Ask ChatGPT "who are the best HVAC companies in Phoenix?" once, and you get an answer. Ask it again ten minutes later, and you might get a different one. The list reshuffles. A name that was third drops off entirely. A company you've never heard of shows up. If you've spent any time watching what AI assistants say about a business, you know this already: the output is noisy.
That noise is why a single query tells you almost nothing. And it's why the metric that matters isn't "where did I rank when I appeared" — it's "how often did I appear at all."
Two different questions
There are two things you can measure when you look at AI answers about a market:
- When my business shows up, where does it land in the list?
- How often does my business show up at all, across many runs?
People instinctively reach for the first one. It feels like SEO — rank position, top of the page, everyone understands it. But in AI answers, the first question is a trap, because it only counts the times you appeared. If you show up in 3 out of 25 runs and you're ranked #1 in all three, "average rank: 1st" is technically true and completely misleading. You were absent 88% of the time.
The second question — presence frequency — is the one that survives the noise. It's the share-of-voice number worth reporting.
A worked example
Say you're an agency managing a mid-size plumbing company in Denver. You want to know how it stacks up in AI answers against its two biggest local competitors.
You can't ask once. One run is a coin flip. So you run the same prompt 25 times across the assistants — ChatGPT, Claude, Gemini, Perplexity, and Google AI — and you count appearances. Not rank. Just: did this company get named, yes or no, in each run.
Here's what you might get back:
| Company | Appears in | Frequency | |---|---|---| | Rival A | 18 of 25 | 72% | | Rival B | 14 of 25 | 56% | | Your client | 3 of 25 | 12% |
Now you have something stable. Rival A gets mentioned nearly three times out of four. Your client shows up in about one run out of eight. That gap — 72% versus 12% — is real, it's repeatable, and it doesn't move much if you run the sample again next week. That's the whole point. A single number that bounces around is a distraction. A frequency built from 25 runs is a measurement.
Notice what the rank-based view would have done to this. In those 3 runs where your client appeared, suppose it was listed second each time. "We rank second in Denver AI results" sounds like a win. It's not. It's second place in the 12% of the time you're in the room at all. The frequency number keeps you honest.
Why frequency is the stable one
Individual AI answers vary run to run because the models sample from a distribution. Any single response is one draw. But the rate at which a business appears across many draws converges — the more runs you take, the tighter the number gets. This is the same reason you don't judge a conversion rate off five clicks. Five clicks is noise. Five thousand is a rate you can plan against.
Three runs is not a sample. Twenty-five starts to be one. The frequency you get from a real sample is what turns "the AI said something weird about us today" into "we hold 12% presence in this market, our top rival holds 72%, here's the gap." That second sentence is something you can put in front of a client without hedging.
What this does and doesn't tell you
Presence frequency tells you your competitive position — how visible you are in AI answers relative to the people you compete with, tracked as a percentage over time. It's a scoreboard. Run it monthly and you can see the gap widen or close.
What it doesn't tell you is why, and it doesn't change the answers. Counting appearances is measurement, not optimization. It's the step before anyone tries to move the number — and you can't move a number you haven't measured. Most businesses I talk to have never counted at all. They've asked ChatGPT about themselves a handful of times, gotten a mix of flattering and alarming answers, and formed a gut impression. That gut impression is built on a sample of three.
The discipline here is the same one that works everywhere else in acquisition: define the metric that actually drives the outcome, measure it across a big enough sample that the noise washes out, and report the stable number instead of the anecdote. In paid search you'd never report CPA off a single conversion. Treat AI visibility the same way.
The takeaway
If you're going to tell a client where they stand in AI answers, give them a frequency, not a snapshot. Appearing in 3 of 25 runs versus a rival's 18 is a clear, defensible statement about competitive position. "We ranked second that one time I checked" is not.
Count appearances across a real sample. Separate how often you appear from where you rank when you do. The frequency is the number that holds still long enough to report — and it's usually the number that tells the harder, more useful truth. This is roughly what LLMClarity does under the hood: run the sample, count the appearances across all five assistants, and turn the noise into a percentage you can track month over month.
Start by measuring. Everything else comes after.