What is deep research? Deep search vs deep research for AI agents, explained

TL;DR

• Deep search is a short agent loop inside a single API call. It searches, opens the pages it finds, decides what to look up next, and returns within seconds. It answers "find this specific thing, even if it takes a few hops."
• Deep research is a long-running investigation that runs as a background job. It plans the question, follows many threads, cross-checks sources, and returns a verified answer or a cited report after several minutes. It answers "what is the complete, verified answer to this complex question?"
• Both search more than once. They differ in how much work they do and how your application waits for the result. Deep search fits inside a tool call. Deep research is a job you submit and come back to.
• Linkup offers both: /search with depth: "deep" for deep search (about 5-30 seconds), and the asynchronous /research endpoint for deep research (about 2-20 minutes). Linkup /research scores 61% on Seal-0, the hardest set of the SealQA benchmark. That is the highest published Seal-0 score we have found as of October 2026.

"Deep research" and "deep search" appear everywhere in AI tooling in 2026. Google calls a research feature in AI Mode "Deep Search." OpenAI, Gemini, and Perplexity all ship products called "Deep Research." Exa has a "Deep" search mode and a separate long-running research product, Exa Agent. A 2026 arXiv survey, Deep Research of Deep Research, notes that deep research "currently lacks a unified definition" and that its relationship with "deep search" "remains ambiguous."

This article gives you a working distinction that holds across providers. It is based on what each tool does and how you call it, not on what it is named.

What plain web search does

A standard web search takes a query, runs it against an index, and returns a list of results: URLs, titles, and short snippets of text from each page. The snippets come from the index. The search engine crawled and processed those pages ahead of time, so your query retrieves pre-computed text and does not open any page live.

This is fast, cheap, and correct for most factual lookups. The limitation is that you only get what the index extracted, usually a passage around the matching text. If the answer is buried in a table on page 6 of a long document, a snippet will usually not carry it.

In Linkup, this is depth: "flash" (under 200 ms) and depth: "fast" (about 1 second). Both pass your query straight to Linkup's index with no language model in the loop.

depth: "standard" (1-3 seconds) adds one planning step. A model interprets your request, can run several searches in parallel, and can read one URL you include in the query. It is still a single pass. It cannot use what it found to decide on a next step within the same call.

What deep search adds

Deep search turns one search into a short loop. It runs a search, opens and reads the pages that look relevant, judges whether it has what it needs, and then searches again or opens a page it just discovered. It returns when it has the answer or runs out of budget, typically within seconds.

Reading the full page matters as much as iterating. Deep search works from the page itself, including JavaScript-rendered content and tables. A snippet only tells you the page probably has the answer. Deep search tells you what the page actually says.

In Linkup, this is depth: "deep" on the /search endpoint. It takes about 5-30 seconds and follows step-by-step instructions literally. For example:

First find Datadog's pricing page URL.
Then scrape that URL.
Then return plan names, per-host prices, and included features as JSON.

Deep costs more and takes longer than the other depths. It is the right setting when the answer depends on a page the search has to find first, or on data that only appears inside the page.

Linkup's /fetch endpoint is a related but different tool. You give it a URL you already have, and it returns the full page as clean Markdown. Use /fetch when you know where to look. Use depth: "deep" when the search needs to discover the page and then read it.

What deep research adds

Deep research is a full investigation, not a longer search call. It has a budget of minutes, not seconds, and spends it on planning, breadth, and verification.

A deep research run on a complex question does something like this:

1. Break the question into sub-questions and plan the investigation.
2. Search several threads in parallel.
3. Read the results.
4. Decide whether the evidence is sufficient, conflicting, or raises new questions.
5. Follow the new leads with searches shaped by what step 4 revealed.
6. Cross-check claims across independent sources.
7. Synthesize the findings into a cited report or a verified answer.

Deep search also runs steps 3 to 5 in a small way. Deep research runs many more rounds across many more threads, and adds steps 1, 6, and 7 at scale. That is why it takes minutes and returns a finished deliverable, not a set of pages for your agent to reason over.

Linkup's /research endpoint runs this as an asynchronous task. POST /v1/research returns a task ID immediately. You poll GET /v1/research/{id} every 5-10 seconds until the status is completed, then read the output. You choose:

• A mode. answer returns a precise answer verified across sources. investigate returns a focused report on one subject. research returns a structured report across many topics or entities.
• A reasoning depth. S (2-5 minutes, $0.25) through XL (10-20 minutes, $2.50). The default is L (5-10 minutes, $1.50).
• An output type. A cited answer, or JSON that matches a schema you provide.

How Linkup /research performs on SealQA

SealQA is a public benchmark of fact-seeking questions written so that ordinary web search returns conflicting, outdated, or misleading results. Its hardest set, Seal-0, has 111 questions. A system answers each one, and a judge marks the answer correct or incorrect against a verified reference answer. The score is the share of questions answered correctly.

Linkup /research scores 61% on Seal-0. It answers about 68 of the 111 questions correctly. For context, the SealQA authors report 43.2% for GPT-5 with search at its best setting. Parallel reports 56.8% for its most expensive Task API tier. Each vendor runs the benchmark with its own harness, so treat cross-vendor comparisons as indicative, not as a controlled head-to-head.

