Grok SEO: get cited from the web and from X

Grok answers from its training when it can and searches when a question needs something newer. Its current flagship, Grok 4.7, has a knowledge cutoff of May 2026, per xAI’s models page, so a price change, launch or incident after that reaches an answer only through search. SpaceX’s May 2026 S-1 describes Grok Chat as offering “real-time integration of web search, X data, code execution, and multimodal analysis” on grok.com, the Grok apps, inside X and through the API.

xAI’s API exposes the same two search tools, and their documentation is the clearest public account of how Grok gathers sources. The model works in a loop: at each step it “decides what to do next” — call a tool or answer — and “continues until sufficient information is gathered,” per the tools overview. Nothing stops it calling both tools for one question, and the side-by-side below is what each one does.

On grok.com both kinds of citation sit inline in the answer. The web client, as served on 26 October 2026, types every citation as web or x_post, labels each web chip with the site’s name taken from its domain — reddit.com shows as Reddit — and folds several sources behind one chip as “+N”. Through the API, the citations field lists every URL the agent encountered, and inline citations mark the ones it used; xAI’s citations docs warn that “not every URL in this list will necessarily be directly referenced in the final answer.”

The two studies measure different things. Ahrefs’ Brand Radar counted every domain Grok cited across 1.9 million broad US queries. Presenc AI, which sells Grok monitoring, ran 4,400 prompts and classified each citation as an X post or a web page; it reports that the mix “shifts toward X heavily for current-events queries and toward web for definitional or how-to queries.” Read together: a question about what happened this week leans on X, and a question about what something is or how to do it leans on the web. Work out which kind your buyers ask, and weight the next two sections accordingly.

X Search is four lookups, and each rewards something specific about how a brand posts. None of this is a ranking secret: it’s the capabilities in xAI’s X Search docs, mapped to what an account owner can change.

Two more documented behaviours shape what gets picked. Grok’s published chat prompt says it “can analyze individual X user profiles, X posts and their links,” so a post that links your canonical page is also a route to the page. And the from_date and to_date filters let the agent restrict X results to a window, which suits the current-events questions X is cited for most: post on the day a fact changes, not a week later.

• One fact, one post. “Pro is now $19 a month, from today” can be matched by keyword and by meaning; a launch-day meme can’t.
• Link the page that proves it. The post gets the speed; the page holds the detail Grok’s web search can open and quote.
• Keep one thread per question you get every week. Pricing, availability, migration steps — one thread, kept current, is a ready-made source for x_thread_fetch.
• Keep the handle and bio boring. Searchable beats clever when a lookup runs by account name.
• Correct in public. If @grok answers wrongly about you under a post, reply with the primary source. The next search has something to find.

When Grok answers in public: @grok and Explain

Grok also answers inside X itself: people tag @grok under a post, and the Explain button summarises one. xAI’s published system prompts — last updated in November 2025, so they may lag the current models — tell @grok to use search “to confirm facts and fetch primary sources for current events,” that “parallel search should be used to find diverse viewpoints,” and that it “must use the browse page to verify all points of information you get from search.” Replies are capped under 550 characters with no markdown. Explain is told to “prioritize peer-reviewed research.” Both are public posts in the thread, so a wrong answer about your brand is visible to everyone reading it.

xAI names no search provider for Grok’s web results, publishes no index size and offers no webmaster console or URL submission. It isn’t on IndexNow’s list of participating engines either. What is visible is the output: Ahrefs’ June 2026 ranking of the 50 domains Grok cites most, built from Brand Radar data, is led by user-generated platforms.

So the web half is two jobs: make your own pages openable by a fetcher that doesn’t announce itself (next section), and be part of the conversations on the platforms Grok already cites. Rankbox’s Reddit Presence finds the threads worth joining and drafts replies posted under your own name.

OpenAI, Anthropic, Google and Perplexity publish their AI crawlers. xAI publishes no crawler page, user-agent token, IP list or robots.txt statement for Grok’s search or its live page fetches. Cloudflare wrote in September 2025 that “xAI’s bot, grok, does not self-identify at all, making it impossible for website operators to block it.” A controlled test by StackFox on 6 February 2026 found the same: asked to read a news site, 30 requests hit the server within a second — 19 from one datacenter IP, 10 from another in a third of a second — all claiming to be an ordinary browser, desktop Chrome on both main IPs. “The word ‘Grok’ never appears. Neither does ‘xAI.’”

The consequence: robots.txt can’t target Grok, and neither can a CDN’s “block AI bots” switch, because nothing in the request says it’s Grok. Nothing documents whether Grok’s fetcher reads robots.txt at all. Naming the unverified tokens does no harm, but don’t mistake it for control — your default group is what applies:

# xAI documents no user agent for Grok's search or page fetches.
# These names come from third-party lists and are unverified:
# a rule for them may never match a real request.
User-agent: GrokBot
User-agent: xAI-Grok
User-agent: Grok-DeepSearch
Allow: /

# Grok's browser-like fetches fall under the default group
User-agent: *
Allow: /
Disallow: /account/

Server-rendered HTMLRequired

In Search Engine World’s June 2026 test, one node of Grok’s fetch pipeline ran the page’s script, but the answer still quoted the raw-HTML decoy: “The machinery ran the script; the answer ignored what it found.” Put the content in the HTML response — see server-side rendering.

