Every performance budget eventually meets a skeptical CFO, and the industry’s answer has long been a handful of famous statistics, milliseconds costing percent of revenue, quoted far beyond their original context. The honest version is more useful: the studies are real, directional and non-transferable, and your own number is measurable.
What the canon actually established
The celebrated experiments, retail giants delaying pages artificially, search engines slowing results, travel sites measuring conversion against latency cohorts, consistently found the same shape: measurable engagement and conversion loss as latency rises, with effects compounding at the slow tail rather than the median. What they did not establish is a universal exchange rate; the measured coefficients varied by product, funnel step, user intent and era. Directionally settled, numerically local: that is the honest reading.
Why transferability fails
Latency sensitivity concentrates where intent is fragile: browsing, comparison, impulse. It flattens where intent is committed: nobody abandons a tax filing over 300ms. Your product mixes both, in proportions no other company’s study shares, which is why borrowed coefficients produce budgets nobody trusts. The good news buried in this: the same RUM infrastructure this series keeps building measures your own coefficient as a side effect.
The tail-concentration finding deserves emphasis because it redirects spending: in nearly every credible study and every cohort analysis we have run, the revenue damage lives disproportionately in the worst-experience percentiles, the p90-plus sessions, not in shaving medians that were already fine. That inverts the common optimization instinct (median-chasing demos well; tail-fixing pays) and it re-derives this series’ entire technical agenda from commercial first principles: congestion control, protocol upgrades, shielding, warming and steering are all tail interventions, which is precisely why they kept appearing here. The millisecond worth buying is the ugly one at the distribution’s edge, and now you know how to price it.
Measuring your own exchange rate
The method is cohort correlation done carefully: segment real sessions by experienced latency (from your RUM percentiles), compare conversion and engagement across cohorts, and control for the confounders that fake the effect, geography and device correlate with both latency and purchasing power, so naive correlation flatters latency’s importance. Stronger designs use natural experiments (regional incidents, provider migrations, deploy-induced regressions) as instrumented variation. The output is the only statistic that survives your CFO: milliseconds priced in your own funnel’s currency.
In practice
Run the cohort analysis on your last quarter of RUM and conversion data; most estates find a real, product-specific latency-revenue slope concentrated in specific funnel steps and specific percentile bands. Then budget precisely: spend on the steps and tails where your slope is steep, decline gracefully where it is flat, and retire the borrowed statistics from your slide decks, your own number is smaller than the legends and infinitely more persuasive.
Latency-revenue cohort analysis is an assessment module here: your RUM, your funnel, your exchange rate, documented.