ONE BUDGET, THREE PLATFORMS
We built Bach AI to find revenue leaks inside Meta ad accounts. It did that well. It was also the wrong product, and the way we found out is worth writing down.
For about a year we ran audits for e-commerce brands and handed back a number: here is what your Meta account is wasting. The number was real and the brands acted on it. Then, repeatedly, something uncomfortable happened. We'd fix what we found, and the business wouldn't move as much as the fix suggested it should.
The reason, once we stopped avoiding it, was obvious. We were handing brands a Meta answer to a question that had never been about Meta. The larger money wasn't inside the account we were looking at. It was sitting between that account and their Google and TikTok spend, in a space our product could not see.
We had built exactly the thing I spend my time criticising: software that looks inside one platform while the money leaves between them.
WHAT A SINGLE-PLATFORM VIEW CANNOT SEE
Four things, specifically — the same four I wrote about earlier this month.
It cannot see that your Meta lookalike, your Google Display audience and your TikTok interest targeting are chasing substantially the same people, so you're paying a premium to compete with yourself across three auctions. It cannot see that your branded search spend is collecting a toll on demand your TikTok creative generated two days earlier. It cannot see budget stranded in a channel that got expensive on Tuesday while another hit its cap on Thursday. And it cannot see that the creative you killed for a mediocre in-platform ROAS was the one producing your branded search lift.
Each of those has a dollar value. In most accounts, together, they're larger than anything a single-platform audit finds.
THE REFRAME
Here's the sentence the whole rebuild came out of.
You don't have a Meta budget, a Google budget and a TikTok budget. You have one budget and three places to put it.
That sounds like semantics until you follow it through, at which point it changes your reporting, your allocation logic, your review cadence and your org chart. Four rules follow from it, and none of them require buying anything.
RULE ONE: ONE SCOREBOARD
Blended MER — total revenue divided by total ad spend across every channel — goes at the top of every report, above every channel number. Platform ROAS gets demoted to a diagnostics section, in smaller type, further down.
This looks cosmetic and it is not. People optimise whatever sits at the top of the page. Put three channel ROAS figures up there and you will get three channels each optimising itself at the expense of the other two, which is precisely the behaviour that produced the problem.
RULE TWO: ALLOCATE ON MARGINAL RETURN
Stop asking which channel performs best. Ask where the next thousand dollars does the most.
These give different answers more often than not. A channel showing 4.8 average ROAS that has already saturated its addressable audience will return very little on the next thousand. A channel at 3.2 with room to grow will return considerably more. Average ROAS is a summary of everything you have ever spent; the decision in front of you concerns money you have not spent yet.
Get Abhishek Pinninti’s stories in your inbox
Join Medium for free to get updates from this writer.
You can estimate the marginal curve crudely and still be far ahead: plot each channel's daily spend against daily contribution over sixty days and look at where the slope flattens. Crude and directionally right beats precise and about the wrong question.
RULE THREE: REBALANCE WEEKLY
The auction reprices daily. A monthly budget split is a bet that the market will hold still for thirty days, and it never has — competitors launch, seasonality shifts, a platform changes its delivery, someone else's Black Friday moves your CPMs.
Weekly rebalancing isn't optimal either. It's roughly four times better than monthly, and it's achievable by a human, which matters more than optimality.
RULE FOUR: ONE CHANNEL ON TRIAL EVERY QUARTER
Pick your least certain channel. Run a fourteen-day matched-geo holdout. Read total revenue, not platform revenue. Four a year.
After a year you have four pieces of actual evidence about what your channels do, rather than four quarters of accumulated assumption. That compounds in a way no dashboard does.
WHY ALMOST NOBODY DOES THIS
Not ignorance. Every competent media buyer I know already agrees with everything above.
It's capacity. Doing this properly every week means six data pulls, reconciling three attribution models that disagree by design, estimating marginal return per channel, making a call, executing it across three interfaces, and writing down why — so that in six weeks somebody can tell whether it worked. That's most of a day, every week, forever, and it is always the first thing to die when a launch or a stockout or a board meeting eats the calendar.
And there's a structural reason underneath the capacity one. Your Meta buyer owns Meta. Your Google person owns Google. Someone freelance owns TikTok. Nobody on the org chart owns the space between them, so in most brands the person holding the whole picture is the founder, at eleven at night, with three tabs open.
WHAT WE REBUILT
Bach AI now works across Meta, Google and TikTok in one view, reconciled against real store revenue rather than platform-reported revenue, and built specifically around the four leaks above — because those are the ones that show up nowhere else.
The second change matters more than the first. We stopped building something that reports and started building agents that act: reconcile, find the overlap, rebalance on marginal return, flag the creative the platform reports are mispricing, and then do it again next week. Finding the leak was never the scarce thing. Closing it every Monday, across three platforms, forever, is the scarce thing.
What the agents don't do is set strategy. They work inside a mandate you give them — your margin floor, your growth appetite, your no-go channels. Anything claiming to choose your positioning or your offer is selling you a model of a business it has never met.
WHERE TO START
Honestly: not with us. Start with the ninety-minute audit I published last week. It's free, it's complete, and it will tell you whether any of this describes your account.
If it does, and if the problem turns out to be that nobody has a Monday long enough to keep doing it, that's the conversation I'd like to have.
—
I'm Abhishek, founder of Wittelsbach. Five years of running paid acquisition for e-commerce brands, and one year of building the wrong product before building the right one.
