[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-sample-ratio-mismatch::en":3,"gloss-cluster-sample-ratio-mismatch::en":23,"gloss-next-sample-ratio-mismatch::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"sample-ratio-mismatch","analytics","Sample Ratio Mismatch (SRM)","A sample ratio mismatch is a difference between the traffic split an experiment was configured for and the split it actually delivered. A test set to fifty-fifty that finishes with 51.4 percent of users in one arm looks like rounding and is not: at any meaningful sample size a deviation that large is wildly unlikely by chance, which means something assigned users non-randomly. The check itself is one line of arithmetic against the expected proportions, and it is the most valuable integrity test an experimentation program can run. It matters because the mismatch is a symptom rather than a rounding problem, and whatever caused it almost always biases the result too. A redirect that loses slow connections on one arm. A bot filter that treats the variant differently. Assignment logged after the page renders, so the users who bounced fastest are missing from one side. An arm that errors more often, whose failed sessions therefore never report at all. In every one of those cases the missing users are not a random sample, they are the ones who had the worst experience, so the surviving arm looks better than it is and the effect is measured on a population that was quietly selected. The discipline is to run the check before reading the result and to treat a failure as disqualifying rather than as a caveat. An experiment with a mismatch cannot be rescued by adjusting for it, because the mechanism that caused the imbalance is unknown and nothing tells you which direction it pushed the metric. The correct response is to find the cause, fix it and re-run. Running the check automatically on every experiment, rather than when a result looks surprising, is what keeps it honest: a mismatch on a result that confirms expectations is exactly the one nobody goes looking for.","A sample ratio mismatch means the traffic split was not the one you configured, and why that disqualifies an experiment instead of adding a caveat to it.",null,[11,14,17,20],{"slug":12,"name":13},"ab-testing","A\u002FB Testing",{"slug":15,"name":16},"guardrail-metric","Guardrail Metric",{"slug":18,"name":19},"statistical-significance","Statistical Significance",{"slug":21,"name":22},"tracking-plan","Tracking Plan",[24,26,29,33,36,39,42,43,46,49,52,55],{"slug":12,"category":5,"name":13,"updated_at":25},"2026-08-24T02:46:38+00:00",{"slug":27,"category":5,"name":28,"updated_at":25},"active-user","Active User (DAU, WAU, MAU)",{"slug":30,"category":5,"name":31,"updated_at":32},"autocapture","Autocapture","2026-08-24T02:46:37+00:00",{"slug":34,"category":5,"name":35,"updated_at":25},"cost-per-resolution","Cost per Resolution",{"slug":37,"category":5,"name":38,"updated_at":32},"customer-data-platform","Customer Data Platform (CDP)",{"slug":40,"category":5,"name":41,"updated_at":25},"deflection-rate","Deflection Rate",{"slug":15,"category":5,"name":16,"updated_at":25},{"slug":44,"category":5,"name":45,"updated_at":32},"identity-resolution","Identity Resolution",{"slug":47,"category":5,"name":48,"updated_at":32},"multi-touch-attribution","Multi-Touch Attribution",{"slug":50,"category":5,"name":51,"updated_at":25},"novelty-effect","Novelty Effect",{"slug":53,"category":5,"name":54,"updated_at":32},"retention-curve","Retention Curve",{"slug":56,"category":5,"name":57,"updated_at":25},"seat-utilization","Seat Utilization"]