DeepSearch: the research agent, steered properly.
DeepSearch is Grok's research agent — it breaks your question into sub-questions, runs waves of searches across the web AND X, follows the promising links, and synthesizes a briefing. Used lazily it produces impressive-looking mush; steered properly it's a genuine analyst. The steering is this lesson.
01 How the loop works (so you can steer it)
When you toggle DeepSearch (or prefix a prompt with it), Grok plans sub-queries, searches in parallel, reads results, follows fresh links, summarizes, and repeats — up to a step limit. DeeperSearch is the same loop with more iterations and deeper link-following: noticeably slower, meaningfully more thorough. Two practical consequences fall out of this design:
- Your prompt is the research brief, not a search query. The agent plans sub-questions from what you wrote — so what you write should look like what you'd hand a junior analyst.
- It inherits the X superpower and the X problem. Real-time social signal that web-only research agents can't see — and a rumor mill in the source mix. Your prompt sets the evidence bar.
02 The DeepSearch brief pattern
Write the brief like you mean it
That closing line — "what you could not find" — is the single highest-value instruction for any research agent. It converts silent gaps into visible ones.
03 Three research workflows that earn their keep
The decision dossier
Vendor selection, tool adoption, market entry: ask for the case for, the case against, and — uniquely Grok — what current users are saying on X right now. Documented capability + live sentiment in one pass is the combination web-only research can't produce.
The landscape sweep
"Map the options for [problem] as of this month" — DeepSearch's iterative link-following genuinely shines at finding the newer entrants that a static index hasn't ranked yet. Ask for a table: option, maturity, pricing if public, who's vouching for it, who's complaining.
The claim autopsy
Someone forwarded you a confident claim. DeeperSearch, explicitly adversarial: "Find the strongest evidence FOR this claim, the strongest evidence AGAINST it, and trace where the claim originally came from." Tracing origin is the killer move — half of viral claims die on inspection of their birthplace.
04 The honest comparison
Grok DeepSearch wins when the question has a "right now" component — sentiment, breaking developments, what practitioners say this week. The X integration is unmatched, full stop. Perplexity wins on citation discipline and source transparency for documented topics — it remains the cleanest "show me exactly where this came from" experience (our research comparison). ChatGPT's Deep Research wins on very long, exhaustive structured reports from the documented web. The pro pattern, as ever, isn't loyalty — it's matching the question: live signal → Grok; verifiable citations → Perplexity; encyclopedic depth → ChatGPT.
Run one real dossier
Pick a decision you're actually facing and run the full brief pattern through DeepSearch. Grade the output: did it cite? Did it show conflicts? Did it confess what it couldn't find? Then run the same brief through whatever you used before — and you'll know exactly where Grok sits in your personal research stack.
What you can do now
- Explain the DeepSearch loop — and why your prompt is a research brief, not a query
- Use the brief pattern: purpose, source guidance, conflict-surfacing, and the "what you couldn't find" close
- Run decision dossiers, landscape sweeps, and claim autopsies
- Choose DeeperSearch when thoroughness beats speed — and budget your metered queries
- Place DeepSearch honestly against Perplexity and ChatGPT research by question type