You opened ChatGPT, asked the exact question your dashboard asks, and your hotel wasn't in the answer — even though your Visibility Index says 72. Both things are true at the same time. Understanding why is the difference between reacting to noise and reading real signal.
The root cause: AI doesn't give the same answer twice
Large language models are probabilistic. To sound natural and varied, they sample from a range of possible next words rather than always picking the single most likely one. The practical consequence: ask "best boutique hotel in Lisbon" ten times and you can get ten slightly different lists. Sometimes you're first. Sometimes you're third. Sometimes you're not there at all. Nothing about your hotel changed between 10:00 and 10:01 — the model simply rolled the dice again.
This is not a defect to be engineered away. It's the nature of the thing being measured. Any honest approach to AI visibility has to start by accepting that the signal is inherently noisy — and then deal with it.
When you test a prompt yourself, you're taking a single sample of a noisy system — one die roll. Your dashboard score is built from thousands of samples across every engine and persona, every day. A single miss in your own test is completely consistent with a healthy score, the same way one coin landing tails tells you nothing about a coin that's 72% heads. The systematic measurement exists precisely because spot-checking can't be trusted.
So why does my score move between scans when I changed nothing?
Three forces are at work, and it helps to name them:
1. Sampling noise
Even averaged over thousands of answers, the day-to-day result will drift a little simply because each scan is a fresh set of rolls. This is the normal "breathing" of the signal — usually a few points either way — and on its own it means nothing.
2. The providers update their models
OpenAI, Google, Anthropic, and the rest ship model updates constantly, often without announcement. A new version can shift how it phrases recommendations and which properties it favors — for everyone at once. A genuine drop that appears on one engine and only that engine, on a particular day, is very often a model update rather than anything you did. We watch for this pattern (we call it model drift) precisely because it's systematic, not personal.
3. The world the AI reads changed
Assistants that browse the web are reading a moving target: a new review, a competitor's fresh landing page, an updated OTA listing. As the source material shifts, so does the answer.
The core idea: separate signal from noise
If the signal is noisy, the job isn't to pretend it's smooth — it's to know how big "normal" is, so you can spot when something real happens. Sigtrip estimates a noise band for your specific hotel: the size of the ordinary day-to-day wobble, learned from your own history during periods when nothing was changing. We show that band right on your trend line, and we only call a move "real" when it breaks out of it.
A 3-point dip inside the band is not a story. A 15-point drop that clears the band is. The band is what lets you ignore the jitter and pay attention to the moves that warrant it.
- Moves bigger than your noise band — that's the whole purpose of the band.
- Sustained moves — a trend over several days beats any single-day spike.
- Moves concentrated on one engine — usually a provider model update.
- Moves concentrated on one persona — usually a genuine coverage gap worth closing.
- Anything right after you edited prompts, personas, or facts — that's your change, not noise. (Edits apply on the next daily scan, so give it a day before you read the result.)
Why we don't just smooth it away
It would be easy to run a long rolling average and hand you a reassuringly straight line. We don't, on purpose. Heavy smoothing hides exactly the moves you most need to see — like a model update that drops you ten points overnight, or a fix that lifts you the day after you ship it. We'd rather show you the real point and be honest about the uncertainty around it than present a smooth number that's quietly wrong. The noise band labels the uncertainty instead of burying it.
Want to see the noise for yourself?
Ask the same question five or ten times in a single assistant — then try it in another. You'll watch the answer shift in front of you. That spread is exactly what your score integrates and averages out. Once you've seen it directly, the daily wobble on your dashboard stops being mysterious and starts being something you can read.
For the full picture of how those samples become a score in the first place, see how AI visibility is measured. And when you're ready to move the number rather than just read it, that's how to improve your AI visibility.
Don't chase the jitter. Watch the band-breakers, the sustained trends, and the gaps — and act on those.