Grounded vs Ungrounded: The One Question to Ask Any AI Visibility Tool

If you use a tool to track whether AI recommends your brand, there's a question worth asking it: when it queries ChatGPT, does ChatGPT search the web first?

It sounds like a technical detail. It isn't. It decides whether the number you're looking at can move at all when you improve your site.

Two very different questions

When you send a prompt to a large language model through its API, you can do it two ways.

Ungrounded means the model answers purely from what it absorbed during training. No searching, no fetching. The answer reflects a snapshot of the web frozen at the model's training cutoff — often many months in the past.

Grounded means the model runs a real search before answering, reads what it finds, and builds its answer from those pages. This is what the consumer products actually do. When someone types "best CRM for a small law firm" into ChatGPT today, they get a retrieval-augmented answer built from pages fetched at that moment.

So an ungrounded API call and a real user's ChatGPT session are answering two different questions:

• Ungrounded: "What does this model remember about my industry from its training data?"
• Grounded: "What would a real person see if they asked this right now?"

Why this decides whether your work counts

Here's the uncomfortable consequence. A model's training data is fixed. Nothing you publish today changes what a model already memorized.

So if a visibility tool measures ungrounded responses, the score it shows you is a measurement of a frozen artefact. You can rewrite your homepage, add schema, publish ten new pages, earn a mention in an industry roundup — and the number will not respond, because the thing being measured cannot respond.

A grounded score behaves the opposite way. The engine re-searches every time. New pages become eligible immediately. A mention on a site the engine likes to cite can show up in days. The number becomes a feedback loop instead of a readout.

How to tell what your tool is doing

Vendors rarely state this outright, so here are three tests you can run yourself.

A grounded answer knows where its information came from and can cite it. If your tool shows engine responses with no source links for any engine except Perplexity, that's a strong signal the rest are ungrounded — Perplexity browses by default, so it's often the only one that looks "connected".

Track a question about a recent event, a newly launched competitor, or a product released in the last few months. An ungrounded engine will either not know or will answer confidently with stale information.

There's a real cost difference. Grounded calls are slower and more expensive — search adds latency and per-call fees on top of tokens. A tool running ungrounded calls is cheaper to operate, and that saving is exactly why it happens.

What "no simulations" should actually mean

Plenty of tools promise "real API calls, not simulations". That claim is usually true and still misses the point. An ungrounded API call is a real API call. It's just a real call to something that doesn't behave like the product your customers use.

The honest bar is higher: real calls, in the mode real users experience, with the sources the engine actually drew on.

What we changed

AEO Track now runs every engine search-grounded — ChatGPT, Gemini, Claude, Perplexity and Grok. Each check uses the provider's own web-search capability, and we store the source URLs each engine returned alongside the answer.

Two things fell out of that change that are worth knowing about.

First, grounded answers are longer. A search-backed "best options for X" response often runs well past a thousand tokens once each item gets a sentence of context. We raised our response limit accordingly, because a truncated answer that cuts off before your brand appears is a false negative — and it's biased against exactly the mid-ranked brands most likely to be paying attention to their visibility.

Second, we now record whether each stored result was grounded. That sounds like bookkeeping, but it's what lets a score be audited later: you can tell how any given data point was produced.

The practical takeaway

If you're evaluating AI visibility tools, grounding is the first question, not a footnote. A tool that measures ungrounded model memory can tell you something interesting about how a model was trained. It cannot tell you whether your work is paying off — and it will quietly imply that it can.

Frequently asked questions

What does 'search-grounded' mean for AI visibility tracking?

It means the AI engine runs a real web search before answering, then builds its answer from the pages it found. This matches what consumer AI products actually do. An ungrounded call answers only from training data, so it reflects a frozen snapshot of the web and cannot respond to anything you publish today.

Why can't ungrounded AI visibility scores improve?

A model's training data is fixed at its cutoff date. Publishing new content, adding schema or earning citations does not change what a model already memorized, so an ungrounded score cannot react to your work. It still drifts with model updates and randomness, which makes it look responsive when it is not.

How can I tell if my AI visibility tool uses grounded queries?

Check whether results include real source URLs for every engine rather than just Perplexity, track a question that only the live web could answer correctly, and ask the vendor directly which model and mode they call. Grounded calls cost more and are slower, which is usually why tools skip them.

Which AI engines does AEO Track query with web search enabled?

All five: ChatGPT, Gemini, Claude, Perplexity and Grok. Each check uses the provider's own web-search capability and stores the source URLs the engine returned alongside the answer.