Add up the conversions every platform claims for last month. Then open the order system. If the first number is bigger than the second, and at real spend it almost always is, you already know the problem. Your marketing mix model will tell you a third story. You just haven't priced the gap.
Here's the part that bothers me more. Three teams will spend the next budget meeting arguing about whose ledger is right, the CFO will ask which number is broken, and not one of them will mention the thing that actually moved the number: what the ads said.
Up to three in four buyers say their measurement falls short
A recent industry report [1], built on a survey of more than 400 senior planning and analytics leaders at US brands and agencies and released in February 2026, is blunt about where measurement stands. Up to three quarters of buy-side leaders said attribution, incrementality tests and mix models all fall short on rigor, speed, trust or efficiency. No paid channel came out as fully represented in their mix models. Half said commerce media was underrepresented, and close to half said the same about creator and influencer spend.
The same report found most mix models and attribution studies get refreshed once or twice a year, while buyers want monthly updates and expect to roughly triple the number of incrementality tests they run. Its authors estimate that better, AI-assisted measurement could shift around $26 billion in ad spend toward channels today's models undercount.
That's a real problem. But notice what everyone surveyed wants: faster models, more tests, more data. More instruments. Nobody asked for better ads.
Steering by the wake
I started with words and learned the math second, and that order matters here.
The technician looks at three ledgers that disagree and sees a data problem. Refresh the model monthly. Add a clean room. Run more tests. Every fix is another instrument with its own owner and its own seat at the budget meeting. That's honest work and I want it done.
The strategist looks at the same ledgers and asks a different question: what were people actually responding to? A channel doesn't sell anything. A message sells, and the channel carries it. When modeled contribution jumps, something said something better to somebody. When it drops, the words wore out or started speaking to the wrong buyer.
None of the three ledgers can see that, because none of them were built to. So the technician keeps tuning the instruments while the actual cause of the swing walks out the door with the creative team.
Every ledgermeasures where thesale landed. None ofthem measures whatmade it happen.
The numbers matter. They're just downstream of the words, and if you only manage the downstream, you're steering by the wake.
Nine numbers, one job
I learned what disagreeing numbers do to a team the hard way. At Hyundai Capital America I was Western Region Regional Sales Manager, measured on nine core KPIs at once, and I was ranked #1 nationally across all nine for 15 of 24 months.
Nine numbers, each a partial truth. Picking a favorite would have been useless. The job was to keep them reconciled so nobody could hide in the one that flattered them, then go fix the conversation driving all nine. The numbers told me where to look. What got said across the desk moved them.
Marketing measurement at scale is the same job.
Marketing mix model, platforms, bank: the Three Ledgers
Each ledger answers one question well and every other question badly.
- The bank ledger answers: how much did we actually make? Orders, revenue, contribution. It's the only ledger that can't double-count, and the only one that doesn't know where anything came from.
- The model ledger answers: roughly how much of that total did each channel cause, over months? It's slow, it misses channels it wasn't built to see, and it can't see individual ads at all.
- The platform ledger answers: within this channel, which ads and audiences drove the most response? It's fast and granular, and every platform grades its own homework.
The rule is simple. Each ledger is only allowed to answer its own question.
The bank sets thetotal. The modelsplits the total.The platforms rankinside their slice.
Keeping each ledger in its own lane, month after month, is a habit the iFix Lab drills.
The math: 14,000 claimed sales, 10,000 real ones (illustrative)
These are modeled numbers, built for illustration. They're a model to reason with, and they won't predict your own split. A business does 10,000 orders in a month.
- Platform A claims 6,200 conversions. Platform B claims 4,600. Search claims 2,400. A fourth partner claims 800. Total claimed: 14,000.
- Over-claim ratio: 14,000 claimed divided by 10,000 real, which is 1.4. The platforms collectively claim 40% more sales than happened.
That ratio is your first monthly number. Watch its direction. If it climbs while spend climbs, the platforms are increasingly claiming the same buyers, which usually means you're spending deeper into overlapping audiences.
Now the model. Say it estimates that 6,000 of the 10,000 orders were driven by advertising at all, with the rest coming from repeat buyers, word of mouth and existing brand demand. It credits 22% of those 6,000 incremental orders, about 1,320, to Platform A. Platform A's own share of claims is 6,200 out of 14,000, about 44%. That's a 2 to 1 gap in share, and a far bigger one in raw orders.
Neither number is the truth. So you run a geo holdout on Platform A. Suppose it shows Platform A drove about 1,620 incremental orders, roughly 27% of the 6,000. Now you have a calibration: the platform's claim of 6,200 is about 3.8 times what it actually caused, and the model under-credits it by about 5 points of share. You adjust the model and set a platform correction factor, both date-stamped.
At $10M a month, the gap between 22% and 27% is real money. If Platform A gets $3M a month and the model's undercount pushes you to cut it by a fifth, you've moved $600,000 a month away from a channel that was earning it.
Now the strategist's footnote to that same illustration. Suppose that holdout ran during the six weeks Platform A carried a new concept that spoke straight to a buying moment nobody else in the category was naming. The 27% isn't a property of Platform A. It's a property of Platform A carrying those words. Swap the concept for a tired one and the same channel, same budget, same audience could read back at 18%. The math is right. It's just measuring the message and labeling it a channel.
