count or lose the event.
The model does exactly what you ask. Optimise for Add to Cart or raw Lead and it will find people who do that cheap thing and never buy.
We reconcile Meta against Shopify, Stripe or your CRM. Platform-reported ROAS commonly runs well above de-duplicated reality.
These are published in advance so the teardown is checkable rather than an opinion. You can hold the reading on your own account against every one of them.
If your reported numbers fell off a cliff this year and nothing else changed, that was probably not your performance. On 12 January 2026 Meta removed the 7-day and 28-day view attribution windows. On 3 March 2026 it redefined click-through to require an actual link click, moving likes, saves and video views into a separate 1-day engage-through bucket. When defaults tighten, reported conversions drop even where sales did not, worst in long-consideration categories. This part rarely gets explained. It is checkable in your account in about ten minutes.
Station two
Meta retrieves creatives, not audiences. Your creative library is the closest thing to targeting you still control. Which turns the whole thing into an arithmetic problem, not a talent problem.
C14 has the highest hook rate on this board and it is dead. It stopped the scroll and lost everyone at the click. C07 is the only concept in the cycle earning more budget.
Same concept, new hook and new format. Iterating off something already proven is not the same bet as a cold idea, which is why the winner is the asset and the ad is only the current expression of it.
| Frequency | Outbound CTR | CPM |
|---|---|---|
| 1.5 | 1.74% | $18.90 |
| 2.5 | 1.21% | $22.60 |
| 4.0 | 0.68% | $31.90 |
Our standard on prospecting: frequency at or below 3 per 7 days.
Plan on roughly one in twenty new concepts becoming a winner. The top one to two percent of ads take about half the spend. So five new ads a month produces approximately nothing, and forty produces two winners. The benchmark is one new ad per three thousand dollars of monthly spend, refreshed every seven to fourteen days, because concepts that used to last six weeks now burn out in two or three.
A composite board, not a real client. Every figure on it illustrates the shape of a testing cycle, not a promised result.
Most accounts we audit are shipping four or five new ads a month. We will tell you exactly how many yours shipped, which of them are past their efficient life, and write you the next three.
Eight to twelve conceptually distinct concepts per cycle, two to three variants each. Distinct means a different angle, format or hook. Three colourways of the same ad is one concept, not three.
Ten distinct concepts, three variants each. Thirty assets shipped in a cycle, against the four or five a month most accounts manage.
Angle first, because it has by far the widest variance. Then format. Then hook. Production polish last, and it is routinely inverse-correlated with performance.
Hook rate and hold rate are diagnostics, not goals. You can buy a great hook rate with a fake scroll-stopper and destroy everything downstream. We read the shape of the funnel, not one number.
How we report
In-platform ROAS is a delivery-optimisation readout. It is useful for comparing two ads to each other in the same account on the same day. It is not comparable across months where Meta changed a window, and it is not comparable to your accountant. We optimise against it because the system trains on it, and we report to you on blended MER, new-customer CAC and contribution margin. If you want to know whether Meta is actually incremental, that is a geo holdout, and it takes three weeks.
Optimised against
Reported to you
The refusals
We will not guarantee a ROAS or a lead count.
A number promised before anyone reads your account is a guess.
We do not charge a percentage of your ad spend.
We are never paid more for telling you to spend more.
We will never ask you to run ads from our ad account.
You own the Business Manager, the dataset and every audience.
We will tell you when Meta is the wrong channel for you, before you pay us.
Who this is not for
You are spending more than roughly $300,000 a month.
At that level you can afford a buyer, a creative strategist and an editor in house, and in-house will beat us on attention. Hire them.
Your budget is small and your target cost per acquisition is high.
An ad set needs roughly fifty optimisation events a week to leave the learning phase. If your target CPA is $250 and your budget is $3,000 a month, Meta cannot stabilise on your real conversion event, and no agency can fix that with cleverness. Fix the offer or the funnel first, or start on a channel that harvests existing demand.
We would rather say this now than take three months of your money to arrive at it.
Get your teardown
Five business days. A senior goes through your signal layer and your creative library, reconciles what Meta reports against what your business recorded, and sends back a ranked fix list plus three new ad concepts written for your brand. Free, and yours to use whether you hire us or not.
view-only partner access to your Business Portfolio, on the ad account, the dataset in Events Manager, your catalog if you sell products, and your Page and Instagram. Plus your own numbers, because no cost per acquisition means anything without your margin. You grant it in a few clicks and revoke it in one.
review your live ads from the public Meta Ad Library against your three closest competitors, and check your landing pages and whether your pixel fires correctly. That is genuinely useful and it is not an audit. We will not call it one.
Capacity is limited on purpose: seniors do the work, so we take a few new engagements at a time. The audit is how we choose.
Five business days. Mutual NDA before access. You own the Business Manager and keep every file.
Your current winners are already decaying. That is not a sales tactic, it is how the auction works now: concepts that used to run for six weeks burn out in two or three, and the ones still running are costing more each week to serve. The teardown tells you which of yours are past their efficient life, and gives you three replacements to put in.
Rather talk first? Book a 30-minute call with the senior.
3 free concepts · 5 days
Read from your account, written for it