Grounding is when an AI model bases its answer on retrieved, verifiable sources (live web results or documents) instead of relying only on what it learned in training.
An ungrounded answer comes from the model's parameters: a compressed memory of its training data, frozen at its knowledge cutoff. A grounded answer folds in fresh retrieval (web search, a document, a database) before responding.
Grounding is why AI rankings move. A model's trained opinion of a category changes only when the model is retrained, but its grounded answers shift as the underlying sources change. That is also why recommendation answers differ between the same model with and without web access.
Why it matters now
Grounding is why AI rankings move at all. A model's trained opinion of a category changes only when the model is retrained, which is rare. Its grounded answers shift whenever the underlying sources change, which is constant. For a brand, that means the recommendation is not frozen: today's answer can differ from the last one because the model read something new.
It matters now because grounded, retrieval-backed answers dominate buying questions. Buyers ask for current recommendations, and the engines that fold in fresh sources give better ones. That is the layer where a brand's recent work actually shows up.
How WhereDoIRank measures it
We ask the models the category's buying questions with retrieval in play, then record and score the answers at each close. Because grounded answers change as their sources change, the close-over-close movement on the index is largely a picture of the grounding layer: what the models read at this close versus the last.
Comparing the four models sharpens the read. When ChatGPT, Claude, Gemini, and Perplexity diverge on the same brand, the difference usually traces to what each grounded on, which is exactly the layer a brand can influence.
A worked example
Ask the same model the same category question twice, once with web access off and once on. Off, it answers from training memory and may favor the brands it "knew" at its cutoff. On, it can retrieve current sources and surface a brand that launched or improved since then. The two answers can name different leaders.
That contrast is grounding at work. For a challenger, it is also the opening: the ungrounded answer is hard to change, but the grounded one responds to sources you can influence this quarter.
Common misconceptions
Grounding does not mean an answer is correct. Retrieving a source reduces guesswork, but a model can still ground on a wrong or outdated page, so grounded answers hallucinate less, not never. Grounding is also not the same as training: it does not teach the model anything durable, it just informs one answer. And it is not automatic on every query; when a model answers from memory alone, there is no grounding to influence, which is why the same question can behave so differently with browsing on versus off.
Grounding is also not something a brand controls directly. You cannot make a model retrieve a page; you can only make the pages it does retrieve represent you accurately. The influence is indirect, exercised through the sources, which is exactly why reading the citations behind an answer is the practical place to start.
How it connects
Grounding is the mechanism that retrieval-augmented generation (RAG) implements, and citations are its visible output: the sources an answer grounded on. It stands opposite the knowledge cutoff, the frozen training memory a model falls back on when it does not retrieve, and it is a check on hallucination, since a grounded, cited answer is easier to verify.