What is the best way to measure performance marketing?

Triangulate three layers. Platform attribution for daily steering, incrementality tests for proof, and media mix modeling for quarterly budget strategy. No single layer is trustworthy alone.

How often should you run incrementality tests?

One clean test a month or quarter, starting with your biggest or most doubted channel. Write the kill-or-scale thresholds down before the test starts.

Is media mix modeling only for big brands?

Not anymore. Open tooling such as Google's Meridian made quarterly model refreshes realistic for mid-size brands.

RGM® Learn · Measurement

Dashboards steer.
Tests prove. Models plan.

How to Measure Performance Marketing — the Triangulation Playbook

Three measurement layers, each checking the others' blind spots. Here is how to run them together so your budget follows truth, not the loudest dashboard.

You measure performance marketing with three layers. Platform attribution answers "what should I change today?" Incrementality tests answer "did this spend cause new revenue?" Media mix modeling answers "where should next quarter's budget go?" Use one alone and it will eventually lie to you.

By David Schaefer · LinkedIn · Updated June 2026

The numbers behind the problem

3x+

How far observational attribution can mis-state true lift · Marketing Science 2019

Short-term incremental sales from brand-keyword ads in eBay's experiments · Econometrica 2015

52%

US marketers already running incrementality testing · eMarketer 2026

60%

Senior decision-makers who trust independent lift tests most · eMarketer 2026

Why one measurement source is never enough

Every tool answers a different question, and every tool has a blind spot. Platforms grade their own homework. Models smooth over the details. Experiments are slow and expensive. The fix is not picking a winner — it is making them check each other.

Here is the uncomfortable evidence. Researchers ran 15 large advertising experiments on Facebook — 1.4 billion impressions — and compared the true, experimentally measured lift against what standard attribution methods reported. The attribution numbers were often wrong by a factor of three or more, in both directions.

Claim: Common observational attribution approaches often mis-state experimentally measured ad lift by 3x or more. Source: Gordon, Zettelmeyer, Bhargava & Chapsky — Marketing Science (2019). Context: if budget moves on dashboard ROAS alone, you are scaling some channels that deserve cuts — and cutting some that deserve scale.

The line attributed to retailer John Wanamaker — that half his advertising money was wasted, he just never knew which half — is more than a century old. The tools changed. The question did not.

A short history of knowing what worked

Marketing measurement swings on a pendulum: from direct response, to statistical models, to user tracking, and now back to models plus experiments. Knowing the history tells you why triangulation won.

1923

Coupons and split runs. Claude Hopkins publishes Scientific Advertising: track every response, test every claim. Direct response becomes a discipline.

1960s

Econometrics arrives. Consumer giants begin modeling sales against media spend statistically — the ancestor of today's media mix modeling.

2000s

The last-click era. Search ads and web analytics make per-click tracking cheap. Last-click attribution becomes the default — and quietly over-credits the bottom of the funnel.

2010s

Multi-touch attribution rises. User-level journey stitching promises credit for every touch. It works only as long as users can be tracked across the web.

2021–25

The signal collapses. Apple's App Tracking Transparency prompts (2021) cut mobile tracking; Google retires Privacy Sandbox and keeps third-party cookies user-controlled (Oct 2025). User-level MTA loses its raw material.

Now

Triangulation. The industry returns to models and experiments — modernized: Bayesian MMM (Google open-sourced Meridian), always-on geo holdouts, and server-side event routing. 46.9% of marketers plan to invest more in MMM (eMarketer, 2026).

The three layers, in depth

Think of it as an instrument stack. Layer one is the speedometer — always on, roughly right. Layer two is the inspection — slower, but it finds the truth. Layer three is the map — it plans the route the other two cannot see.

L1Daily

Platform attribution

The reporting inside Google, Meta, TikTok. Fast, free, directional. Trust it for in-channel calls: creative, bids, audiences. Never for budget shifts between channels — it cannot see the other channels and it grades its own homework.

L2Truth

Incrementality & holdouts

Pause or vary spend in matched geographies and compare against baseline. The closest thing media has to a clinical trial. Slow and costly — so you aim it at the biggest or most doubted line item first.

L3Strategy

Media mix modeling

A statistical, top-down read of how every channel drives revenue — including TV, audio, and anything attribution cannot track. Privacy-proof by design: it uses spend and outcomes, not user data.

