Grounding is when an AI assistant searches the web before answering and builds its response from what it just read, rather than from what it memorized during training. It’s the biggest quiet change in how AI talks about businesses: an ungrounded model answers from a frozen snapshot that barely registers most local businesses, while a grounded one answers from your live profiles, reviews, directories, and website — sources that exist right now and can be fixed right now. If an assistant describes your business wrong today, grounding is the reason that’s fixable — and the reason it’s measurable.
Here’s what grounding actually is, why the assistants your customers use adopted it, and what it changed for a small business.
Two places an answer can come from
An AI model is trained once, on an enormous snapshot of text. Whatever stuck from that training is the model’s long-term memory — researchers call it parametric memory, because it’s baked into the model’s parameters. It’s remarkable for general knowledge and nearly useless for your business: training snapshots skew toward the most-written-about things on the internet, and a local tint shop or landscaping company isn’t written about enough to leave a durable imprint. Even if it were, the snapshot ages from the moment it’s taken — a model trained last year holds, at best, last year’s version of your hours.
The second place is retrieval. A grounded assistant runs web searches on the user’s behalf, reads what comes back, and composes its answer from those sources — fetched seconds before the reply appears. The answer is grounded in documents the assistant can actually point to, which is why grounded answers come with citations attached.
Same model, two entirely different answers about your business, depending on which mode it’s in. That distinction — memory versus lookup — explains almost everything confusing about AI and local businesses. Why doesn’t ChatGPT know my business? is that story in full.
Why assistants started looking things up
Grounding exists because parametric memory fails exactly where users need reliability: facts that are local, specific, or newer than the training snapshot. “Who does ceramic tint near me, open Saturdays?” is all three at once. A model answering that from memory can only do two things — admit it doesn’t know, or guess. Both are bad products, and the second one is dangerous: a confidently invented answer looks identical to a real one.
So the major assistants wired search in. ChatGPT, Google Gemini, Perplexity, Claude, Meta AI — on questions like these, they search the web, read a handful of results, and answer from what they found, citing as they go. For questions about local businesses on the assistants your customers actually use, grounded answering is now the norm, not the exception.
The result is worth stating plainly: when a customer asks an assistant about businesses like yours, the answer is not a mystical AI opinion. It’s a fast, automated read of a specific set of web sources — a set you can see in the citations, and largely influence.
The web-off test
You can watch the difference yourself in about two minutes. If an assistant lets you turn web search off, ask it about your business with search disabled. You’ll almost certainly get one of three things: an honest “I don’t have information about that,” vagueness, or a confident guess with details subtly wrong. That’s parametric memory, unassisted — what AI “remembers” about your business without looking is usually nothing.
Then ask again with search on. Suddenly there are specifics, and citations under them. Everything useful in the second answer arrived in the seconds before it was written — retrieved from your profiles, your reviews, your listings, your site.
We run a version of this test deliberately in our own measurement, keeping one web-off run alongside the grounded ones, because it isolates the point: the value isn’t stored in the AI. It’s stored in the sources — so the sources have to be right. The full DIY walkthrough includes this test.
What grounding changed for a small business
Before grounding, if a model held a wrong or missing impression of your business, there was nothing to be done — the snapshot was taken, training was over, and nothing you published would reach that model until some future retraining you didn’t control and couldn’t schedule.
Grounding inverted that. Your visibility in AI answers stopped being frozen history and became a live property of sources that are editable today:
- The inputs are ordinary. The sources grounded assistants read are the same profiles, listings, reviews, and pages local businesses have always maintained. No new platform to master — a new reader for the old ones.
- Fixes can actually land. Correct the stale listing, resolve the duplicate profile, publish the page that answers the question customers keep asking — and those changes are in the pile of material the next grounded answer is built from. Not instantly and not guaranteed, but reachable, which the training snapshot never was.
- The failure modes moved somewhere inspectable. A wrong AI answer used to be a shrug — “the model just says that.” Now a wrong grounded answer usually traces, through its citations, to a specific source holding a specific bad fact. That’s a to-do list, not a mystery.
One honest caveat: grounding makes answers traceable, not automatically true. An assistant that retrieves a stale directory will faithfully repeat the stale directory — it can only compare sources against each other, not against reality. Grounding moved the problem from “inside the model, unfixable” to “in your sources, fixable.” It’s still your problem. It’s just finally an addressable one.
Citations are receipts
The last change is the one we build on: grounding made AI visibility measurable. An ungrounded answer is unfalsifiable — there’s no way to know why the model said what it said. A grounded answer shows its work. The citations attached to it are receipts: which sources were read, whose material earned the recommendation, where the wrong fact came from.
Receipts make honest measurement possible — asking real customer questions repeatedly, recording what assistants answer and cite, checking each claim against verified facts, and re-testing after fixes to show what actually moved. That’s the discipline behind the three numbers we track — Discovery Rate, Answer Accuracy, Citation Share — defined in What is AI visibility?, with worked examples in our case studies.
Grounding didn’t make AI answers predictable — assistants still vary their responses, and nobody can promise you a mention. What it made them is inspectable: built from sources you can see, largely fix, and verifiably improve. For a small business, that’s the whole story in one sentence — your AI visibility is no longer something that happened to you in a training run. It’s something you can work on, starting with finding out where you stand.