Quick interview answer
Model realistic user journeys with thread groups and pacing, parameterise data with CSV Data Set, correlate dynamic tokens, then read 90th/95th percentile latency, throughput and error rate — never the average.
Detailed explanation
Structure: thread group with ramp-up matching production arrival rate, HTTP Request Defaults, cookie and cache managers, CSV Data Set for unique users, regex or JSON extractors for tokens, and assertions so failures are counted properly. Run non-GUI (jmeter -n -t) with listeners off, generate the HTML dashboard, and compare against a baseline. Interpretation focuses on percentile latency, throughput plateau, error onset and resource saturation on the server side.
1jmeter -n -t checkout.jmx -Jusers=200 -Jrampup=120 -l results.jtl -e -o report/Real-world example
A 200-user ramp exposed a connection-pool limit at 140 users: throughput flattened while latency tripled.
Interview tip
Explicitly reject averages in favour of p95 — it is the fastest credibility signal in performance rounds.