Making Trade and Retail-Media Spend Earn Its Place on the CPG P&L

A quarter of your gross revenue leaves through trade promotion, and the retail-media line is gaining on it fast. Ask the room which of those dollars actually moved units that wouldn’t have sold anyway, and the honest answer is a shrug and last year’s calendar copied forward. The display, the feature, the sponsored placement — most of them looked like they worked, because volume rose during the week they ran. Whether that volume was incremental or just pantry-loading a base you were never going to lose is the question your reporting cannot answer — and the next annual plan is about to re-bet on it at full scale.

This paper is for the VP of Revenue Growth Management or the CMO who has to defend that line to a CFO. By the end you will have a way to reallocate trade and retail-media dollars on measured incrementality rather than relationship history, and a single governance rule for what spend gets approved — one a finance partner can sign without taking your word for it.

The leak, in numbers you can put in front of the board

Trade promotion runs 15–25% of gross revenue for most packaged-goods brands — typically the second-largest line after cost of goods. Retail media is the fastest-growing line on the commercial budget, climbing toward the kind of money trade has always consumed. Between them they form the discretionary spend pool, and the discretion is the problem: a large fraction of trade dollars funds events whose “lift” is forward-buying or stocking up a household that was going to buy anyway. Industry incrementality work routinely finds that a third or more of promoted volume is non-incremental — paid uplift the brand would have earned at full margin.

The waste persists because it is invisible where decisions are made. Aggregate ROI looks defensible; the event underneath it never gets inspected. So the calendar renews: this feature ran last year, the buyer expects it, and pulling it is the harder conversation. That is how a brand runs a thousand events, funds the same share of dead ones every cycle, and never sees the line item that is bleeding — because no single one is. The leak is spread across every event that was never measured.

Why event-level attribution fails today

The honest reason brands cannot answer “which dollars worked” is that the question requires a counterfactual, and the data resists it. To know an event’s incremental lift you have to know what the base would have done without the promotion — a quantity that, by definition, never happened. Estimating it well is the entire job, and three things break it:

• Baseline is contaminated. The pre-promo weeks used to anchor the base are themselves shaped by prior promotions, seasonality, distribution swings, and competitor activity. Naively subtracting a flat baseline from promoted weeks counts pantry-loading as uplift and pulls future demand into the event window.
• Syndicated data lags and gaps. Nielsen and Circana are the category’s external scorecard, which is exactly why they are over-trusted. They arrive weeks late, miss coverage, and smooth over the in-quarter signal that retailer POS already holds — so a dead promotion gets discovered a month after it could have been cut.
• The signals disagree. Shipment, retailer POS, and syndicated views of the same week rarely reconcile cleanly, and the gaps between them are where the argument over “did it work” goes to die. Reconciliation forecast-to-POS often sits at 60–88%, which is too loose to settle an event-level fund/kill call.

Retail media compounds the failure. The networks bill against their own attribution while the trade team credits the same week against its model — so a single incremental unit gets claimed twice, paid for twice, and booked into a margin number that was never real.

The bet: causal accountability before the plan locks

The position is narrow and testable. The win is not a better dashboard and not more volume — it is making each promotion and each placement carry a measured incremental result the moment its spend is committed, so the next annual plan is built from causal evidence instead of inherited habit. Keep the systems you already run — TPM, demand planning, ERP, the syndicated subscriptions, the retail-media networks. They hold the authoritative numbers and should keep holding them. Add one governed layer above them where policy and AI estimate incrementality per event, flag the spend that earned nothing, and route only the genuine judgment calls to people. The authoritative plan stays where it is; what changes is that nothing gets funded on faith.

What you accumulate this way is not copyable. A competitor can buy the same Circana feed and rent the same retail-media network tomorrow. What they cannot rent is your library of which events actually drove incremental lift — in your categories, at your retailers, against your contaminated baselines — encoded as versioned rules instead of living in a senior trade marketer’s instinct. That library is the compounding asset, and it is what lets you defend a reallocation to a board.

What AI actually does here: separate base from uplift, then flag the dead spend

The AI job in CPG is not detection and not content quality. It is causal incrementality at the event level — and it earns its keep on exactly the two tasks humans cannot do at this volume.

First, attribute the lift. Across messy syndicated, POS, and shipment signal, the model builds a defensible counterfactual baseline for each promoted SKU-retailer-week — controlling for seasonality, prior-event pull-forward, distribution, and competitor activity — and reports the incremental units the event created versus the base that would have sold regardless. That is the number the P&L has been missing.

Second, flag the non-incremental spend for reallocation. Once every event carries a measured incremental ROI, the ones below the line stop being invisible. The model surfaces them as a reallocation list — this much trade money, on these events, bought no uplift — and reconciles retail-media buys against trade so the doubly-claimed unit gets paid once. That turns trade and retail media from a relationship cost defended by anecdote into a measured, governed investment defended by evidence.

These tasks scale only with agents driving them — no human team reconciles millions of SKU-retailer-week cells inside a quarter. But authority is granted by stages, gated on how well the model’s estimates hold against realized POS, not switched on across the board.

1. 01Ingestsyndicated, POS, shipments
2. 02Baselinemodel the counterfactual base
3. 03Attributeincremental lift per event
4. 04Flagnon-incremental spend — RGM gate
5. 05Reallocatemedia netted against trade
6. 06Retireauto-cut proven-dead events

The gate sits where it does on purpose. Ingesting feeds and estimating baselines can run unattended early — they only inform. Cutting seven figures of trade capital, or reallocating a retailer’s media, stays under human sign-off because the cost of being wrong is lopsided: hold a genuinely incremental event and you lose a cycle of upside; fund a dead one and you burn the cash and teach the buyer to expect it again. Each gate is tuned to that lopsidedness, not to a model accuracy score.

