Search On vs Search Off: Why AI Answers About Your Brand Differ

Ask ChatGPT what your company charges, and you might get two different answers in the same week. That's not random noise. It's two different systems answering.

In a search-off answer, the model replies from parametric knowledge: what it absorbed during training, frozen at a cutoff date. In a search-on answer, the engine runs web queries, pulls pages into its context and writes from them. The glossary entry on retrieval-on vs retrieval-off answers covers the terms in more detail.

The two answers can disagree about your price, your products, your headcount or whether you still exist. For brand accuracy, the first job is knowing which one you're looking at.

No major engine searches on every query. As of October 2026, the documented behavior looks like this:

• ChatGPT decides at the model's discretion whether to search. OpenAI's help page says ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to search providers. If memory is on, saved memories can shape that rewrite.
• Claude on claude.ai searches only when the user's web search toggle is on, and Anthropic publishes no default. On the API, the integrator's setup and system prompt steer it.
• Perplexity's Agent API requires developers to opt web search in. Perplexity retired its Sonar Chat Completions API on 27 September 2026, and code that was not migrated can return ungrounded answers without any error.
• Google AI Overviews often don't trigger at all. Google says AI Overviews and AI Mode "may use different models and techniques", so they are separate surfaces even when they share a page.

Nobody outside OpenAI knows how often ChatGPT actually searches. Published estimates disagree sharply: Profound reports about 18% of conversations trigger a web search, Semrush's US clickstream data put it at 34.5% of queries in February 2026, and cloro measured 80.5% globally on a 100-prompt set it chose. They measure different things on different populations. Treat the range as proof that the mix is unknown, not as a number to plan around.

A search-off answer can only contain what was in the training data before the cutoff. OpenAI's model documentation gives the GPT-5.6 family a knowledge cutoff of 16 February 2026. The Instant path behind the chat-latest alias carries an older cutoff of 31 August 2025.

That means "ChatGPT's knowledge" is not one thing. Routing between model paths changes the cutoff by months. A price change, rebrand, acquisition or product launch after the relevant cutoff simply does not exist in a search-off answer.

It also means old material can stick. If an outdated article was in the training data, the model may repeat it long after you've corrected it on your site. Our explainer on whether old news can shape what AI says about your brand goes deeper.

To see which of your recent changes fall after each cutoff, the free knowledge cutoff check compares your news dates against model cutoffs.

Citations are the main signal. When an engine searches, it usually attaches links to the sources it used. When it doesn't, the answer arrives with no sources.

Two caveats keep this honest. First, a missing citation tells you the answer is probably search-off, but you can't be certain from outside. Second, cited is not the same as used. OpenAI's own docs say the number of sources the model consulted is often greater than the number it cites. Engines also cite differently: Claude cites short text spans, while Perplexity cites whole sources by number.

For a step-by-step method, see how to tell if an AI answer used live web search.

Most brand errors fall into one of two buckets, and each bucket has a different fix.

• Search-off errors come from training data. The model states a price from two years ago or a product you discontinued. You can't edit a trained model. You can only make sure the search-on path has correct material, and wait for a future model with a later cutoff.
• Search-on errors come from what the engine retrieved. Either it couldn't reach your page, couldn't read the fact on it, or found a third-party page that says something different. These are fixable because the cause sits on pages someone controls.

If you blend the two, you'll misdiagnose. A team that sees a wrong price and rewrites its pricing page has wasted the effort if the wrong answer was parametric. A team that blames the model when its firewall is blocking the search crawler has missed the fix entirely.

Search does not guarantee accuracy either. A search-on answer is only as good as the pages retrieved. Our piece on whether web search makes AI answers more accurate covers where grounded answers still go wrong.

Be clear-eyed about the split.

Out of your control

• Whether an engine searches for a given question.
• The model's knowledge cutoff and what was in its training data.
• The queries the engine actually runs after rewriting the user's prompt.
• Which retrieved pages it chooses to cite.

