How Grok searches: web searches, page browses and X

Published 7 November 2026. Updated 7 November 2026.

Grok searches the web when it decides a question needs current information. It writes its own queries, searches the web, opens the pages it wants to read in full and searches posts, accounts and threads on X, often several times for one answer, then marks its sources with numbered links.

Our page on how AI assistants search the web compares Grok with ChatGPT and Claude. This page stays with Grok: when it searches, its three kinds of search, how many queries it writes, what one answer looks like in xAI's own data, and what that means if you want Grok to read and cite your pages.

When does Grok decide to search?

Grok searches inside a loop that runs until it has enough to answer. xAI's tools overview says the model "analyzes the query and determines what information or actions are needed", then either calls a tool or answers, and repeats the cycle with whatever came back.

The search tools are built in and run on xAI's servers, so the queries are written and sent there, away from your browser. A question Grok can answer from what it already knows gets no search, and an answer with no search has no sources to cite.

What are Grok's three kinds of search?

Grok can search the web, open a page from those results, and search X. xAI's Web Search documentation says the web tool lets Grok "search the web in real-time and browse web pages", and its X Search page covers the rest.

Search	What Grok does	What comes back	What it tells you about your site
Web search	Runs a query Grok wrote against the web	A list of result pages	Whether your page came up for that query
Page browse	Opens one page to read it in full	The text of that page	Whether Grok read your page itself, beyond its listing in the results
X search	Searches X posts, accounts and threads	Posts, profiles and whole threads with replies	Whether posts about you, or your own account, were part of the answer

X search splits into four calls, and xAI's X Search documentation names each one:

• x_keyword_search runs a keyword search with X's advanced operators, sorted by Top or Latest.
• x_semantic_search searches posts by meaning rather than exact words.
• x_user_search finds accounts by name or description.
• x_thread_fetch pulls one post with the posts above it and the replies under it.

For a brand, the keyword call on your name and the user call for your account show whether Grok went looking for you on X at all.

Which search engine does Grok use for the web?

xAI does not name one. Its documentation describes Web Search as its own server-side tool and says nothing about an outside index behind it. Reports from 2024 and 2025 linked Grok to Bing or to a search API such as Tavily, and we have not seen xAI confirm either. The part you can check is the output: the queries Grok wrote for a turn and the pages that came back for them.

How many searches does Grok run per question?

More than ChatGPT, on the published numbers. xAI gives no figure, but Peec counted 5 million fan-out queries from ChatGPT, Grok and Perplexity between 1 and 21 April 2026 and found Grok writing 6.8 per prompt, against 2.1 for ChatGPT and 1.4 for Perplexity.

What Peec measured	Grok
Fan-out queries per prompt	6.8 (ChatGPT 2.1, Perplexity 1.4)
Chats with a site: search	18.3%
Chats that searched Reddit	10.5%, nine in ten of them as site:reddit.com
Wirecutter and Consumer Reports in the queries	Added by Grok every time, never typed by the user

Peec's write-up reads Grok's pattern as a research brief: a broad search first, then queries with the year added, brand against brand comparisons and searches aimed at named sources. These are one company's figures from its own prompt set, so treat them as a guide to the pattern.

The site: share matters for site owners. When Grok writes site:reddit.com or names a review site in the query, it has picked where to look before any ranking happens, so threads and reviews that mention you on those sites are part of what it reads.

What does one Grok answer look like in the data?

xAI's API returns two lists of sources with every answer, and the gap between them is where your page gets read but not cited.

Field	What it holds	What it tells you
citations	Every source the search came across	Whether Grok found your page at all
Inline citations	Numbered links in the answer text, starting at 1, with a number reused when a source comes back	Which sources the answer leaned on
web_citation or x_citation	Whether a citation points to a web page or an X post	Whether you were cited from your site or from a post about you
start_index and end_index	Where the citation sits in the answer	Which sentence of the answer your source backs

xAI's citations documentation says of the full list that "not every URL in this list will necessarily be directly referenced in the final answer", and that switching inline citations on "does not guarantee that the model will cite sources on every answer". Web pages and X posts sit in the same list, side by side. ChatGPT and Claude keep the same two lists under other names, and our post on fetched vs cited sources sets them out field by field.

Does Grok identify itself when it reads your page?

Not on the evidence so far. xAI publishes no crawler page for site owners, where OpenAI and Anthropic each list their bots by name. In a test on 6 February 2026, StackFox asked Grok to open pages on a site while logging the requests, and saw about 30 requests from two IP addresses at a VPN and a proxy provider, all sent as an ordinary Chrome browser, with no Grok name anywhere.

That is one small test, and xAI could change it. But it means a robots.txt rule that names a Grok bot may never match the request that reads your page. Whether Grok reads you depends mostly on whether your page comes up in the searches it writes.

How can you see what Grok searched in the Grok app?

The API gives developers the fields above. In the Grok app, Only Fan Out reads the responses grok.com already sends to your browser and lists, for each turn, the web searches, page browses and X searches Grok ran, its numbered citations resolved back to their sources, and the X posts quoted or referenced in the answer.

Searches Grok runs against a connected Gmail, Drive or Calendar are counted, but their query text is never stored. If Grok shows a Cloudflare block page, reload grok.com and ask again. For how the extension reads all this with no server involved, see how Only Fan Out works, and for the same view of Claude, how Claude searches the web.

What does this mean for your site?

Grok reaches pages through its own searches, and it writes more of them than ChatGPT, so start with the queries.

1. Read the queries Grok writes for prompts on your topic, and rank for them. Expect the year added and brand against brand comparisons.
2. Check the sites Grok names in its queries, such as Reddit and review sites, for threads and reviews that mention you.
3. Post on X under the name people search for, since Grok runs keyword, meaning and account searches there.
4. Treat robots.txt as a weak lever for Grok, since its page fetches may not carry a bot name.

None of this guarantees a citation. xAI publishes the tools and fields above, not how Grok chooses which of the sources it found to cite.

Related questions