Web search API glossary

An interface that runs a web search on request and returns the results as structured data, usually JSON, instead of a page made for human eyes.

Search engine results page: what a search engine shows for a query. A SERP API returns that page as data, without reading the pages it lists.

The list of pages a search returns, in the order the engine judged most relevant. LLMWebSearch keeps that order and numbers each result with position.

The short extract under each result's title. Useful for choosing what to read, too short to answer most questions on its own.

The words sent to the search. Agents often rewrite a user's question into one or more shorter queries before searching.

A special token inside a query, such as quotes for an exact phrase or site: to stay on one domain. The q parameter passes them through as written.

A search run at the moment of the request, as opposed to a lookup in an index built earlier. 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.

A search limited to news articles, usually with dates. Set topic to news, and add time_range to keep only recent stories.

A rule that keeps results to some sites or removes others. The include_domains and exclude_domains parameters take up to 20 domains each.

The part of the service that opens the top result pages and extracts their text. Ours introduces itself as LLMWebSearch/1.0 (+https://llmwebsearch.com/bot/); the page reader page describes it.

The name a client sends with every web request. A reader that states its real name lets site owners recognise it and set rules for it.

A text file at the root of a website that tells robots which paths they may fetch. The standard is RFC 9309. Our reader respects robots.txt: it introduces itself as LLMWebSearch, follows the rules a site writes for it or for all robots, and leaves out the content of any page it is not allowed to read.

Lines in robots.txt that state how content may be used, for example ai-input=no to refuse use as AI input. A site that sets Content-Signal: ai-input=no in its robots.txt keeps its pages out of your results' content.

Keeping the article or documentation text of a page and dropping menus, footers, cookie banners and ads, so a model reads what matters.

A light text format where # marks a heading and - a list item. Models read it well, and it costs fewer tokens than HTML for the same page.

A browser with no window, driven by code. It lets the reader see the text of pages that only appear after JavaScript runs.

A gate that keeps content for paying subscribers. The reader never signs in and never works around a paywall; such pages come back without content.

The field that says whether a result's page was read and, if not, why: ok, not_requested, blocked_by_robots, target_refused and others listed in the docs.

Large language model: a model trained on text that writes text. Its knowledge stops at a training date, which is why it needs search for anything recent.

Giving a model sources to base its answer on, then asking it to cite them. Search results with page text are the most common way to ground on the web.

Also called function calling. The model asks to run a tool, such as web_search, with arguments it chooses; your code runs it and returns the output.

Model Context Protocol: an open standard for exposing tools to AI clients such as Claude Desktop or Cursor. A small MCP server can wrap a search API as a tool.

A program where a model decides the next step, calls tools, reads their results and repeats until the task is done. Web search is often its first tool.

Retrieval-augmented generation: fetch relevant text first, put it in the prompt, then generate. With a web search API, the retrieval step is the live web.

An agent pattern that plans many queries, reads many pages and writes a report with citations. It can run dozens of searches for one question.

The amount of text a model can take in at once, measured in tokens. Set max_chars to cap how much of each page enters that window.

The unit models read and bill by, roughly a short word or part of one. Clean Markdown spends fewer tokens than raw HTML.

A link from a statement in an answer to the source it came from. Returning sources rather than a written answer lets your model cite exactly what it read.

A confident statement a model makes up. Grounding on fresh sources and asking for citations reduces it; it does not remove it.

What each call spends. One credit pays for one search: up to 20 ranked results, with the pages of the first 3 read and returned as Markdown. Reading 4 to 10 pages makes it 2 credits. A call that ends in an error costs nothing.

A prepaid batch of credits, from 5,000 searches to 1,000,000 searches. Search credits have no expiry date: what you do not use this month is still there next year.

A crypto payment that funds your balance, at least $25. Packs are then bought from the balance.

The number of calls an account may make per minute. Every account has a per-minute call limit, shown in the dashboard; ask support and it can be raised.

A protocol that lets a client pay for one HTTP request in a stablecoin, with no account. Some search APIs accept it; LLMWebSearch is paid from a prepaid balance instead.

Know your customer: identity checks such as an ID document or a selfie. LLMWebSearch asks for none; the no-KYC page explains what still applies.