Do AI model APIs search the web by default?

Most model APIs do not browse the web unless you ask them to. Web search is a web search tool that you attach to the request. If you leave it off, the model answers from what it learned in training.

Even with the tool attached, search is not guaranteed. The model decides per query whether a search is worth running. The evidence base behind this page describes retrieval as "per-query and discretionary on every platform."

That gives you two different kinds of answer from the same API:

• Retrieval off (parametric). The model answers from its training data. Your current website plays no part.
• Retrieval on. The model runs one or more searches, reads results, and may cite them. Your site's crawl access and index status matter here.

What it means: an API answer about your brand tells you little until you know which of the two you got. The next move is to treat "did search run" as a required field, not a footnote.

Each vendor exposes search differently. The common thread is that you opt in and the model decides.

Some details worth knowing:

• OpenAI says the number of sources "is often greater than the number of citations." Retrieved and cited are different sets. Our playbook on logging retrieved vs cited sources covers how to store both.
• Anthropic charges $10 per 1,000 searches plus tokens. Claude chooses whether to search, and the integrator's system prompt steers that choice. See what Claude's web search tool docs reveal.
• Gemini grounding is a separately billed, developer-invoked path. It bills per search query on Gemini 3 and per prompt on 2.5 and older. It is not the same thing as Google AI Overviews or AI Mode.
• Perplexity retired its Sonar Chat Completions endpoint on 2026-09-27. On the Agent API that replaced it, web search is opt-in. Our write-up of the Sonar to Agent API migration has the detail.

An answer with no search behind it is a snapshot of the model's training data. It stops at a knowledge cutoff the vendor publishes.

OpenAI's model docs list a Feb 16, 2026 knowledge cutoff for the GPT-5.6 family. The Instant path, chat-latest, carries an Aug 31, 2025 cutoff. So "ChatGPT's knowledge" is not one value. Which model you call changes what it can know by several months.

For a brand, this has a hard consequence. If you changed your pricing, launched a product, or rebranded after the cutoff, an ungrounded answer cannot reflect it. No page edit, robots.txt change or markup reaches that answer this quarter.

What is in your control is the retrieval-on path. That path depends on access: whether search crawlers may fetch your pages, whether your content is in the HTML before JavaScript runs, and whether your pages are indexed. Fixing those does not promise a citation. It removes the blocks that make one impossible.

Search enabled, search not run

Attaching the tool does not mean a search happened. The model can answer directly. If you only log the request settings, you will record grounded intent and ungrounded output.

Silent migrations

Perplexity code that was not migrated after the Sonar retirement can return ungrounded answers without an error. Nothing in the response tells a careless pipeline that web access disappeared. Our piece on whether the Sonar retirement affected tracking walks through the risk.

Consumer apps behave differently

The ChatGPT app searches at the model's discretion. Claude.ai search is a user toggle with no published default. An API call with search forced off, or forced on, will not match either. Compare the two in API answers vs consumer app answers.

The queries are not your prompt

OpenAI says ChatGPT search "typically rewrites your query into one or more targeted queries." Gemini grounding generates its own queries. Your prompt is the input you control, not the query the search engine sees.

Any visibility number built on API answers averages two different things unless you separate them. Log these fields for every answer:

1. Model and version string, plus the call date.
2. Search tool enabled: yes or no.
3. Search actually ran: yes or no, taken from the response, not the request.
4. Search queries issued, where the API returns them.
5. Retrieved sources and cited sources, stored separately.
6. Errors and empty responses, so failed calls are not scored as clean answers.

Then report retrieval-on and retrieval-off results as separate series. A drop in mentions might mean the model changed, or that search stopped firing. You can only tell them apart if you logged it. For the case for testing both states, see whether to test with web search on or off.

When you evaluate any tracking tool, ask one question first: does each result show whether search ran? If it can't answer that, its trend line mixes two measurements.

Does the Perplexity Agent API search the web automatically? No. On Perplexity's Agent API, web search is opt-in. You have to enable the `web_search` tool in the request. Code that skips this gets answers from the underlying model's training data, with no live results. This matters most for teams that moved off the Sonar Chat Completions endpoint, which retired on 2026-09-27, because an incomplete migration fails quietly rather than with an error.

If I enable the web search tool, will every answer be grounded? Not necessarily. Enabling the tool gives the model permission to search. The model then decides per query whether a search is needed, and on Claude the system prompt also steers that decision. Check the response for search results or retrieved sources. If none are present, treat the answer as ungrounded even though the tool was on.

Is an API answer the same as what users see in ChatGPT or Claude? No. The apps make their own retrieval decisions. The ChatGPT app searches at the model's discretion, and Claude.ai search is a user toggle with no published default. Apps can also route to different models than the one you call. API answers are a useful, controllable proxy, but they are not a copy of the consumer experience.

Can I improve how an ungrounded answer describes my brand? Not this quarter. An ungrounded answer comes from training data frozen at the model's knowledge cutoff. Nothing you change on your site reaches it until a future model is trained. What you can control is the retrieval-on path: crawler access in robots.txt and at your CDN, content present in the raw HTML, and index eligibility. Those remove blocks; they do not guarantee a mention.

Does Gemini grounding mean my result matches Google AI Overviews? No. Grounding with Google Search on the Gemini API is a separate, developer-invoked path. Google states that AI Overviews and AI Mode may use different models and techniques, and the Gemini API writes its own search queries. A grounded Gemini API answer is a proxy for Gemini, not a measurement of what AI Overviews show for the same question.