Quantifying Impact

"It made things faster" is a claim. "p95 went from 1.9s to 380ms on the search endpoint, measured in production over the following week" is evidence. The second is not more impressive because the number is big — it's more impressive because it demonstrates you were measuring at all, which is the actual senior behaviour.

Where numbers come from when you think you have none. Latency and error rates from dashboards. Build and deploy times. Bundle size and Core Web Vitals. Support ticket volume on a flow. Conversion or drop-off at a step. Time-to-first-PR for new hires. Incident count or on-call pages per week. Cost per thousand requests — increasingly load-bearing for LLM features, where a prompt or model change moving spend from $4k to $900 a month is a genuinely senior result. Even "three engineers stopped spending Mondays on this" is quantified.

When you truly don't have a number, say what you observed and how: the queue stopped growing, the flaky test disappeared from the last 200 CI runs, the on-call rotation stopped getting paged for it. Direction plus evidence is honest and lands fine. Inventing precision is the one unrecoverable move here, because interviewers probe numbers and a fabricated one collapses immediately.

Be honest about attribution. If the redesign shipped alongside a pricing change, don't claim the revenue. "Conversion went up 12%; I can only attribute part of that to the perf work, but the flow I owned went from 6% to 2% drop-off" is a far stronger answer than a clean claim you can't defend, and it demonstrates exactly the analytical care the number was supposed to signal.

Note also that the interesting number is often the counterfactual — what the thing cost, not just what it gained. "It took three weeks and we deferred the mobile work to do it" shows you knew the price.

Senior candidates are expected to connect work to outcomes, and it's the most common gap in otherwise strong stories. Numbers also protect you in the deep dive: an answer anchored in measurement invites follow-ups you can answer, whereas an unquantified claim invites the one you can't.

• Pair every result with how you know it — the dashboard, the window, the population.
• Sources you probably already have: latency, error rate, build time, bundle size, ticket volume, drop-off, pages, cost per request.
• LLM cost and token spend are first-class impact metrics at these companies.
• No number is fine; fabricated precision is not, and it fails on the first follow-up.
• Direction plus observable evidence ("stopped paging", "no flakes in 200 runs") is a legitimate result.
• Split attribution honestly when other changes shipped alongside — the caveat strengthens the claim.
• Say what the work cost as well as what it returned.
• Capture numbers while you still have dashboard access; they're unrecoverable after you leave.