Model media spend, not impressions. Spend is the input marketers and finance teams actually control, it reconciles to the books, and it translates directly into iROAS and budget decisions. Impressions describe what a platform says it delivered.

Your CFO asks one question: if we move $1 million, what happens to incremental revenue? A model built on spend answers in dollars. A model built on platform-reported impressions has to convert back to dollars first, and every conversion adds potential for error.

That does not make impressions worthless, and cost per thousand impressions (CPM) does move. But the answer to price changes is a well-specified spend model, not a less auditable input. Here is the case for spend, including the strongest argument against it.

Spend And Impressions Measure Different Things

Spend is the money invested in a channel. Impressions are ad deliveries counted under each platform’s own reporting rules. An impression is not a unique person reached, a viewable ad, or a person who paid attention.

The two are linked through CPM:

\text{Impressions} = \frac{\text{Media Spend}}{\text{CPM}} \times 1{,}000

When CPM holds steady, impressions rise in proportion to spend and the two inputs carry nearly the same information about advertising volume.

Week Media spend CPM Impressions
1 $10,000 $10 1,000,000
2 $15,000 $10 1,500,000
3 $20,000 $10 2,000,000
4 $25,000 $10 2,500,000

The interesting question is what happens when CPM moves. That is the strongest argument for impressions, and we take it on below.

Five Reasons Spend is the Stronger Input

1. Spend is the Lever You Actually Control

Marketers do not buy impressions directly. They set budgets, and the platform decides what those dollars deliver. Spend is the one media input you can set, change, and be held accountable for. Impressions depend on what the platform delivers, what the market charges, and how the platform chooses to count.

A response curve built on spend reads straight into the decision: what does the next $100,000 return in incremental revenue?

2. Spend Reconciles. Impressions Do Not.

Spend lives in invoices, purchase orders, agency reports, and the general ledger. Finance can audit it. Impressions are counted differently by each platform, and television exposure often arrives as gross rating points (GRPs) estimated from audience panels.

Combining those definitions in one model introduces measurement error, and inaccurate inputs distort channel effects and the budget recommendations built on them. Spend data can have errors too, especially in bundled buys, but it can be checked against a record the business already trusts.

3. Spend Produces iROAS Directly

Incremental ROAS (iROAS) is the number marketing and finance both need:

\text{iROAS} = \frac{\text{Incremental Revenue}}{\text{Media Spend}}

If a channel drives $300,000 in incremental revenue from $100,000 of spend, its iROAS is 3.0. Impression counts cannot tell you that. Any impression-based model still needs spend as the ROI denominator, so the dollars end up in the model either way.

4. Spend Compares Across Channels

A search impression, a connected TV impression, a retail media impression, and a display impression are not the same unit of anything. Comparing raw volumes across them misleads.

Dollars give planners one unit to move between channels. That does not mean a dollar does equal work in every channel. The model still estimates channel-specific effectiveness, saturation, and delayed effects. It just does so in a currency the CFO already uses.

5. Spend Makes Optimization Simple to Act On

Budget optimization asks where the next dollar earns the most. A spend-based model answers in those terms: marginal return by channel at the current investment level, and where each channel starts to saturate.

Exposure-based models can optimize too, but they add a step: translating exposure back into spend. Every conversion between the two adds assumptions, and those assumptions land in the budget recommendation. Modeling spend removes that translation step.

The Strongest Objection: What Happens When CPM Moves?

The best argument for impressions is that prices move. When CPM rises, the same dollars buy less delivery, and a naive spend model can read higher prices as more advertising.

CPM is also not yours to set. It comes from an auction or a negotiated rate, and it moves with how many other advertisers want the same audience and inventory: holiday demand, a competitor launch, a rival’s budget push. You control your budget. You do not control what the market charges for delivery. An impressions input blends your decision with everyone else’s. A spend input isolates the part you can act on, and the market’s effect can be modeled separately.

Here is the case. A brand spends $10,000 at a $10 CPM and gets 1 million impressions. The next week it spends $15,000 at a $15 CPM and still gets 1 million impressions. Spend rose 50%. Delivery did not move. A model that reads that as 50% more advertising pressure will get the channel’s effect wrong.

That risk is real. It is also a specification problem, and a spend-based model has four answers to it.

  1. Treat the price shift as information. If a dollar buys half the delivery in November, the planning answer is that November dollars return less. A spend model with a seasonal or price-pressure control reports exactly that. An impressions model can capture it too, but then needs a CPM forecast to turn its answer back into a budget.
  2. Model response curves, not straight lines. Saturation and carryover are estimated per channel, so a spend jump with flat delivery shows up as diminishing return instead of a large lift.
  3. Control for known cost shocks. Peak-season auction pressure, major promotions, and platform pricing changes can enter the model as control variables.
  4. Calibrate with incrementality tests. Experiments measure the actual incremental effect at the spend levels run. If cost inflation made dollars less effective, the test shows it, and the model is corrected toward it.

The conclusion is not that CPM never matters. It is that CPM volatility is a reason to build the spend model carefully, not a reason to hand the budget conversation to a less auditable input.

When A Platform Changes How It Counts Impressions

Impression definitions belong to the platform, and platforms change them. Amazon has told advertisers that from January 1, 2027, display ads on Amazon Store placements will count as impressions when they appear on the shopper’s screen, not when they begin downloading. Amazon estimates about 63% fewer reported impressions on average, according to Digiday and PPC Land. The notice frames it as a reporting change, and Google is making a similar move for some display inventory in February 2027.

