To find out whether ChatGPT recommends your business, ask it the way a scientist would: the same real buyer questions, on a schedule, in a clean session, with every answer logged — because a single check of a probabilistic system proves nothing. This is the full protocol. It costs nothing but discipline.
The casual version is a founder ritual: open ChatGPT, type "best [your category] for [your market]," hold your breath. The answer matters — buyers increasingly start research there — but one answer is one draw from a probability distribution. Ask again in an hour and the citations can change.
Why one check lies to you
- Answers are probabilistic. Same question, same engine, same day → different citations. Sampling variation is built in.
- Answers are personalized. Chat history, custom instructions, and memory shape what you see. The answer you get logged-in isn't what your buyer gets.
- Engines change silently. Retrieval sources have shifted drastically within weeks; last month's result may say nothing about today.
The fix for all three: repeated, controlled sampling. Not "did we appear?" but "in what fraction of scans did we appear?"
Step 1: Choose your questions
Don't test keywords; test questions — full sentences a buyer would type. Pick 5–10:
- 2–3 category questions: "best [category] for [your specific segment]"
- 2–3 problem questions: the pain as a buyer phrases it before knowing the category
- 1–2 comparison questions: "[you] vs [competitor]", "alternatives to [market leader]"
- 1 reputation question: "what is [your company] and is it any good"
Write them down exactly and never edit them mid-experiment. Changed wording is a changed experiment.
Step 2: Control your conditions
- Use a fresh, dedicated account — memory off (ChatGPT: Settings → Personalization → Memory), temporary chats. Pick one setup and stick with it.
- Confirm web search fired. ChatGPT runs live search on only about a third of queries; if the answer has no citations, it came from training data — log that as its own outcome, not a miss.
- Test at least two engines. You can be visible in one and absent in the other. Perplexity shows its sources most plainly.
Step 3: Log like you mean it
One spreadsheet, one row per question per engine per scan:
| Date | Engine | Question | We appear? | Position | Competitors named | URLs cited | Notes |
|---|
The columns that matter long-term are We appear? (your citation rate) and Competitors named (theirs). The URLs cited column is the most useful diagnostically: it shows which pages are winning the answer — often a roundup you could be in, or a competitor page worth studying.
Step 4: Sample on a schedule
- Baseline: every question, both engines, daily for 7 days — 70–140 data points, enough to see a pattern.
- Ongoing: 2–3 scans a week catches trends and flips.
Week one gives you something like: "Cited in 20% of scans for question 3, 0% elsewhere; competitor X cited in 80% for question 1." Still noisy at 7 samples (treat small gaps as tentative), but a different universe from a screenshot.
Step 5: Read the results
- Never cited, don't rank on Google either → retrieval problem. Build the page. (The five failure modes walk through diagnosis.)
- Never cited, but you rank → the page isn't citable — no statistics, sources, or quotable specifics. Fixes: How to get cited by ChatGPT.
- Cited 20–60% of the time → you're on the shortlist; this middle zone is where content changes move the rate most.
- Citations are third-party roundups that omit you → the battle is inclusion in those sources, not your own site.
- No citations at all → training-data answer; log it separately and focus where you have influence.
Step 6: Re-scan after every change
Illustrative, not measured data: weekly citation rate for one buyer question, before and after a citability rewrite. Expect the flat weeks — recrawl lag is real. Judge the trend, never a single scan.
Baseline → change a page → keep scanning → watch the rate. Give it a few weeks (recrawl timing is undocumented; don't judge under two). If the rate doesn't move, the change didn't work; try the next tactic. Citation rate is your instrument.
The honest cost accounting
10 questions × 2 engines × daily ≈ 30–40 minutes a day of asking, reading, and logging — indefinitely. Most teams last two weeks before the spreadsheet dies. That failure mode is why our product exists: we run this exact loop daily and draft the pages that fix what the scans find. But run the DIY version for a week first — you'll learn more about your market from 7 days of honest scans than from any sales pitch, including ours.
Read next: Why ChatGPT recommends your competitors · How to get cited by ChatGPT