Retrieval-on vs retrieval-off answers, defined

Every AI answer comes from one of two retrieval states.

• Retrieval-off (also called search off, ungrounded or parametric): the model answers from what it learned in training. It reads no web pages for this answer. Its knowledge stops at a cutoff date.
• Retrieval-on (search on, grounded): the model runs one or more searches, pulls pages into its context, and writes from them. It may cite some of those pages.

This entry stays short. For the full comparison, read why AI answers differ with search on and off. For the underlying mechanism, see what retrieval means.

Retrieval is discretionary and set per query on every major platform, as of October 2026.

• ChatGPT decides at the model's discretion whether to search. When it does, OpenAI says it rewrites your prompt into one or more targeted queries.
• Claude.ai treats web search as a user toggle, and Anthropic publishes no default.
• Perplexity's Agent API requires developers to opt web search in. Its Sonar endpoint retired on 2026-09-27, and code that was not migrated can return ungrounded answers without any error. See what the Sonar retirement meant for tracking.
• Google AI Overviews often do not trigger at all for a query.

The model behind an answer also varies. OpenAI's model documentation lists a Feb 16, 2026 knowledge cutoff for the GPT-5.6 family and Aug 31, 2025 for the Instant path ChatGPT uses. So a retrieval-off ChatGPT answer can reflect two different snapshots of the world.

The two states respond to different things.

A retrieval-on answer depends on access. If your robots.txt, firewall or JavaScript rendering blocks a search crawler, your pages cannot enter the context. Those are fixes you control.

A retrieval-off answer depends on training data you cannot change this quarter. Anything you publish after a model's cutoff is absent from it.

This is why a blended score misleads. Published estimates of how often ChatGPT searches range from roughly 18% of conversations (Profound) to 80.5% of a 100-prompt set (cloro), because each measures a different population. A tool that never records retrieval state is averaging an unknown mix. The next move: test each question with search forced on and forced off, using a method like testing AI answers with search on and off.

• "No sources means the AI made it up." Not necessarily. It usually means the answer was retrieval-off. Read why some ChatGPT answers have no sources.
• "Allowing GPTBot fixes retrieval-off answers." GPTBot governs training use only. Any effect on future models is unmeasured by anyone, and it would arrive a model generation later.
• "A cited answer shows everything the model used." OpenAI's API returns retrieved sources separately from citations, and the retrieved set is often larger. You can log retrieved vs cited sources to see the gap.
• "You can always tell from the answer." Visible citations are the main signal. Without them, you do not know which state you are looking at.

• Retrieval: how engines pull pages into context at answer time.
• Knowledge cutoff: the date a model's training data ends, which bounds every retrieval-off answer.
• Grounding: another name for writing an answer from retrieved sources.
• Citation: a source the answer credits, which is a subset of what was retrieved.

How can I tell whether an AI answer used web search? Look for citations or source links. Their presence strongly suggests retrieval was on. Their absence does not prove it was off, because some interfaces hide sources and some answers mix both. Through an API you can control the setting directly and log the sources returned. For the signals by platform, see [how to tell if an answer used live web search](/questions/tell-if-ai-answer-used-web-search).

Can I change what a model says in retrieval-off answers? Not this quarter. Retrieval-off answers come from training data frozen at a cutoff. Nothing a publisher changes today reaches them. What you can control is whether your pages are reachable and readable when retrieval is on: crawler permissions in robots.txt, firewall or CDN rules, and having key facts in the HTML before JavaScript runs.

Is a retrieval-off answer less accurate? It can be out of date, because it cannot see anything published after the model's cutoff. That makes recent prices, products and changes the most likely to be wrong. Retrieval-on answers can still be wrong too, since they depend on which pages were fetched and how the model used them. Test both states rather than assuming either one is correct.

Should AI visibility tools test with search on or off? Ideally both, recorded separately. Each state tells you something different: search off shows what the model already believes about you, search on shows whether your pages make it into live answers. A single number that mixes them cannot tell you which problem to fix. The key requirement is that the tool discloses which state each answer came from.