The Platforms Grading Their Own Homework: Why Your Attribution Data Is Structurally Broken
A peer-reviewed paper from NeurIPS 2025 formally proves what performance marketers have suspected for years — the mechanism that decides which of your ad platforms gets credit for your conversions is mathematically designed to be gamed.
The Number You're Building Your Business On
Every day, marketing leaders look at a dashboard that tells them which channels drove their sales. Google claims credit for some of them. Meta claims credit for more. TikTok takes a slice. CTV, email, and display each put a flag in the ground.
The numbers almost never add up to 100%. Sometimes they add up to 300%.
Most marketing organizations respond to this with a practice called "triangulation" — averaging conflicting attribution numbers, applying judgment factors, and moving on. It feels like sophistication. It is not. It is rounding error with a strategy deck attached.
A paper accepted at NeurIPS 2025, one of the most selective machine learning research conferences in the world, has now formally proved something that practitioners have suspected but couldn't articulate: the standard attribution mechanism is not just imprecise. It is mathematically broken in a specific, game-theoretic way. And the platforms running your advertising are the ones with the most to gain from keeping it that way.
What Last-Click Attribution Actually Rewards
The dominant attribution standard across performance marketing is Last-Click Mechanism, or LCM. The logic is simple: whichever advertising platform most recently reported a conversion gets 100% of the attribution credit for that sale.
The problem, as researchers Nan An, Weian Li, Qi Qi, Changyuan Yu, and Liang Zhang demonstrate in "Beyond Last-Click: An Optimal Mechanism for Ad Attribution," is that LCM is not Dominant Strategy Incentive Compatible.
DSIC is a concept from mechanism design theory. A mechanism is DSIC if every participant's optimal strategy is to report honestly, regardless of what other participants do. If a mechanism is NOT DSIC, then participants have an incentive to behave strategically rather than truthfully.
LCM fails this test. Here is why: if the last conversion report wins all attribution credit, every platform has a rational financial incentive to delay its own report until after its competitors have reported. The platform that times its reporting window correctly captures more credit, more budget, and more revenue. This is not fraud. It is optimization within the rules of a poorly designed game.
The result is attribution data that reflects the timing strategies of competing advertising platforms, not the actual sequence in which customers encountered your brand on their path to purchase. You are measuring which platform is best at gaming a reporting window. You are calling it marketing performance.
The Peer-Validated Mechanism: What an Actually Fair Attribution System Looks Like
The paper proposes a solution called the Peer-Validated Mechanism. The design principle is elegant: a platform's attribution credit is determined by what OTHER platforms report, not what it reports itself.
Under PVM, Google's attribution score is calculated using Meta's and TikTok's data. Meta's score uses Google's and TikTok's data. TikTok's score uses Google's and Meta's data. Because your self-report has zero effect on your own credit, the strategic incentive to game reporting windows disappears entirely. Honest reporting becomes each platform's dominant strategy regardless of what competitors do.
The paper proves three formal results. First, PVM is DSIC: honest reporting is the optimal strategy for every platform under any circumstances. Second, PVM outperforms LCM on both attribution accuracy and fairness in numerical experiments. Third, PVM is the optimal DSIC mechanism in homogeneous settings, meaning platforms with similar audience characteristics and conversion behavior.
This is a genuine theoretical advance. The NeurIPS acceptance alone signals it passed peer review from some of the most rigorous evaluators in machine learning research. The researchers have done what the industry has not: they have formally specified what an incentive-compatible attribution system looks like, and proved one exists.
Where the Mechanism Meets the Wall
Here is where practitioners need to read more carefully than the abstract.
The data-sharing problem. PVM requires each platform to observe and validate other platforms' conversion claims. But Google, Meta, TikTok, and every other major advertising platform operate inside walled gardens with entirely separate identity graphs. Meta cannot see Google's click data. Google cannot see Meta's view-through events. TikTok cannot see your CTV impressions. Post-iOS 14, the identity matching between platforms has degraded further. Peer validation requires peers to share observable data. In the current advertising ecosystem, they do not.
The consent problem. Implementing PVM at scale requires advertising platforms to submit to a neutral attribution mechanism that removes their ability to self-report. Google's Data-Driven Attribution is specifically engineered to credit Google inventory. Meta's Conversions API is built to maximize Meta's reported ROAS figures. These are not neutral measurement tools. They are products. No major platform has voluntarily adopted a third-party standard that reduces its own reported returns. PVM requires either regulatory mandate or contractual leverage that no advertiser currently possesses.
The homogeneity problem. PVM is provably optimal only in the homogeneous setting, where platforms have comparable audience characteristics and conversion behavior. Real advertising channels are profoundly different from each other. Google Search captures users with declared purchase intent, moments before a decision. Meta reaches passive scrollers who may convert a week later. TikTok drives discovery among audiences not yet in-market. CTV builds brand salience in users who will not convert for 90 days. These are not substitutes. Modeling them as homogeneous platforms for credit allocation produces mechanically fair attribution applied to structurally incorrect assumptions.
