Identity Errors Can Erase ~70% of Campaign ROI: The Study

Identity resolution errors can erase roughly 70% of measured campaign ROI — that is the headline finding of a new LiveRamp and MMA study presented on July 20, 2026, and it comes with a caveat most coverage will bury: the number is a synthetic-data simulation outcome, not a measurement from live campaigns.

The stakes are real even if the data is modeled. In the study's 50%-precision scenario, a campaign genuinely returning $1.50 per dollar reads as $0.43 — a result that would get a profitable channel defunded. And the sharpest finding is not the headline number at all: losing just ~1% of impression volume in a non-random way reversed channel rankings in the simulation, while the standard diagnostics teams use to sanity-check their models actually improved.

This guide separates the study's two easily-conflated collapse figures, explains why the pattern of data loss matters more than the volume, walks the measurement-method resilience hierarchy the modeling produced, and names the conflict of interest plainly: LiveRamp co-authored the research and sells the fix it recommends.

The report is titled "The Missing Piece: Improving Confidence in Marketing Measurement," jointly published by LiveRamp and the MMA (Marketing + Media Alliance). It was presented at the MMA CMO + CEO Summit 2026 in Santa Barbara, California, during a session on data loss and measurement, and covered same-day by PPC Land. Worth stating plainly: PPC Land is the sole detailed source for the study's figures at time of writing — the MMA's own event page confirms the summit and session but has not independently republished the study text.

The methodology is fully synthetic. The research team built a test dataset of 1.9 million impressions across 147,941 users and four publishers, with 12,956 transaction outcomes and a 1.82% empirical conversion rate. They then degraded that dataset in controlled ways — deleting exposures, mismatching identities — and watched what each measurement method reported versus the known ground truth. Because the true ROI is defined by construction, the simulation can quantify exactly how far each method drifts. That is the design's strength. Its weakness is the obvious one: no live campaigns were measured.

Most secondary coverage will blur the study's two headline numbers into one. They are distinct metrics from the same scenario. When identity precision drops to 50% in the modeling, a campaign's return — genuinely $1.50 per dollar spent — is measured as $0.43. That is a collapse of roughly 71% in the measured ROI, the basis of the study's ~70% headline. In the same simulated scenario, the campaign's lift — genuinely 25% — is measured as 6.8%, a collapse of roughly 73% in the measured lift figure specifically.

Why does this distinction matter beyond pedantry? Because the two figures answer different budget questions. The ROI collapse tells a CFO the channel appears to lose money when it doesn't. The lift collapse tells a growth team the campaign appears barely incremental when — in the modeled world — it is highly incremental. If you quote one number, label which one, and keep the simulation qualifier attached. The study earns credibility precisely by hedging; coverage that strips the hedge is doing the vendor's marketing for it.

The simulation's most counterintuitive result concerns how data goes missing, not how much. When the researchers randomly removed 20% of impressions, the baseline channel ranking held — the model still identified the right winners. But when loss was concentrated among converters — the users who actually bought — channel rankings reversed at roughly just 1% aggregate volume loss. In the modeled world, a twentieth of the missing data did twenty times the damage, because it was the wrong twentieth.

The study names two distinct failure modes behind that non-random loss, and they are worth keeping separate because they demand different fixes:

The report distills this into three claims it argues CMOs need to internalize: the pattern of data loss matters more than the volume of data loss; identity precision matters more than identity match rate; and randomized controlled trials hold up structurally where other measurement methods break. All three are simulation-derived — but the first two, at least, echo what practitioners see when a "healthy" attribution model quietly starts recommending the wrong channel mix.

Here is the detail most write-ups of this study will skip, and the one operators should sit with: in the simulation, standard model-fit diagnostics — AUC, the workhorse metric teams use to confirm an attribution model is healthy — actually improved even as outcome-correlated data loss corrupted the channel rankings. The model looked better on paper while giving worse answers.

The mechanism is intuitive once stated. When data loss concentrates among converters, the surviving dataset becomes cleaner and more separable — easier for a model to fit — while systematically under-representing exactly the exposures that drove conversions. Fit metrics measure how well the model explains the data it can see. They say nothing about whether the data it can see still resembles reality. In the modeled scenario, those two things moved in opposite directions.

