How Does ChatGPT Decide When to Use Its Search Tools?

What triggers a ChatGPT search?

Semrush lists similar reasons: the user selects web search, asks for sources, the model is uncertain, or the question involves facts after the training cutoff. Nectiv's data by vertical fits the same pattern. ChatGPT searched for 59% of local prompts and 41% of general commerce prompts, while credit card and fashion prompts were the least likely to trigger a search.

How often do ChatGPT, Claude, Gemini, and Perplexity use web search?

• ChatGPT: about a third of prompts, varying by topic and model. Nectiv analysed more than 8,500 commercial-intent prompts across nine verticals in 2025 and found search in 31% of them. Semrush's clickstream data, covering real users from October 2024 to February 2026, found search enabled on 34.5% of queries by February 2026. The monthly share ranged from 15% to 66.3% as OpenAI changed models.
• Perplexity: every question by default, unless the user turns off the Web source. Perplexity describes itself as searching the internet in real time for each question and citing its sources.
• Claude: decides per prompt when web search is available. Anthropic's documentation for its web search tool says Claude "determines when to search based on the prompt", and can search several times in one answer.
• Gemini: decides per prompt when grounding with Google Search is enabled. Google's developer documentation describes the model deciding whether a query needs a search. In Google Search itself, AI Overviews and AI Mode are built on search results by design.

We know of no published study that runs the same prompts through all four assistants and reports their search rates, and we have not measured it ourselves. Figures from single tests should be treated as snapshots, since rates change with every model update.

Is ChatGPT searching more or less than before?

Both, depending on what you measure. Semrush found the share of queries with search enabled fell every month from November 2025 to February 2026. But when ChatGPT does search, it searches much more. Nectiv re-ran about 4,000 of its prompts in 2026 with a newer model, GPT-5.6 Sol. Average fan-out, the number of searches per prompt, rose from 2.17 to 7.61, and the longest chains grew from 4 searches to 29. Queries also got longer, averaging 6.82 words, and ChatGPT began searching for both the current and previous year's content.

So fewer prompts may trigger search, but a searched answer now draws on far more queries, and far more pages have a chance to be retrieved.

Why does the search decision matter for visibility?

Because the two kinds of answer come from different places:

• Searched answers draw on pages retrieved at the time. Updating a page, earning a new review or being added to a comparison can change them within days or weeks, once the page is indexed.
• Unsearched answers draw on training data. They reflect what was written about you before the model's training cutoff, and change only when a new model is trained. You cannot fix them quickly.

That is why the same brand can look strong in one ChatGPT answer and absent in another. For reporting, separate searched and unsearched answers, as covered in whether ChatGPT visibility reports should separate search, non-search and shopping answers. Where the searches go is covered in whether ChatGPT search uses Google, Bing or its own index.

How can you tell whether ChatGPT searched?

• Sources are shown. Searched answers usually display citations or a sources panel.
• A search indicator appears while ChatGPT is working.
• Recent facts appear that postdate the model's training.

Seeing the exact queries depends on the model version. Developers found that ChatGPT's web interface stopped exposing search queries in its page data with GPT-5.3 in March 2026, then exposed them again from GPT-5.6 in August 2026, so check before relying on it. The API's web search tool returns them either way. Methods are covered in how to use browser tools to reveal ChatGPT's search queries and fan-outs.

How can you influence both kinds of answer?

For searched answers:

• Make sure OAI-SearchBot can crawl your site, and that key pages are indexed in the search engines ChatGPT uses.
• Answer the sub-questions ChatGPT searches for, such as prices, comparisons, reviews and features. Nectiv found words like reviews, free, features, comparison and the year were common in its queries.
• Keep dates and facts current, since ChatGPT now looks for this year's and last year's content.
• Get listed on the third-party pages that searches retrieve, such as comparisons and directories.

For unsearched answers:

• Build a lasting presence in the sources models learn from, such as reputable publications, reference sites and widely read discussions.
• Be consistent about what you are and what you offer, so future training data describes you accurately.
• Be patient. Changes show up only with new model versions.

Does the user's plan or setting change search behaviour?

It can. Search is available to logged-in and logged-out users, but the model ChatGPT uses varies by plan and by the mode chosen, and each model has its own tendency to search. Reasoning modes that think for longer may run more searches on a single question, although no study we know of has measured this. Users can also switch search on for every prompt in a conversation, or turn on location sharing, which affects local results. None of this is visible to a brand checking its own visibility, so test with the settings your customers are most likely to use, and record which model and mode you tested. Deep research, a separate mode, searches many times by design, as covered in how deep research chooses which companies to recommend.

What might this look like in practice?

Consider an illustrative project management tool, Raskmere, tracking 60 prompts in ChatGPT. It split them by whether ChatGPT searched. For searched prompts, such as "best project management tool for architects 2026", Raskmere appeared in 40% of answers after it published a comparison page and was added to two review roundups. For unsearched prompts, such as "what project management software do agencies use?", it appeared in only 5%, because it launched after most coverage that models learned from. The team reported the two groups separately, focused short-term work on searched prompts, and kept building long-term coverage for the rest.

What do people ask about ChatGPT search?

Does ChatGPT always search the web?

No. It decides per prompt unless you turn search on. Studies found search in about a third of prompts.

Which prompts make ChatGPT search?

Recent events, local and shopping questions, requests for sources, and topics it is unsure about.

Does Perplexity always search?

By default, yes. Perplexity searches the web in real time for every question unless the user turns off the Web source.

Can I make ChatGPT search?

Yes. Select the web search option, or ask it to search and cite sources.