Definition
Grounding rate is the share of AI answers, for a given set of prompts on a given engine, that were built from live retrieved sources rather than from the model's parametric knowledge alone. A reading looks like 'ChatGPT searched the web on 34% of our tracked prompts last month, Perplexity on nearly all of them'. It is a property of the prompt and the engine, not of your brand, but it determines which optimization levers can work at all.
The split matters because the two kinds of answer respond to different work. When an answer is grounded, retrieval ran, so a page that is crawlable, ranks for the model's grounding queries and contains a clear passage can enter the answer within days. When an answer is ungrounded, the model is drawing on training data, and no on-site change will alter it until a future model is trained. Influencing those answers is a slower game of broad, consistent presence across the web, the substance behind brand depth and share of model.
The engine decides whether to retrieve. Google's Gemini API documentation says the model 'determines if a Google Search can improve the answer' before running any query, and OpenAI's tools guide says the model 'automatically decides whether to use a configured tool'. Published measurements of ChatGPT show how selective that decision is, and how much the number depends on what is counted:
- Profound, an AI visibility vendor, analyzed about 700,000 US ChatGPT conversations from October to December 2025 and reported in February 2026 that about 18% of conversations trigger at least one web search. Opening turns searched far more than later ones: 12.6% of first turns carried citations, against 3.0% of twentieth turns.
- Nectiv tested more than 8,500 prompts across nine industries and found that 31% triggered a search. Local-intent prompts searched most often (59%), credit card and fashion prompts least (18% and 19%) (Search Engine Land).
Retrieval-first products such as Perplexity are designed to search on nearly every question, and Google's AI features draw on Search results by design, so their grounding rates sit far above ChatGPT's. Within any engine, time-sensitive, local, comparative and product questions are grounded more often than definitions and general concepts.
Measuring it is straightforward when the engine exposes its behavior. Citations or a visible search step mark a grounded answer; developer APIs return the executed queries; and information published after the model's knowledge cutoff implies retrieval ran. Results differ between the consumer interface and the API, so measure on the surface your customers use (UI vs API tracking). In Promptwatch, a response with no citations is usually one the model answered from training data, which is expected rather than an error.
Use grounding rate to segment work. Report citation rate only on grounded prompts, since ungrounded prompts can never cite anyone. Put content effort into prompts that ground reliably, and treat persistently ungrounded prompts as a reputation problem: what the model believes about your category comes from the sources that shaped its training data. A related quality check is grounding depth, whether the cited pages actually support the sentences they are attached to; weakly supported answers tend to swap sources more often from run to run.
Examples of Grounding Rate
- A software company finds that ChatGPT searched on 41 of its 120 tracked prompts in September, nearly all of them comparison and pricing prompts, and concentrates its content updates there.
- A definitional prompt ('what is revenue operations') is answered from model memory on every run, so the brand's new explainer page has no effect on the ChatGPT response.
- An analyst recalculates citation rate using only grounded responses and finds the brand's real citation share is three times higher than the blended number suggested.
- A retailer notices that follow-up questions in long ChatGPT conversations almost never trigger search, so it optimizes for the opening questions shoppers ask instead.