The operational translation: if your model's fit metrics look fine but your channel rankings just flipped — a channel that led for quarters suddenly reads as a laggard with no creative, bidding, or market change to explain it — that is exactly the failure signature this simulation describes. The diagnostic that would catch it is not a better fit statistic; it is an independent ground-truth check, which in practice means incrementality testing and causal lift measurement run alongside whatever attribution stack you already trust.

The simulation's second contribution is a resilience hierarchy: it subjected different measurement designs to the same non-random (outcome-correlated) data loss and recorded what each one reported against the known $1.50 true return. Randomized controlled trials sat at the top — the study's gold tier. Quasi-experimental models landed in a moderate-vulnerability silver tier. Fixed-formula, rules-based attribution — the last-touch and position-based logic still underpinning many dashboards — was the least resilient bronze tier. The original coverage presents these as prose findings; the table below repackages them into the decision reference we wanted and couldn't find.

The gap column is our arithmetic from the study's reported values: the reached-only design's $0.91 sits $0.59 below the $1.50 truth, an understatement of about 39%; the quasi-experimental design's −$2.65 represents a $4.15 swing and a sign flip. Directionally, this hierarchy matches what the broader measurement literature has argued for years — see our marketing mix modeling vs. attribution playbook and the fuller guide to choosing the right measurement method. What the simulation adds is a mechanism: it is not that fixed-formula multi-touch attribution modeling is lazy math — it is that rules-based methods structurally cannot detect when their input data has been non-randomly corrupted, while an intent-to-treat RCT is insulated by design.

Now the part most coverage treats as a footnote. The study's recommended remedy for the precision problem it diagnoses is data collaboration — clean rooms and identity resolution services. That is LiveRamp's product category: identity resolution via RampID is the company's core offering. LiveRamp co-authored the research that concludes marketers need what LiveRamp sells. Its VP of Product, Christine Grammier, made the framing explicit at the summit, arguing that "A strong data foundation is critical for unlocking the best data to power your AI and agentic tools." None of this makes the simulation wrong — the methodology is described transparently and the report hedges its own findings — but it is a conflict of interest, and you should read the conclusions the way you would read any vendor-funded benchmark.

There is a quieter irony worth noticing. LiveRamp's own marketing has long advertised a 99.5% figure for its identity graph match rate — and this study's second "truth" is that match rate is the wrong metric to optimize, because precision matters more. In effect, the research reframes the industry's favorite vanity metric — one its own co-author has marketed on — toward a harder standard that happens to favor premium identity infrastructure. That reframing may well be correct. It is also commercially convenient.

Two more sourcing notes. First, LiveRamp's corporate context: the company is NYSE-listed and, per the same coverage, subject to a pending acquisition by Publicis Groupe — meaning the study's co-author is on its way into one of the world's largest agency holding companies, which itself sells measurement and identity services. Second, the widely-quoted companion stat — 67.4% of marketers ranking "proving incremental ROI" as their top challenge — comes from TransUnion/EMARKETER research from October 2025, cited in the same coverage as supporting context. It is a separate piece of research from a separate organization, not a finding of the LiveRamp/MMA simulation.

Strip the vendor framing and a genuinely useful checklist remains. The simulation identifies failure modes that are plausible, cheap to test for, and mostly fixable without buying anything. Here is how we would sequence it:

Looking forward, the precision question is about to get more consequential, not less. As buying itself becomes agentic — the emerging DV/LiveRamp/Pixalate agentic ad-buying stack we covered earlier this month puts optimization decisions inside automated loops — corrupted measurement stops being a quarterly reporting problem and becomes a real-time budget-steering problem. An agent reallocating spend on a channel ranking that flipped because of non-random data loss will compound the error at machine speed. If the simulation's dynamics hold even partially in the field, measurement integrity becomes a prerequisite for agentic buying, not a nice-to-have. That projection is ours, not the study's — but it is the reason this research matters beyond one conference session.

If you want a second set of eyes on your own stack, this is the work our analytics and measurement engagements are built for — auditing attribution inputs, designing holdout tests, and wiring incrementality checks into paid media operations so budget decisions rest on something sturdier than a well-fitting model.