The Budget Reallocation Audit: When MMM and MTA Disagree
Objective: Given a channel-level spreadsheet showing MTA-attributed revenue, a simplified MMM-style contribution estimate, and geo-holdout incrementality test results for the same 6 channels, apply the lesson's triangulation method to decide which model to trust per channel and draft a reallocated budget.
You're the paid media lead at Instacart. MTA says TV and podcast sponsorships contribute almost nothing; a first-pass MMM says they're carrying real weight. Finance wants one answer before next quarter's budget is locked.
Use the lesson's geo-holdout triangulation method to see which model each channel's real-world test result agrees with, then draft the reallocation.
Before you start
What you'll need
Free path (everything below is enough to finish)
Free, sufficient for a 6-channel comparison table with no modeling software required
Free tier provides the touchpoint-level data MTA numbers are built from
Free, connects directly to Sheets for a presentable chart without a paid BI tool
The process
3 steps
Step 01 of 03
The lesson states MTA only sees what it can track, so it systematically undercounts TV, podcasts, organic social, and any touch where the user can't be identified.
MTA attributes 2% of revenue to TV and 1% to podcast sponsorships, both channels Instacart has spent steadily on for two years. Is 3% combined a reliable read on their true contribution?
Procedure
- Sort the export by MTA-attributed revenue share, ascending
- Flag any channel below 3% that has meaningful, sustained spend behind it
- Note that TV and podcast are both untrackable at the user level, this is a structural gap, not a performance signal
MTA REVENUE SHARE (6 channels) Paid Search 52% Paid Social 28% Display 14% Retargeting 3% TV 2% Podcast 1%
Healthy
A near-zero MTA share for a channel with little to no spend behind it, the low number matches the low investment.
Unhealthy
A near-zero MTA share for a channel with two years of steady, meaningful spend, that's a measurement gap, not a performance verdict.
What this means
TV and podcast can't be near-zero in reality given the spend levels, MTA's near-zero number reflects what it can see, not what's actually happening.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Finance is proposing to cut TV and podcast based on the 2%/1% MTA numbers | Flag both as structurally undercounted before any cut decision, and pull the MMM estimate for a second opinion | 5 min |
Step 02 of 03
The lesson notes MMM can measure TV, radio, out-of-home, organic word of mouth, and long-term brand-building effects, precisely the things MTA can't see.
The same export includes an MMM-style contribution column: TV at 11%, podcast at 6%. That's a 9-point and 5-point swing from the MTA numbers. Which set do you trust so far?
Procedure
- Add MMM contribution % next to MTA share % for each channel
- Compute the absolute point-difference per channel
- Rank channels by the size of the disagreement between the two models
MODEL COMPARISON Channel MTA MMM Difference Paid Search 52% 44% 8 pts Paid Social 28% 24% 4 pts Display 14% 9% 5 pts Retargeting 3% 6% 3 pts TV 2% 11% 9 pts Podcast 1% 6% 5 pts
Healthy
A small point-difference between MTA and MMM for a fully-trackable digital channel, the two methods roughly agree.
Unhealthy
A large point-difference concentrated exactly on the untrackable channels (TV, podcast), matching the structural gap identified in Step 1.
What this means
The disagreement isn't random, it's concentrated on the two channels MTA structurally can't see, which is exactly what the lesson predicts, but a model disagreement alone still isn't proof, it needs a real-world tie-breaker.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| MTA and MMM disagree by 9 points on TV's contribution | Don't average the two numbers or pick one by preference, pull the geo-holdout test result for TV specifically | 5 min |
Step 03 of 03
The lesson's pro tip: run a spend-up or spend-down test in a subset of regions, measure the actual sales lift, then compare it against what MMM and MTA each predicted. This triangulation reveals which model is closer to ground truth.
Instacart paused TV entirely in 4 test regions for 6 weeks. Sales in those regions dropped 9.5% versus matched control regions. MMM predicted an 11% contribution; MTA predicted 2%. Which model does the real-world test support, and what's the reallocation call?
Procedure
- Record the geo-holdout sales-lift result (9.5%) next to the MTA (2%) and MMM (11%) numbers for TV
- Note which model's estimate the real-world test result sits closer to
- Draft a one-line reallocation recommendation for TV based on the test, not either model alone
TV, MODEL VS. REALITY MTA estimate: 2% MMM estimate: 11% Geo-holdout result: 9.5% sales lift when TV is on Verdict: MMM was close, MTA was structurally wrong, keep TV funded
Healthy
A geo-holdout result close to MMM's estimate, confirming the aggregate model over the user-level model for this specific channel.
Unhealthy
Splitting the difference between MTA and MMM without running the test, or trusting whichever number supports the budget you already wanted.
What this means
The geo-holdout, not either model, is the actual ground truth here, and it confirms TV should stay funded near MMM's estimate, not MTA's.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Finance wants to cut TV to zero based on the 2% MTA number | Present the geo-holdout result as the deciding evidence and recommend keeping TV near its current budget | 30 min |
Final deliverable
A 6-channel comparison table (MTA share, MMM contribution, geo-holdout result where available) with a written reallocation recommendation per channel, explicitly citing the geo-holdout as the tie-breaker for TV and podcast.
See a reference example
DoorDash, Q3 channel reallocation memo (excerpt) Channel MTA MMM Geo-holdout Verdict Paid Search 48% 41% n/a Trust MTA, fully trackable Display 6% 13% +10% lift Trust MMM, geo-holdout confirms Radio 1% 8% +6.5% lift Trust MMM, keep funded Recommendation: shift 8% of Paid Search budget into Display, which the geo-holdout confirms is underfunded relative to its real contribution.
Success criteria
You're done when you can:
- Correctly identifies TV and podcast as MTA's structural blind spots, not genuine underperformers
- Computes the point-difference between MTA and MMM per channel without conflating disagreement with proof
- Uses the geo-holdout result, not model preference, as the deciding evidence in the final reallocation recommendation