The measurement reckoning: what four new papers reveal about the metrics you are actually trusting

By Mo Touzani · 2026-07-23 · Marketing Measurement

Every growth team has a dashboard they quietly do not fully believe. The attribution number that feels too generous. The LTV projection that assumes last year's customer still exists. The CAC target that got set once and never revisited once the channel mix changed underneath it. Most teams live with that discomfort because interrogating the model feels like a distraction from the work of hitting quarterly numbers. Four recent research papers make that discomfort harder to ignore.

Your attribution number is inflated, and now there is a figure attached

Start with attribution, the metric most growth leaders trust the least and use the most. A team at TikTok built a correction framework that does something almost embarrassingly obvious in hindsight: it measures how much of your paid attribution is actually just capturing demand that would have converted organically anyway, a phenomenon called cannibalization, and corrects for it using real incrementality experiments as ground truth. Deployed at scale, it cut measured cannibalization by 15 percentage points. That number matters because it quantifies something every experienced media buyer has felt but rarely proven: your attribution dashboard has been overstating the case for your own budget the entire time.

The catch is real. This correction needs experiment volume to calibrate against. TikTok can run the incrementality tests required to build it because TikTok has TikTok's scale. A team running two holdout tests a year has no reference points to correct anything. The honest reading is not "go build this." It is "understand why your dashboard number is inflated, and push for the minimum experiment cadence needed to know by how much."

Attribution without a cannibalization correction is not measuring your channels. It is quietly lying to you every single reporting cycle.

A Bayesian bridge, built on a foundation nobody checks

A second paper offers the more realistic path for teams without hyperscale budgets. Instead of relying on user-level tracking, which privacy regulation keeps eroding, a Bayesian framework uses your marketing mix model's channel-level output as a prior, then layers campaign-specific detail on top. It solves a genuine and long-standing conflict: MMM is privacy-safe and stable but too coarse to act on, multi-touch attribution is granular but needs tracking that is disappearing. The framework threads that needle mathematically.

What it cannot do is fix a bad MMM. A Bayesian prior built on a misspecified base model just produces a more confident-looking wrong answer. If your marketing mix model already ignores a competitor promotion or gets seasonality wrong, this framework does not resolve that. It bolts a precise-looking number onto an imprecise foundation and calls it solved. The math is elegant. The dependency it creates on getting the underlying model right is easy to skip past in a headline.

Your average customer does not exist

Then there is lifetime value, where the laziest habit in marketing analytics still lives: one model, one number, applied to a customer base that obviously is not one thing. A team at Baidu tested that assumption directly, at a scale that matters, 15 million real users, not a sanitized benchmark. Their model treats demographic segments as genuinely different distributions and treats behavior as a fast-moving sequence rather than a fixed snapshot, combining both through a graph and transformer architecture built specifically for that shape of problem. It beat every static alternative tested against it.

The finding is simple to state and expensive to ignore: your customer base is not one population, and modeling it as though it is costs you precision you could otherwise have. What the paper does not solve is just as important. None of this touches the CAC side of the ratio. A brilliant LTV model paired with an undisciplined acquisition budget still produces a business that spends its way into a bad payback period. LTV precision without CAC discipline is a scoreboard, not a strategy.

The agent that automates the how, not the why

The final paper should reshape how growth leaders think about AI's actual role here, as opposed to the role vendors keep pitching. AgentLTV does not predict lifetime value directly. It automates the process of building the pipeline that predicts it, using an LLM-driven search across modeling choices, refined through an evolutionary process, deployed as executable code the agent writes and repairs itself. That is a genuine capability shift, and it is also exactly where the danger sits.

An agent optimizing a pipeline will optimize toward whatever metric you told it to hit. It has no opinion on whether that metric is the right one for your business, your retention motion, your risk tolerance. Automating the search does not automate the judgment. Any team adopting this kind of tooling needs a clear, human answer to what it is actually optimizing for before handing the search over to an agent, because the agent will answer that question with total confidence and zero context.

What the four papers say when read together

Put the four together and a pattern emerges that the industry's thought leadership circuit tends to skip. Correlation-based measurement, the click that gets full attribution credit, the LTV model trained once and left alone, is not a simplification anymore. It is a liability. The research converging in 2026 says the same thing from four different directions: attribution needs an experimental anchor, LTV needs to account for who your customer actually is right now, and AI's real value is speeding up the mechanical search for better models, not replacing the judgment about what those models should optimize for.

The framework for acting on this

1. Find the metric your team trusts most without ever asking why. CAC payback, blended ROAS, whatever number gets quoted in the board deck without a second thought. Ask what it would take to prove that number with an actual experiment instead of a model's assumption. If the honest answer is "we have never tested it," that is the finding, not a vendor pitch to buy your way out of it.
2. Segment your LTV model before you upgrade your attribution stack. A single LTV number for the whole customer base is the more common and more fixable error. Split by the two or three segments with the most obviously different value curves before reaching for graph neural networks or agent-searched architectures built for hundred-million-user platforms.
3. Write down what an automated agent is optimizing for before you deploy one. If a modeling agent is going to search pipelines on your behalf, the business, not the agent, decides what tradeoff between false positives and false negatives the model should accept. Make that decision explicit and in writing before the search starts, not after you are surprised by the result.

The actionable takeaway

If you are a founder or CMO deciding where to spend the next quarter's analytics budget, the useful version of all this is not "adopt AI attribution" or "buy an LTV platform." It is narrower and less comfortable. Find the one metric your team currently trusts the most without question, and find out exactly why you trust it. If the answer is not an experiment, it is a story you have told yourselves long enough to believe.