What is a web search API for AI agents?

What happens during one call?

Five things, in order, though vendors hide some of them:

1. Query handling. The service cleans the query and applies your filters: country, language, time range, included or excluded domains.
2. Search. It runs the query against a web index, either its own or a live search engine, and gets a ranked list of URLs.
3. Reading. For the top results, it fetches each page and checks whether the site allows it.
4. Cleaning. It strips navigation, ads and scripts and keeps the main content, usually as Markdown.
5. Packaging. It returns everything as JSON, ready to drop into a prompt.

With LLMWebSearch, step 3 covers the first 3 results by default, and up to 10 with read.

How is it different from a SERP API?

A SERP API returns the search results page; a web search API for agents returns the sources behind it. SERP APIs mirror what a person sees on a search engine: ads, maps, "People also ask" boxes, shopping cards and ten blue links. That suits SEO tools and rank tracking.

An agent needs less layout and more text. It wants a clean list of sources and the words on those pages. With a SERP API, the agent must call a second service to read each link. A web search API for agents does both in one call.

How is it different from a web reader API?

A reader starts from a URL; a search API starts from a question. A reader, or fetch tool, takes a link you already have and returns its content. It cannot tell you which pages to read.

Agents often need both. They search to find candidates, then read the best ones. A search API that also reads saves that second step for the top results. When the agent already has URLs, for example from a user message, a read call is enough. LLMWebSearch offers it as the extract endpoint.

How is it different from a model's built-in search tool?

A built-in tool searches on the model vendor's side and usually returns an answer with citations, not the pages. OpenAI, Anthropic, Google and Microsoft all sell one. As of October 22, 2026, OpenAI and Anthropic charge $10 per 1,000 searches plus the search content as tokens. Gemini 3 grounding costs $14 per 1,000 after 5,000 free searches a month on the paid tier. Grounding with Bing costs $14 per 1,000.

Microsoft states that results of Grounding with Bing are "not directly accessible for use in other applications". A standalone API returns raw sources you can store, filter and show. It also works with any model. The comparison of built-in tools has the details.

Where do the results come from?

Either from the vendor's own index or from a live search engine queried at call time. Brave says every result comes from its own independent index. Perplexity describes an index of "hundreds of billions of webpages". Exa's documentation says its lite mode serves cached content only and never fetches live.

Own indexes can answer fast, but they decide what is crawled and how often. Live search reflects what a search engine shows at that moment. LLMWebSearch works the live way: Results come from a live web search run at the moment of your call, in the order the search engine ranks them; nothing is served from an old index.

How do web search APIs charge?

Four models are common in 2026, often mixed:

• Per search. A fixed price per call, such as $5 per 1,000 at Brave, Perplexity, Linkup or You.com, as read on October 22, 2026.
• Per credit. Plans sell credits and an advanced search uses more of them. Tavily's basic search is one credit and advanced is two.
• Per result or per page. Valyu charges per result returned; Firecrawl adds one credit per page read.
• Per token. Jina's search bills at least 10,000 tokens per request.

Compare like with like: a price that includes full page text is not the same product as a price for snippets. The dated price comparison normalises them.

What should I check before choosing one?

Run twenty of your real queries through each candidate, then look at seven things:

• Content. Full pages, query-focused extracts, or snippets only?
• Freshness. Can you filter by time range? Do results carry dates?
• Control. Can you cap results, pages and characters per page?
• Price per useful search. Include the reads you would otherwise pay for separately.
• Billing terms. Do credits reset each month or carry over? Is a card or a work email required?
• Data rights. May you store results? Does the vendor train on your queries?
• Site etiquette. Does the reader respect robots.txt?

For LLMWebSearch: Search credits have no expiry date: what you do not use this month is still there next year. We do not train any model on your queries or on what the API returns, and we add no rule of our own against storing results; the rights of the sites whose pages you read still apply.