Insights
Articles on conversion optimization, A/B testing, attribution, incrementality, and marketing measurement.
- The measurement reckoning: what four new papers reveal about the metrics you are actually trusting — A new correction framework deployed at TikTok proves attribution overstates its own impact. A Baidu model trained on 15 million users proves one LTV number for your whole customer base is a liability. Read together with two more 2026 papers, the pattern is one the marketing conference circuit keeps avoiding.
- The measurement stack is getting smarter and less useful at the same time — New 2026 research on causal attribution, graph-based lifetime value models, and AI-driven conversion rates is more technically rigorous than anything the industry has produced in a decade. It is also drifting away from what most growth teams can use. Here is the gap, and the one question that closes it.
- Geo-lift testing: the complete guide — Geo-lift testing is the most accessible form of incrementality testing, and the one most teams run first. This guide covers the full mechanics: how to pick test and control markets, how much spend and duration you need for statistical power, how to read the results, and the mistakes that invalidate a test before it starts.
- MMM vs. attribution vs. incrementality testing: which to use when — Marketing mix modeling, multi-touch attribution, and incrementality testing answer three different questions, and teams that treat them as competing answers to the same question misallocate budget. This guide maps each method to the decisions it can support, the decisions it cannot, and the order in which to build them.
- Your attribution model has never run a real experiment — Most attribution models have never produced a causal measurement in their operational history. They find correlations, dress them in causation language, and call the output strategy. This is how budgets get reallocated with precision toward the wrong channels. Here is the framework for telling whether your model is lying to you.
- Why most A/B tests don't change anything — Most A/B tests fail before the experiment runs. The hypothesis is missing, the statistical power is insufficient, and no one has specified what the result means before it arrives. This is not an optimization problem. It is a design problem, and it explains why most experimentation programs produce high test volume and low organizational learning.
- The attribution problem no one talks about — Every attribution model in wide use answers the same question: which channels were present in the conversion path? None of them answer the question that matters: which channels caused the conversion? The gap between those two questions is where most marketing budget gets misallocated.
- What CMOs need from growth analytics — Marketing analytics has produced more dashboards and fewer decisions. The problem is not data availability. It is a structural misalignment between the metrics that are easy to measure and the metrics that would change how a marketing organization allocates budget and headcount.
- Incrementality testing for companies that don't have Netflix's budget — Incrementality testing has a reputation for requiring platform-scale infrastructure and a large media budget. That reputation is wrong. The geo-holdout test is within reach of any company spending $50,000 per month on paid media, and it produces evidence that no attribution model can match.
- Conversion rate benchmarks are mostly useless — Conversion rate benchmarks aggregate data across companies with different products, traffic sources, price points, and funnel structures. The resulting number is not a meaningful reference point for any specific company. Self-benchmarking is almost always more useful, and most teams never do it properly.
- The decision criteria problem in experimentation — Most teams know when a test reaches statistical significance. Almost none have specified in advance what that result means or what action it requires. Without pre-specified decision criteria, positive results get shipped, negative results get rationalized, and experimentation programs accumulate velocity without accumulating knowledge.