The method · measurement notes
The proof is the instrument. Run it yourself.
Anyone can screenshot one good answer. We don’t. If we were paying for this work, this is the standard we’d ask for.
Ask twice and you get two different answers.
AI answers are probabilistic. The same question, asked the same way, can name your firm on Tuesday and skip you on Thursday. A vendor waving one screenshot is showing you their luckiest run, and a scary screenshot proves just as little as a flattering one.
The question that matters isn’t whether AI named you once. It’s how many of the answers buyers get actually name you. That’s a rate, and you can only get a rate from a decent sample.
The question set, sampled
- Q1who's the best interior designer in singapore?
- Q2which ID firm should renovate my 4-room BTO?
- Q3reliable interior designer that won't blow my budget?
- Q4condo renovation under $60k, which firm?
- Q5minimalist HDB renovation, which studio?
- Q6interior designer or contractor for a resale flat?
+ your firm’s own buying questions, added at intake
The sampling protocol
M.1
Questions come from real homeowner phrasing.
We build the question set from the way buyers actually type: budget wording, property type, style words. 48 distinct questions per firm.
M.2
Each question runs many times, never once.
The same engine answers the same question differently on different runs, and the frontier models underneath change without notice, which is one more reason a single answer proves nothing. Mitra repeats until the sample reaches 240 answers, spread across the window rather than taken in one sitting. Measurement runs on Perplexity today, with more engines on the roadmap. Your report states exactly which engines the runs used, and re-measurement uses the same ones, always.
M.3
We record firms named and sources quoted.
For every answer we log which firms were named, in what order, and which pages or publications the engine leaned on. No sentiment scores or invented metrics, just counts.
M.4
The result is a rate with a margin of error.
Your citation share arrives as a rate with a 95% confidence interval. When we re-measure, the same protocol runs again, so movement is real movement and not sampling noise. That automated count is the yardstick. Moving it is our team's work.
What 240 runs looks like.
Early runs swing a lot. By run 240 the estimate has settled and the band around it is tight enough to act on. Without that band you can’t tell a real gap from noise.
“36% ±6” in plain English.
Read it like this: if buyers asked this question 100 times this month, our best estimate is that this firm appears in 36 of those answers, and we’d be surprised if the true number sat outside 30 to 42.
So when your share reads 0% and a competitor reads 36% ±6, that gap is real and not one bad day. When a re-measurement climbs from 0% to 18% ±5, that movement is real too. Those are the numbers the guarantee gets checked against.
What we will not do
- Sell you a screenshot, in either direction.
- Claim we built an AI. Mitra is a measurement system that reads the AI products your buyers use, not a model we trained.
- Show sample results without an “illustrative” label until real case files clear their 90-day windows.
- Move the yardstick between measurements.
The audit runs on this exact protocol.
The free audit uses the same protocol: your buyer questions, run repeatedly until the sample reaches 240 answers. It costs nothing, and the report is yours either way.