SealQA measures whether the final answer is right when the web is noisy. It does not measure the structure or readability of a long report. That makes it a fit for agents that act on a researched answer.

When to use each one

The mistake teams make is sending every lookup to deep research because it sounds more thorough. Deep research takes minutes and costs dollars per run, not cents. For individual lookups in a tool-calling loop, a search depth is the right call. Reserve deep research for questions that need many rounds of searching and verification, and where the caller can wait.

The reverse mistake is asking deep search for a multi-section report. Deep search is built to find and read the right pages quickly. A broad report needs the time budget of deep research.

How the terminology maps across providers

Product names do not line up with these categories, so classify by behavior:

• Google "Deep Search" in AI Mode issues hundreds of searches and returns a fully cited report in a few minutes. By the definitions here, that is deep research, despite the name.
• Grok "DeepSearch" launched in 2025 as a multi-step research agent. xAI later folded it into Grok's Expert and Heavy modes.
• OpenAI "Deep Research" in ChatGPT is a multi-step report agent: deep research. OpenAI shut down its dedicated deep research API models (o3-deep-research, o4-mini-deep-research) on July 23, 2026. It now points developers to its general models with web search.
• Gemini "Deep Research" is a deep research agent. Developers call it through the Interactions API as a background job and poll for the result.
• Exa "Deep" (type: "deep", about 4-12 seconds) is deep search. Exa Agent is deep research. At its highest effort level, Ultra, a typical run takes about 30 minutes.
• Linkup /search with depth: "deep" is deep search. Linkup /research is deep research.

When you evaluate a tool, ask two questions. Does it return inside one call in seconds, or run as a background job over minutes? Does it give your agent pages to reason over, or a finished answer or report? The answers tell you whether you are comparing deep search or deep research.

Why this distinction matters for AI agent architecture

If you are building an agent, choosing between deep search and deep research is an architectural decision, not just a parameter setting.

Search fits inside a synchronous tool call. Your agent sends a query, gets a result in a fraction of a second to a few seconds, and continues. Flash and fast suit ReAct loops, voice agents, and chat. Standard suits requests that need interpreting. Deep suits steps where the agent can afford tens of seconds to get a page-level answer.

Deep research is asynchronous. Your agent submits a question and polls for the answer over minutes. This fits scheduled workflows, background jobs, and any task long enough to run offline and return when ready.

For most agents, the right pattern is to use fast as the default search tool. Escalate to standard or deep when the request needs interpretation or page reading. Route to /research only when the user asks for a comprehensive or verified result, or when the task clearly needs many rounds of retrieval.

The full technical reference for Linkup's endpoints, including /search, /fetch, and /research, is at docs.linkup.so. For how web search accuracy affects the quality of what your agent produces, see Best Web Search API for AI in 2026, Benchmarked on Accuracy.

FAQ

What is deep research?

Deep research is a long-running agent workflow. It plans an investigation, runs many rounds of search where each round is shaped by the previous one, cross-checks sources, and returns a cited report or a verified answer. It takes minutes and runs as a background job. It is the right tool for complex questions that need reasoning across many sources, not for single lookups.

What is deep search?

Deep search is a short agent loop inside one search call. It runs several searches, reads the full content of the pages it finds, and can open a page it discovered in an earlier step, all within seconds. It is deeper than standard search because it reads pages and follows leads. It is still a single synchronous call that returns pages and an answer, not a long investigation.

What is the difference between deep search and deep research?

Both iterate. Deep search runs a few rounds within one call and returns in seconds, so it fits a synchronous tool call. Deep research runs many rounds across many threads, adds planning and cross-checking, and returns after minutes, so it runs as an asynchronous job. Deep search gives your agent evidence to use. Deep research gives it a finished answer or report.

Does Linkup offer deep research?

Yes. Linkup's /research endpoint runs asynchronous deep research in three modes (answer, investigate, research) and four reasoning depths (S to XL, about 2-20 minutes, $0.25-$2.50 per run). It scores 61% on SealQA's Seal-0 set. Linkup's /search endpoint with depth: "deep" offers deep search, which follows leads and reads full pages within about 5-30 seconds. They are different endpoints for different jobs.

When should an AI agent use deep research instead of web search?

Use deep research when the question needs many search rounds, where each round depends on the previous one, and when the caller can wait minutes. Examples include competitive analysis, due diligence, research reports, and hard factual questions where sources conflict. Use web search for lookups in agent tool-call loops where the user is waiting. Use deep search when one of those lookups needs to find a page and then read it.

Which providers offer deep research APIs?

Linkup /research, the Gemini Deep Research agent (through Google's Interactions API), Exa Agent, and Parallel's Task API all run multi-step research workflows through an API. OpenAI's ChatGPT Deep Research is a consumer product, and its dedicated deep research API models were shut down in July 2026. The main differences are latency, accuracy on public benchmarks, output format, and compliance options.

On SealQA's Seal-0 set, Linkup /research reports 61%, the highest published score we have found as of October 2026. Zero Data Retention, where queries and results are never written to persistent storage, is available from Linkup on request.