Over-claim ratio 1.4. The bank sets the total; everything above the line is double counting.
Where creative breaks all three ledgers
Here's the column nobody keeps. None of the three ledgers measures creative as a cause.
The bank doesn't know which ad sold. The model credits the channel for whatever its creative did. The platforms rank ads only against each other, inside their own walls.
So when a channel's modeled contribution jumps after a creative change, the model learns “this channel got better.” Next year's plan funds the channel. The writers who caused the jump don't get the budget, and the next creative cycle inherits a target built on an idea that's already wearing out.
The fix: tag every channel-level read with the concepts that were live during it, in one plain sentence each. What did the ad say, and to whom? Model results become “channel plus message set,” same as holdout results. When the message turns over, the read gets flagged for refresh.
Two warning signs hiding in the ledgers
- The over-claim ratio is a frequency alarm. When it climbs, check frequency and audience overlap first. Rising overlap is often the cause, and overlap is usually what happens when every concept in the account is saying the same thing to the same person.
- Teams start writing for the ledger. When the platform ledger runs the meeting, people write ads that win attributed clicks. The dashboard says you're doing great, and the bank ledger doesn't move.
Why one source of truth turns dangerous between $1M and $10M a month
At $1M a month the usual advice works: pick one source of truth, usually the analytics platform or a simple blended number, and get on with it.
Between $3M and $10M, the single source of truth becomes the most dangerous number in the building. Whoever controls it controls the budget. Teams start optimizing for the measurement and lose sight of the business, and the over-claim ratio drifts up unnoticed because nobody's tracking it.
And “build a mix model and you're done” ages badly. A model built once and refreshed yearly can't see creative cycles, new channels or a change in the auction. It quietly credits last year's best headline to whatever channel happened to carry it.
Past $10M a month, measurement is a governance fight
- Measurement becomes a governance question. Somebody needs authority to declare which ledger answers which question, and to settle disputes between the CMO's team and finance well before the budget meeting.
- The CFO should co-own the calibration calendar. When finance helps design the holdouts, finance trusts the results. That trust is worth more than any model upgrade.
- The person who owns the words needs a seat at the measurement table. At this spend, the biggest swings in your ledgers are message swings. If the head of creative never sees the calibration table, you're paying for precision on numbers nobody's allowed to act on.
The monthly reconciliation page
One page a month, owned by someone without a channel budget, with finance in the room and the agency owner reading every version. It carries four things: the bank total, the trend in how much the platforms over-claim against it, the model's split by channel, and what the ads said while each of those numbers was being made. That last line is the one most teams leave off, and it's the one that explains the other three.
The page itself, with the calibration table and the rule that builds your test calendar, is a template members get inside the iFix Lab.
Questions senior buyers ask
What is a marketing mix model, in plain words?
A statistical model that looks at months or years of spend and sales history, plus outside factors like season and price, and estimates how much of your total sales each channel caused. It's good at the big split over time. It can't see individual ads.
The platforms say we're doing great and finance says we aren't. Which one's broke?
Usually neither, exactly. Each one is answering its own question. The platforms count every sale they touched, finance counts every sale that happened, and both are right about what they count. The over-claim ratio shows you the size of the gap, and the reconciliation page decides which ledger answers which question.
Which ledger should the CMO report to the board?
The bank ledger for results, with the model split as the explanation. Platform numbers belong in operating reviews.
Can AI-assisted mix models fix this?
They can make models faster and cover more channels, which the survey work says buyers badly want. They can't make a model see a message it was never given, and they can't settle a turf fight. Speed helps. Governance decides. The words cause.
How do we stop platforms from double-counting the same buyer?
You can't stop them, because each platform only sees its own touchpoints. You can stop it from reaching the budget. The over-claim ratio caps the total, the model divides it, and any platform claim above its calibrated share gets marked down before anyone uses it to argue for money.
What should the first holdout test be?
The channel where platform share and model share disagree most, weighted by spend. Hold the message steady while it runs, so you're testing the channel and not a creative change.
Who owns the reconciliation sheet?
Someone who doesn't own a channel budget. Analytics, finance, or a measurement lead who reports to both, with the head of creative copied on every version.
The meeting I'd rather sit in
Three ledgers on the table, each with its question written above it. One reconciliation sheet everyone signed off on before anyone walked in. And a fourth column the technicians didn't ask for: what we said, to whom, while each number was being made.
Getting there takes a few uncomfortable conversations with finance, with your partners and with the people who'd rather the dashboards stay in charge. The iFix Lab is where operators learn to walk into those conversations as strategists. The weekly Signal Briefing gives you what moved in search, signals and conversion, each item with a verdict (hype, watch or working) and a move, so you show up with the fourth column already filled in. The next budget meeting will pick a ledger whether you're ready or not. Join the iFix Lab for $97 a month. Cancel any time, and month one carries a 14-day refund. Next note: what happens to targeting when the signal goes dark, and why the words in the ad are now doing the job the pixel used to do.