Consider what that does to an impression-based model. If delivery and spend are unchanged across the changeover, the input still falls by roughly two thirds. The model reads a collapse in advertising activity, history on either side of the date stops being comparable, and the channel’s estimated effect breaks at the seam. Digiday reports that analysts are already advising teams to recalibrate media mix models that use impressions as an input.

A spend series does not have that problem. A dollar on the invoice means the same thing the day before and the day after a platform rewrites its counting rule. If prices move as a result, which buyers quoted by Digiday expect, that shows up as a price shift the model can control for, not as a change in what the input means.

Attribution has the same exposure. In January 2026, Amazon moved view-based attribution for DSP ads on Amazon Store inventory to a model using shopping signals and a shorter window, which changed how reported ROAS reads. Incrementality-adjusted attribution anchors platform-reported credit to experiment results, so a rule change in one platform’s report does not rewrite your view of what worked.

Where Impressions Fall Short As a Model Input

Impressions look closer to the mechanism, because people respond to ads, not to budgets. In practice they carry problems a model cannot fix on its own.

1. Counts do not equal meaningful exposure

Two campaigns can each deliver 1 million impressions and differ widely in placement quality, audience relevance, and attention. A raw count treats them as identical.

2. Definitions vary by platform

Digital platforms, video publishers, and television measurement each count delivery differently. Putting them in one model without validation builds inconsistency into the inputs.

3. Some exposure metrics are estimates

Television GRPs come from audience panels and carry sampling and weighting error. Digital impressions carry their own limits around invalid traffic and viewability.

4. Input error flows downstream

Noisy inputs distort estimated channel effects, and optimization then amplifies the distortion into budget moves.

A clean, consistent input often produces a more reliable model than a theoretically closer measure that arrives with substantial reporting error.

Where Exposure Data Still Earns a Look

Spend is our default, not a universal rule. There are cases where we would test an exposure metric against it:

  • Reach and frequency matter. Upper-funnel campaigns can behave differently depending on how many unique people see an ad and how often. Total spend or total impressions cannot show that.
  • CPM swings are large and persistent. If a channel’s cost per thousand moves enough that dollars and delivery diverge for long stretches, an exposure metric may describe it better.
  • Exposure data is clean and consistent. High-quality, well-defined delivery data across campaigns can improve a model, as long as spend is kept for the ROI denominator and budget simulation.

Even then, the decision should come from testing both inputs on your own data, which the next section shows how to do.

Spend vs. Impressions At a Glance

Consideration Spend Impressions
Represents Investment made Delivery as reported by each platform
Reconciliation Matches invoices and finance records Definitions differ by platform and format
Link to iROAS Direct: it is the denominator Needs spend added back in
Cross-channel comparison One monetary unit Units are not equivalent
Budget optimization Marginal return in dollars Needs a spend-to-exposure step
CPM volatility Needs controls and calibration Reflects delivery directly
Main limitation A dollar does not always buy the same delivery A count does not mean attention

Test It On Your Own Data

  1. Plot weekly spend and impressions for each channel and calculate CPM by week.
  2. Flag channels where CPM swings enough that spend and impressions diverge.
  3. For those channels, add price-pressure or seasonal controls to the spend model.
  4. Compare spend-based and impression-based versions on out-of-sample forecast accuracy and coefficient stability.
  5. Check the winner against an incrementality test. Correlation alone does not settle which input is better, and neither input establishes incrementality by itself.

How Lifesight Models Media

Lifesight is the Agentic Unified Marketing Measurement Platform. Our causal Marketing Mix Modeling models media spend by channel, so results arrive in the unit marketing and finance already share: iROAS.

Incrementality tests calibrate the model against what experiments show, and incrementality-adjusted attribution applies that calibration to platform-reported credit. Together they check the cost and counting shifts described above. Ask Lifesight and AI agents then help investigate the findings and prepare the next step: a recommendation your team reviews and approves.

Across 300+ brands, Lifesight customers see ~9.8% average sales lift.

Want to see what a causal, spend-based model says about your channels? Book a Demo.

Frequently Asked Questions

1. Is spend better than impressions for Marketing Mix Modeling?

For budget decisions, yes, and Lifesight models spend. It is the input you control, it reconciles to financial records, and it produces iROAS directly. Exposure data can still help when it is clean and consistent, so test both on your own channels.

2. What if CPM changes during the year?

Then a naive spend model can mistake higher prices for more advertising. Price-pressure controls, saturation curves, and incrementality calibration are how a spend-based model accounts for it.

3. What happens if a platform changes how it counts impressions?

An impression-based model sees its input shift even when delivery did not, and has to be recalibrated around the changeover. A spend-based model keeps the same input definition, so the series stays continuous. Amazon’s January 1, 2027 change to display impression counting is the current example.

4. Can a model use both spend and impressions?

Yes. Spend is needed as the ROI denominator in either approach. Using both as separate predictors of the same channel effect creates multicollinearity and identification problems unless the model is built to handle it.

5. Does a higher impression count mean better performance?

No. Placement, audience, creative, frequency, and existing demand all shape outcomes. Performance has to be judged on incremental business results.

6. Does modeling spend establish incrementality?

No. The input choice is not what makes a model causal. Incrementality comes from the model design and from calibration against experiments.