The causality problem. This is the deepest one, and the paper does not engage with it. Attribution credit is not the same as causal lift. PVM distributes credit more fairly across the platforms that touched a conversion. But touching a conversion is not the same as causing one. A user who would have purchased regardless of any advertising will still generate an attribution event. A channel that appeared in the customer journey but contributed nothing to the decision will receive PVM credit proportionate to its touchpoint count. The question that actually drives budget decisions is not "which platform touched this conversion?" but "which platform drove revenue that would not have occurred without the ad?" The only methodology that answers this question is incrementality testing: controlled geographic or time-based experiments with clean holdout groups, run at scale, with sufficient statistical power. Attribution models, whether LCM or PVM, measure presence in the customer journey. Incrementality testing measures necessity. Only the second question justifies moving budget.
The Unit Economics Bomb Hidden in a Game Theory Paper
The paper proves that LCM overcredits platforms that game reporting windows. It stops there. But practitioners need to follow the chain of consequences further.
If your primary attribution model systematically overcredits certain platforms, your cost-per-acquisition figures are wrong. You are dividing ad spend by conversion counts that do not accurately reflect which platform drove which purchase. This distorts your CAC calculation at the source.
If your CAC is wrong, your LTV/CAC ratio is wrong. If your LTV/CAC ratio is wrong, your payback period calculation is wrong. If your payback period is wrong, your budget allocation decisions, scaling thresholds, channel mix, and growth model are all built on a foundation that shifts under real operating conditions.
This is not a theoretical problem. Every quarter in which you optimize spend toward channels with artificially inflated attribution is a quarter in which you are likely underfunding channels with genuine causal impact and overfunding channels that are simply better at claiming credit. The researchers identified the root cause. They did not calculate the downstream damage. That is the practitioner's job.
What to Do Before PVM Is Implementable
The gap between the mechanism the paper proves is optimal and the mechanism the advertising industry is willing to build could be a decade wide. Here is what to do with this knowledge in the meantime.
Stop treating platform-reported attribution as a neutral signal. It is a number produced by entities with a financial stake in making it large. Build decision-making frameworks that treat attribution data as a directional indicator, not a ground truth. The number Google gives you about Google's performance is not a number Google has incentive to minimize.
Run geo-lift experiments on your highest-spend channels. Geographic holdout tests — where you withhold advertising from a matched market and measure sales lift against a control region — are currently the most practically accessible incrementality methodology. The gap between your attributed CPA and your incremental CPA is the most important number in your marketing stack, and most organizations do not know it.
Rebuild your budget allocation model around incrementality, not attribution. Attribution tells you who was present when a conversion happened. Incrementality tells you who caused it. These are different questions. For budget decisions above a certain threshold, only the second question should determine where money goes.
Build a measurement calendar, not just a reporting calendar. Most marketing organizations produce weekly attribution reports. Very few run incrementality experiments more than once a year. The cadence at which you test causal impact should match the cadence at which you make budget decisions. If you reallocate spend monthly, you need monthly incrementality signal, not annual lift studies.
The Longer Implication
The NeurIPS paper is notable not just for what it proves but for what it reveals about the state of an industry. Performance marketing has operated for fifteen years on an attribution standard that the research community has now formally proved is not incentive compatible. The platforms benefiting from that standard have had every reason to build better attribution tools and every reason not to build attribution tools that might reduce their reported returns.
The gap between what is mathematically optimal and what the industry actually uses is not an accident. It is the equilibrium outcome of a system where the measurement providers and the measured parties are the same entity.
PVM is not yet implementable at scale. But the proof that it exists and is optimal changes what practitioners should demand. A neutral attribution infrastructure, operating across platforms with shared observability and no self-reporting, is not a utopian idea. It is a mathematically specified mechanism that has been peer-reviewed and published. The question is whether the industry will build it before regulators mandate it.
Until then, the most honest thing a marketing leader can say about their attribution data is: this tells me which platforms claim credit for my conversions. It does not tell me which platforms caused them. Those are different sentences. The distance between them is where most marketing budgets are lost.
Source
Nan An, Weian Li, Qi Qi, Changyuan Yu, Liang Zhang. "Beyond Last-Click: An Optimal Mechanism for Ad Attribution." 39th Conference on Neural Information Processing Systems (NeurIPS 2025). arXiv:2511.22918.
More articles
View all →Incrementality Testing 101: What Every E-Commerce CMO Needs to Know
Incrementality is the question every marketing team should be asking: would these customers have converted without our ads? Here's how to find out — without a data science team.
The A/B Test You're Running Is Wrong: A Guide to Statistical Power
Most A/B tests stop too early, run with too little traffic, or declare winners on noise. Here's how to design tests that actually tell you something true.
CAC Reduction: The 4-Step Framework That Cut Acquisition Costs by 35%
A step-by-step breakdown of how we helped one DTC brand identify and eliminate non-incremental spend — reducing CAC by 35% without cutting revenue.
Ready to prove your marketing ROI?
Book a free 30-minute consultation. No commitment, just 30 minutes of clarity on what's actually driving your results.
Book Free ConsultationNo commitment. Just 30 minutes of clarity.