See which partners are earning their money while the campaign is still running. Not after.
You can drill into any report as deep as you want: partner, geo, creative, hour. Every platform slice is still counts that platform credits to itself. And the tools that do join across partners, your MMM, your clean rooms, your lift tests, report in weeks and at channel grain. The decisions in front of you are about Thursday, at the grain you buy at. That gap is the product.
The rows you already generate can answer the Thursday questions.
Where to shift spend next. Which segment is saturating. Which path stopped earning its price. Precise reads the log-level rows your programmatic, CTV, and ad-server buys already produce and answers those questions while the money is still moving. Walled-garden spend exports less; where a platform will not hand over rows, the read says so and prices what it can see instead of guessing at the rest.
Same tags, same pixels, same dashboards; the read comes from rows you already have. Your team keeps making the moves; the moves start carrying their evidence.
Every morning, one read: where to shift spend next.
The read covers every unit you can act on: partner, segment, creative lane, supply path. For each one it reports what the unit contributed over the trailing window against its cost, what shifting more spend there would return, with a range around it, and one recommended move with its predicted result written down before anyone acts.
A trailing number tells you how the spend behaved, not which move to make next. The read prices the move itself, before you make it, and that matters most exactly where a unit is saturating, which is where the money is.
When what your company knows joins the read, it stops being generic. It knows which client hates surprises before it recommends a bold move, and it weighs a win the way your team defines one. That knowledge is captured from the work you already do, each piece showing up only in the read that needs it, and it never leaves your control.
Here is what the first read finds.
Some layers of a campaign stay invisible until every partner's logs sit joined in one place; no single platform can price them alone. Below are four such layers from one campaign's first read, one finding each; open a finding for the numbers. The layers are the ones we check first on every campaign. The numbers are invented to show the shape.
Illustrative · numbers invented to show the shape
1Segments nobody chose were billing on every impression.open
The plan named six audience segments. The delivery logs showed fourteen. One platform had auto-appended lookalike and expansion segments, each carrying its own data fee on every impression. Priced by contribution, the eight appended segments were taking 11% of campaign cost and producing about 2% of the outcomes.
The team turned them off, with the predicted effect written down first: outcome change inside the noise band. The grade came back in six days. Seven segments were dead weight, as predicted. The eighth had been quietly feeding conversions, so the read flagged its own miss, that segment went back on, and every read since prices appended segments instead of assuming about them in either direction.
2Vendor fees were stacking per impression until the working share shrank.open
Verification, brand safety, pre-bid filtering, and a data pass were each billing per impression on the same supply paths. On the worst path, five vendors took $1.90 of an $11 CPM before any media ran. Wherever the fee stack passed 15% of the CPM, additional spend was buying outcomes at roughly 30% above the account average.
The team consolidated to two vendors on those paths, predicting no measurable outcome loss. Ten days later the grade confirmed it: outcomes flat, inside the band, and the fee savings moved to the cheapest outcomes on the board. Right this time. The point is the team would have known inside the flight either way, not at the post-mortem.
3The same inventory was arriving twice under different names.open
About 9% of impressions were the same publisher avails reaching the campaign through two exchanges under different path names, at two different prices, with the campaign sometimes bidding against itself for them. Platform reporting showed two acceptable paths. The joined logs showed one path plus a markup.
The team cut the pricier path, predicting reach would hold. It held, and cost per outcome on that inventory came down about 12%. Booked into memory: the duplicate-path check now runs on every read, because paths get renamed more often than they get cheaper.
4Two overnight corners were quietly absorbing spend.open
Overnight hours in two geos were absorbing 7% of spend at a marginal cost near three times the account average: the corner of the plan nobody reviews daily. The read priced it on day one.
The team capped overnight delivery, predicting the savings would redeploy cleanly. The grade said otherwise: part of that overnight audience was feeding a retargeting pool that converted the next afternoon, so the cheap-looking cap carried a real downstream cost. Half the cap came back off within the week, and the read now prices that handoff. Being right the first time was never the promise. Knowing by Thursday was.
A fifth layer returned no finding at all: the newest partner on the plan was below the support floor, so the read printed "not enough data yet" instead of a number. A guess there would have been treated as a defect.
Four findings, four moves, each predicted before acting and graded after. Two grades confirmed the call, one sharpened it, one partly reversed it. That mix is the loop working.
Some calls are wrong. The loop catches them in days.
Every recommendation is written down before you act: what was known at the time, what was chosen, what was predicted at what confidence. After the outcome matures, the prediction is graded against what actually happened, and the lesson is booked into the next read. The predictions exist so the loop can learn, not so anyone can frame the wins.
- Append-only. The log is never rewritten. Wrong predictions stay with their grades, because the misses are what teach the next read, and the record is how you know when to lean on the system and when not to.
- Reproducible. Any past read re-runs against the same inputs and settings and produces the same answer. Your analysts can audit any number back to its run.
- Yours. The measured contribution dataset is handed back to you. Your team can fit their own models against it and disagree with ours; the measurement layer is the shared ground.
How long does a wrong call survive in your current plan: days, or until the post-mortem? Here it survives until the next grading window. The budget moves, and the system remembers why.
The system proposes. People move the money. Authority to act directly is something the track record earns at the pace you choose, never something the software assumes.
It refuses more than it claims.
This is an observational read of live spend, not a randomized experiment, and it says so. Your lift tests and holdouts still answer whether a channel works at all; this answers which parts of the spend you are already running are earning their price right now. The two compose. Where they disagree, the disagreement is surfaced to you, argued in writing, and settled by the next measurable window.
A pilot starts with one campaign that already ended.
We run the read on a closed campaign first, and you check it against what you already believe before anything is acted on. If the read cannot earn your analysts' respect on a campaign you know cold, it has no business touching a live one.