Governance: approve spend by measured incrementality

The accountability is only real if it is a rule, not a memo. The governing principle is one sentence: no trade or retail-media dollar is approved unless its event clears a measured incremental-ROI threshold, or a finance partner explicitly overrides on the record. The authoritative plan in TPM stays exactly where it is; this layer decides what is allowed to enter it. As versioned, executable policy:

RULE TradeAndMediaApproval
WHEN event.type in ["promo","retail_media"]
THEN require incremental_roi >= 1.2
  AND baseline_confidence >= "medium"
  ELSE action = "FlagForReallocation"

WHEN media.spend AND trade.event share incremental_unit_id
THEN action = "DeduplicateAttribution"        # pay the unit once

WHEN baseline.source == "syndicated_only" AND qtd_pos_available
THEN require reconciled_baseline = true        # no lagged-only calls

WHEN promo.margin_floor_pct < 15
THEN require finance_override = true

Every approval decision emits one auditable, reproducible record — the artifact a CFO trusts over a number negotiated in a deck:

{
  "eventType": "TradeMediaSpendDecided",
  "eventId": "RGM-2026Q1-00417",
  "brand": "NorthHarbor",
  "retailer": "Target",
  "decision": "FlagForReallocation",
  "incrementalRoi": 0.91,
  "baselineConfidence": "high",
  "baselineSources": ["circana", "retailer_pos", "shipments"],
  "doubleCountFlag": false,
  "policyVersion": "cpg.spend.v13",
  "modelVersion": "incrementality-2.4.1",
  "decidedAt": "2026-02-24T12:31:22Z",
  "correlationId": "cpg-28f7c1"
}

Accountability needs a named owner per call, so the approval runs on a fixed responsibility map rather than on whoever negotiated hardest:

Role	RACI	Owns
Commercial Data Director	A	The reconciled baseline every decision is measured against
VP Revenue Growth Management	R	Event-level approve / reallocate against the ROI threshold
Retail Media Lead	R	Media netted to trade, no double-counted unit
Finance / Trade Spend	C	Margin floors and the override of record
CMO / COO	I	Reallocation, incremental-ROI, and mix trends

R = Responsible, A = Accountable, C = Consulted, I = Informed. Privacy and retailer-confidentiality limits are a design rule, not an appendix: minimization, purpose limitation, and retention apply to the evidence store and the syndicated/POS joins, not only the master record. Keeping cost terms, margin logic, and JBP detail inside your own boundary is also why the high-frequency scoring runs local — you do not pay a per-call tax to estimate millions of SKU-retailer-week cells.

The economics: reallocation upside, and the cost of being wrong on baseline

Put the planning ranges in front of the board as the size of the prize, calibrated to your own categories and baselines.

• Non-incremental trade spend20–40%→<15%Reallocation upside
• Measured promo ROI0.8x–1.6x→1.4x–2.5xMargin from killing dead events
• Retail-media vs. trade attributiondouble-counted→netted, paid onceRecovered margin
• Forecast-to-POS reconciliation60–88%→90–99%Baseline trust

The upside is concrete: move non-incremental spend from a third of trade toward the mid-teens, and on a 15–25%-of-revenue budget that reallocated money is margin you already owned and were giving away. But the figure that bounds the risk is baseline confidence. The whole reallocation rests on the counterfactual being right: estimate the base too low and you kill events that were genuinely working; too high and you keep funding dead ones with a confident-looking number attached. That is why the rule carries a baseline_confidence gate and a finance override, and why the ladder holds the reallocation under human sign-off until estimates have tracked realized POS over enough cycles. The cost of being wrong about what would have sold anyway is the one number a CFO will press on — so it is measured, gated, and on the record, not assumed.

Where capital goesNowPilotWatch Event-level incrementality modelfund Spend-approval rule (policy-as-code)fund Trade / retail-media de-duplicationfund Clean rooms with retailer datapilot Another commercial dashboard

The board’s reallocation mandate

The board’s job is to allocate capital across trade and retail media, and the same four questions apply at every funding gate: How much spend did we reallocate off measured non-incrementality, and what margin did it return? Which dollars moved versus merely shifted between budgets? Which controls — the ROI threshold, the de-dup check, the baseline-confidence gate — are now automated and testable? What reusable assets — the incrementality model, the approval policy, the evidence record — did we build?

That last answer is the one that compounds. The first category carries the full cost of the model, the policy gate, and the evidence store; the second reuses all of it and pays only for its own events, and the model is sharper for having seen more of the portfolio’s real lift.

1. First 90 days — measure one categoryStand up the incrementality model across syndicated, POS, and shipments for one category. Baseline today's non-incremental %. Run the approval rule in flag-only mode so every event gets a measured ROI and a logged decision, with no spend cut yet.
2. Months 4–12 — reallocate and netTurn the flags into reallocations under RGM sign-off. Net retail media against trade so the doubled unit is paid once. Extend to adjacent categories and retailers. Let auto-retire graduate only where estimates have held against realized POS.
3. Ongoing — make it the planning inputEach cycle: reallocate events below threshold before the calendar repeats them. Monthly: the reallocation / incremental-ROI / baseline-confidence review. Quarterly: refresh the model and re-rank where capital goes.

Get this in before the next annual plan locks, and the calendar stops copying forward — every promotion and every placement enters the plan already carrying proof that it earned the dollar, or proof that it didn’t.

---

References: Circana / Nielsen syndicated measurement · retailer POS and shipment reconciliation · retail-media network attribution · trade-promotion incrementality · NIST AI RMF · NIST Privacy Framework · NIST CSF 2.0 · GDPR · CISA Secure by Design.