In your control

• Crawler permission per bot. robots.txt rules are independent for each OpenAI crawler. OAI-SearchBot controls ChatGPT search, while GPTBot controls training use only. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
• Network reachability. OpenAI's help page says eligibility requires allowing OAI-SearchBot and confirming your host or CDN allows traffic from its published IP addresses. A firewall can block a crawler that robots.txt allows.
• Rendering. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Facts that only appear after JavaScript runs may be invisible to them.
• Consistent, specific facts. A price or definition stated plainly in fetchable HTML gives the engine a sentence it can quote.

A minimal robots.txt that allows ChatGPT search while opting out of training looks like this:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

None of these guarantees a citation. They remove the blocks that make a correct search-on answer impossible.

1. List the facts AI must get right. Price tiers, product names, founding year, location, ownership. Check that your own site states each one the same way everywhere.
2. Check access for each search crawler. Confirm robots.txt lets OAI-SearchBot, PerplexityBot and Claude-SearchBot in. Then confirm your firewall returns a 200 to them, not a 403. The AI Crawler Checker requests a page as each crawler and shows which ones get turned away.
3. Put the key facts in server-rendered HTML. If your pricing loads through client-side JavaScript, most AI crawlers won't see it.
4. Test the same questions with search on and off. Run them logged out, with memory off, and record whether sources appear. The playbook on testing AI answers with search on and off gives a repeatable method. The free checker on this site gives a quick snapshot for three buyer questions.
5. Sort each error by bucket. Search-off errors go on a watch list. Search-on errors go to the fix list, starting with access and rendering, then the third-party pages the engine cited.
6. Handle facts that changed after the cutoff. Follow the steps for correcting outdated facts past a knowledge cutoff, and recheck when engines ship new models.

• Reading one answer as the truth. A single response is one draw from a mix of search-on and search-off behavior. Repeat the test and note retrieval state each time.
• Trusting a blended score. Any single AI visibility score, ours included, averages answers given with and without live retrieval. Read platforms and retrieval states separately.
• Allowing GPTBot to get into ChatGPT search. GPTBot governs training only. Search uses OAI-SearchBot, and OpenAI says each setting is independent.
• Testing logged in. Memory and custom instructions can change how ChatGPT rewrites a query, so your results won't reproduce for anyone else.
• Treating the Gemini API as AI Overviews. A grounded Gemini API call is a different path from AI Overviews or AI Mode. Results from one don't stand in for the others.
• Rewriting content to fix a parametric error. If the wrong answer had no sources and the fact predates the cutoff, editing your page won't change it. Fix the search-on side and wait for the next model.

Does turning on search make ChatGPT's answer about my brand correct? Not automatically. A search-on answer reflects the pages the engine retrieved and chose to use. If those pages are outdated, contradictory or come from a third party with wrong facts, the answer can still be wrong. Search makes a correct answer possible when your accurate pages are reachable and readable. It doesn't guarantee the engine picks them.

If I block GPTBot, will ChatGPT search stop showing my site? No. OpenAI documents GPTBot as its training crawler and OAI-SearchBot as the crawler for ChatGPT search, and says each setting is independent. You can disallow GPTBot to opt out of training while allowing OAI-SearchBot so your pages remain eligible for search answers. Blocking OAI-SearchBot is what removes you from ChatGPT search answers.

Why does the same prompt give different answers on different days? Several things vary between runs. The engine may search one time and not the next. When it searches, it rewrites your prompt into its own queries, and the results change as the web changes. ChatGPT can also route to model paths with different knowledge cutoffs. Memory and custom instructions add more variation if you test while logged in.

When will my recent news show up in search-off answers? Only after an engine ships a model whose training data includes it, and vendors don't announce what any given model learned. Until then, recent news can only reach users through search-on answers. That's why making the news readable on pages search crawlers can reach matters more than waiting for the next model.

Should AI visibility tools test with search on or off? Ideally both, reported separately, because users see both kinds of answer and each needs a different fix. A tool that only tests one state misses half the picture. A tool that blends them hides which bucket an error belongs to. Our explainer on [testing with search on or off](/questions/ai-visibility-testing-search-on-off) covers the tradeoffs.