FIG. 01 — Triangulated measurement: each layer corrects the others' blind spots

"All models are wrong, but some are useful."
— George E. P. Box, statistician · on models

Where the other frameworks fit

You will meet four other frameworks in the wild. Each has a job — and a failure mode. Here is the honest scorecard.

Measurement frameworks — what to use, what to retire
Framework	What it does	Status in 2026
Last-click attribution	Full credit to the final click	Retire. Systematically over-credits search and brand terms
Multi-touch attribution (MTA)	Splits credit across tracked touches	Diminished. Signal loss removed its raw material; useful mainly on owned, logged-in journeys
Blended metrics (MER, blended CAC)	Total revenue ÷ total spend	Keep. The un-gameable scoreboard — pairs with everything
Marginal analysis	Return on the next dollar, not the average dollar	Keep. Scaling decisions live on the margin; averages flatter big channels

One distinction does heavy lifting: average versus marginal return. A channel can show a strong average ROAS while its next dollar earns almost nothing — auctions saturate. Holdouts and MMM response curves both estimate the margin; dashboards mostly report the average.

The six-step playbook

Triangulation is a calendar, not a slogan. Six steps put it on rails: instrument, define one truth, baseline, test, recalibrate, govern.

Instrument. Send conversions server-side alongside the pixel with shared event IDs so platforms deduplicate (Meta's CAPI dedup; Google enhanced conversions). One conversion definition everywhere. Alert on event-volume drops — broken tracking starves the algorithms silently.
Define one truth. Appoint a referee for orders: your analytics or commerce backend. Platform numbers become claims to verify, never the scoreboard. This single decision ends most attribution arguments.
Baseline. Know organic traffic, repeat-purchase rates, and seasonality before crediting any ad. Without a baseline, every channel "works."
Test. One clean geo holdout at a time, sized for statistical significance, kill-or-scale thresholds written down before launch. Start with the biggest or most doubted line item — branded search is the classic first target.
Recalibrate. Apply each test's correction factor to dashboard ROAS until the next test. If the holdout says retargeting is 30% incremental, its dashboard 5.0x is a true 1.5x — and every weekly report should say so.
Govern. Dashboards daily. One holdout a month or quarter. Model refresh quarterly. Decisions documented with the layer that justified them.

Worked example — recalibration in action. A brand spends $100k/mo. Retargeting takes $30k and claims $150k in attributed revenue — a 5.0x ROAS. A four-week geo holdout pauses retargeting in matched markets: revenue there falls only $45k versus baseline. True incrementality: $45k ÷ $150k = 30%. True iROAS: $45k ÷ $30k = 1.5x. If breakeven is 2.0x, the "best channel on the dashboard" is under water — and roughly $30k/mo was waiting to be reallocated to channels that survive their own holdouts. (Illustrative model — RGM analysis; method per the studies cited above.)

The mistakes that undo it

Testing once, then coasting. Incrementality decays — auctions, creative, and competitors move. A correction factor older than two quarters is folklore, not measurement.

Holdouts too small to read. An underpowered test returns noise, and noise gets read as whatever the loudest stakeholder hoped. Size for significance or do not run it.

Treating MMM as an oracle. A model is a hypothesis generator. Validate its biggest claims with an experiment before moving budget — remember Box.

Measuring everything, deciding nothing. The point of three layers is faster, braver budget decisions. If reports grow but allocations never change, the system failed.

Quick answers

What is the best way to measure performance marketing?: Triangulate: platform attribution for daily steering, incrementality tests for proof, media mix modeling for quarterly strategy. No single layer is trustworthy alone.
How often should you run incrementality tests?: One clean test a month or quarter, starting with the biggest or most doubted channel. Thresholds written down before launch.
Is media mix modeling only for big brands?: Not anymore. Open tooling such as Google's Meridian made quarterly refreshes realistic for mid-size brands.

Frequently asked

Why does platform ROAS differ from real results?

Platforms only see their own touchpoints and grade their own homework. Published research found observational attribution can mis-state true lift by 3x or more.

What happened to multi-touch attribution?

Signal loss broke it. App tracking prompts and browser privacy changes removed much of the user-level data MTA needs, so the industry moved back to models and experiments.

What is triangulated measurement?

Using attribution, incrementality experiments, and media mix modeling together, letting each method check the others' blind spots.

What should a small team start with?

A referee system for orders, blended CAC as the scoreboard, and one geo holdout on the biggest spend line. That trio beats most enterprise stacks in honesty per dollar.