Web search for deep research agents: many queries, real sources

How many searches does one report take?

It depends on your design, and it is worth deciding on purpose. A light report might run two rounds of four queries: eight searches. A thorough one might run five rounds of eight: forty searches.

More searches do not always mean a better report. The gain from each new round falls once the main sources are in hand. A stop rule helps: end the loop when a round adds fewer than two new sources, or when the budget is spent.

With read at 3 or less, each search is one credit. A forty-search report therefore costs forty credits, whatever the length of the pages read.

Why read pages instead of working from snippets?

Because research is about details. Snippets give the gist of a page, but the figures, conditions and dates that make a report useful are in the body. A model that writes from snippets tends to fill gaps with guesses.

Reading the pages also improves the next round of queries. The agent learns the terms the sources use, the names of laws or products, and the dates of events. Those words make sharper queries than the original question.

LLMWebSearch reads the first pages of every search for you. Each result carries its position, title, URL, host, snippet and, when it was read, the page's main content as Markdown or plain text with its word count and publication date when known.

How do I stop the agent from citing the same page twice?

Key your source list by URL. Before adding a result, check whether its URL is already there. Different queries often return the same strong page, and you only need it once.

Some pages appear under two URLs, for example with and without tracking parameters. Strip query strings that only track clicks before comparing. For news, the same story often runs on several outlets; keep the one with the earliest published_at or the most detail.

Number the sources once, at the end, so the citations in the final answer match a clean list.

When should I raise read above 3?

When one query must cover a narrow subject in depth and the first three pages are not enough. Setting read between 4 and 10 makes that search 2 credits instead of 1.

In most loops it is better value to run another query. A new query brings new pages from a different angle, while pages 4 to 10 of the same query often repeat the first three. Keep the higher setting for niche topics where few pages exist at all.

How does this compare with ready-made deep research endpoints?

Several vendors sell research as a single call, priced per task. As of October 22, 2026, Exa's deep search types cost $12 to $15 per 1,000 requests, and Exa retired its separate /research endpoint on April 1, 2026. You.com prices research calls from $12 (Lite) to $1,200 (Frontier) per 1,000. Parallel's Task API runs from $5 (Lite) to $2,400 (Ultra8x) per 1,000 tasks. Linkup's deep search is $50 per 1,000.

These endpoints hide the loop, which is convenient. Building the loop yourself on a search API costs less per report at moderate depth, and you control the model, the prompts and the stop rule. The step-by-step guide has a working script of about 60 lines.