Ask ChatGPT who the best orthodontist in Scottsdale is. Then ask Claude. Then ask Gemini the same thing a week later. You will not get the same answer three times. Sometimes your practice is named. Sometimes a competitor two miles away is named and you aren't mentioned at all. Sometimes the assistant describes you accurately, and sometimes it confidently says something about your hours or your services that stopped being true two years ago. None of that is a bug. It's how these systems work. And it means the question most businesses are actually asking — "am I showing up in AI?" — has a hidden second half nobody warned them about: showing up when.
A check is a snapshot, not a status
Here's what trips people up. You run one test, you see your name, you exhale. Problem solved, you think. Except you didn't solve anything. You took a photograph of a moving thing.
An AI assistant doesn't store a fixed verdict about your business the way a filing cabinet stores a folder. Every time someone asks, the answer is built fresh — pulled together in the moment from whatever sources the model trusts, weighted against whatever it can find about the businesses it's choosing between. That means the answer you saw on Tuesday was true on Tuesday. It carries no promise about next month. The industry has a word for this quiet movement: drift. And drift is the whole reason a single check tells you almost nothing durable.
Consider a plumbing company in Tampa. In March, a customer asks Gemini for an emergency plumber near them and that company comes up first, described warmly, pulled from a run of recent five-star reviews. The owner checks once, sees it, feels good, moves on to actually running the business. By September, the answer has quietly rearranged itself. A competitor across town spent the summer gathering reviews and tightening up how they describe their services online. The model now finds more recent, more consistent, more agreeing information about that competitor. So it reaches for their name instead. Nothing broke on the Tampa company's end. No alert fired. The owner only finds out when a longtime customer mentions, almost in passing, that they "looked it up and some other guy came up."
That's the problem with a snapshot. It can't show you motion. And the motion is the whole story.
Why the answer moves even when you don't
To understand drift, you have to understand what the model is actually doing when it decides whose name to say.
AI systems don't take your website's word for who you are. A business can claim anything about itself, and the model knows it. So instead of trusting the claim, these systems look for corroboration — information that appears the same way across more than one independent source. Agreement is the signal. When a model finds the same facts about you repeated consistently across places it already trusts, it grows confident, and confidence is what lets it recommend you by name. When it finds contradictions — your website says one thing, a directory says another, a third source is years out of date — it hedges. And a hedging model tends to recommend someone it's more sure about. Someone whose story lines up cleanly.
Now notice what that means over time. You are not the only one in the picture. Every competitor in your category is also a source of information flowing into the same system, and that information is changing constantly. New reviews land. Listings get updated or go stale. Someone rewrites their service description to be clearer about exactly what they do. The model gets retrained on a fresher slice of the web. Each of those shifts changes the balance of who looks most corroborated right now — which is why your answer can move while you personally changed nothing. The water rose. You just didn't see it because you were looking at a photo.
Take Healthgrades, for a medical example worth sitting with. A dermatologist in Columbus might treat it as a set-and-forget profile — filled it out years ago, never thought about it again. But a profile like that isn't a static business card to an AI model. It's one of the independent voices the model cross-checks a doctor against. If the hours there, the listed specialties, the practice name, and the location all match what shows up on the practice's own site and elsewhere, the model reads that agreement as a strong signal this doctor is exactly who they claim to be, and it names them with confidence. If that profile still lists a since-closed second office, or a specialty the practice dropped, it quietly becomes a source of contradiction instead — a reason for the model to hesitate. The dermatologist didn't do anything wrong. They just stopped watching a source that kept speaking on their behalf after they walked away from it. That's drift wearing a specific face.
The miss is near-total, not a lower rank
People carry an old mental model over from search, and it's costing them. On a traditional results page, slipping from third to seventh still left you on the page. Someone scrolling could still find you. There was a long tail of visibility.
An AI answer doesn't work like that. When a customer asks an assistant to recommend a roofer or an estate attorney or a pediatric dentist, what comes back is short and named — a handful of businesses spoken as a recommendation, not a page of ten blue links to browse. Some of those answers now carry inline ads alongside the organic names, and the exact shape shifts by question, by session, by which assistant and which week. But the hard truth underneath the variation holds: being left out of that short answer isn't a lower ranking. It's closer to absence. The customer never scrolls to find you, because there is nothing to scroll. They get a name, they act on it, and if the name wasn't yours, you were never in the room where the decision happened.
This is why drift bites harder in the AI era than it ever did in search. The gap between "named" and "not named" is a cliff, not a slope. And that edge moves, month to month, with the information tides — which means the distance between you and invisibility is being quietly renegotiated while you're busy doing the actual work your customers hired you for.
What a real scan actually looks at
This is where a one-time curiosity check falls apart, and where watching it properly starts. RankCommander's scan doesn't ask one assistant one question one time. It asks the questions your actual customers ask — the real-world phrasings a patient or client or homeowner types when they're looking for exactly what you do — across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot. It reads whether you're named, whether a competitor is named instead, and whether what the assistant says about you is actually true. Then it does it again. That repetition is the entire point, because a signal you capture once is a snapshot, and a signal you capture on a running basis is the thing that tells you the ground shifted in time to do something about it.
Underneath that sits the RankCommander AI Visibility Index — tens of thousands of real AI answers evaluated across local-business verticals, and still counting. That's first-party data on how assistants actually behave for businesses like yours, not borrowed survey numbers or a competitor's estimate. You can see the vertical benchmarks at /ai-visibility-index. What it shows, over and over, is that drift is the norm and stability is the exception. The businesses that stay named aren't the ones who checked once and relaxed.
The part nobody tells you until it's too late
The cruel thing about drift is how quiet it is. There's no email. No dashboard flashing red. A competitor doesn't call to tell you they're now the name Gemini gives for your category in your city. You find out the way the Tampa plumber found out — sideways, from a customer, after the client you should have gotten already called someone else. By then you're not preventing a loss. You're trying to win back ground someone else already stood on.
You spent years earning the reputation that makes you the right answer. The reviews, the referrals, the work done well enough that people recommend you by name. An AI assistant can route around all of it in a sentence, naming someone who simply looks more corroborated this month than you do. The right moment to catch that is before the customer does, not after. RankCommander watches the thing you can't watch by hand — every assistant, every month, the drift as it happens — so the name those systems say stays yours. Start your scan and find out what they're saying about you right now, before you hear it from someone who already called the other guy.