No challenge on content pagesRequired

A JavaScript challenge, CAPTCHA or datacenter block hands Grok’s fetcher a page with none of your content on it.

Rate limits that allow burstsHelps

StackFox logged 19 requests in about a second from one IP. Per-second limits tuned for people will cut some of them off.

Titles that work on a source cardHelps

grok.com’s chips show only the site’s name; the card behind each one shows the page title and a short preview. The title is what a reader weighs before clicking.

Links from your X postsHelps

Grok can analyse the links in the posts it reads, so a post that links the canonical URL is a second path to the page.

Structured dataUnconfirmed

xAI says nothing about schema.org. Keep every fact that matters in visible text as well.

llms.txtUnconfirmed

xAI hasn’t said Grok reads llms.txt, and no published test has shown it requested.

robots.txt rules naming GrokNo effect

There’s no documented token to name. A rule for GrokBot or xAI-Grok may never match a request.

IndexNow pingsNo effect

xAI isn’t a participating engine, so pings never reach Grok.

Test pages the way Grok’s fetcher appears to see them — the exact user agent StackFox logged, no JavaScript — ideally from a cloud VM, since that’s where its requests come from:

# The user agent StackFox logged from Grok, and no JavaScript
curl -s -A "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/139.0.0.0 Safari/537.36" \
  https://yoursite.com/pricing | grep -c "Pro is \$19"

# 0 = the text needs JavaScript, or your WAF served a challenge page

Training: no website opt-out, one X setting

The Grok 4 model card says Grok 4 is pre-trained on a recipe that “includes publicly available Internet data.” With no documented crawler token, a website has no robots.txt line xAI says it honours for training. X posts are different: X’s privacy policy, effective 15 January 2026, says X may use what it collects “and publicly available information to help train our machine learning or artificial intelligence models.” Account holders can switch that off under Settings and privacy → Privacy and safety → Grok & Third-party Collaborators, per Roboin’s 2026 walkthrough. As far as anything documents, that setting governs training, not whether X Search can find and cite your public posts.

Expect little Grok referral traffic in analytics, and know that some of the shortfall is measurement, not absence. Two facts about grok.com, both checked on 26 October 2026, explain most of it.

GA4’s AI Assistant channel won’t rescue the Direct visits. Google’s channel definitions describe it as traffic “from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok,” and classify a session there when the medium is ai-assistant — which GA4 sets only “if the referrer matches a list of AI Assistants.” No referrer, no match. Catch the visits that do carry one with a custom channel, and treat a Direct spike on a page Grok cites as a lead worth checking:

# Session source: grok.com, when a referrer survives
^(.*\.)?grok\.com$

# X itself: t.co wraps links in X posts
^(t\.co|x\.com)$

Server logs are the better instrument, with a catch: you can’t filter Grok by user agent. Two methods work instead. Look for its signature — a burst of requests for one URL within a second, a desktop Chrome user agent, a datacenter IP, no referrer — and run a controlled test: publish a URL nobody knows, ask Grok to read it, and see exactly what arrives. Rankbox’s free AI crawler log analyzer handles the named bots; the bursts below are the part only a query like this finds.

# Bursts: 10+ no-referrer requests for one URL in one second (combined log format)
awk '$11=="\"-\"" {print substr($4,2,20), $7}' access.log \
  | sort | uniq -c | awk '$1 >= 10' | sort -rn | head -20

# Controlled test: ask Grok to summarise https://yoursite.com/grok-test-7f3a, then
grep "/grok-test-7f3a" access.log

For visibility itself, run a fixed prompt panel through Grok on a schedule. Mix current-events prompts, where X dominates, with how-to and product prompts, where the web does, and record which domains and @handles get cited. Answers vary between runs, so compare across repeats rather than trusting one screenshot. The free AI visibility prompt generator gives you a starting set.

Grokipedia launched on 27 October 2025 with about 885,000 machine-generated articles, and its homepage now lists more than six million. Nobody edits it directly: readers submit suggestions through a form, and Grok is meant to review each one. That review stalled for months — Lawfare found no article updated after 24 April 2026 and 13,002 suggestions stuck “in review” as of August. Grokipedia v0.3 shipped on 30 September 2026 with a redesign; whether the backlog has cleared isn’t documented.

Its documented reach is mostly outside Grok. Ahrefs’ March 2026 comparison counted 356,200 AI citations of Grokipedia against 24.9 million for Wikipedia, with the most coming from ChatGPT and Google’s AI Mode, then Copilot and Gemini. grokipedia.com isn’t among the 50 domains Grok cites most in Ahrefs’ June 2026 data, and xAI doesn’t document Grok using it in answers.

• Search Grokipedia for your brand, products and founders. Generated pages can exist without anyone asking for them.
• Suggest corrections with a primary source attached — your documentation, a filing or a dated announcement — then check back on the suggestion’s status rather than assuming it was applied.
• Don’t build a Grok strategy on it. Its citations show up in ChatGPT and Google’s AI answers far more than in Grok’s.

Grok was built by xAI, which SpaceX acquired effective 2 February 2026, according to its S-1. The AI business has carried SpaceX branding since: SpaceXAI, and then SpaceX Super Intelligence. Musk said on 4 October 2026 that SpaceXAI would be renamed SpaceXSI, Reuters reported, and the company’s X account became @SpaceXSI around 10 October. The developer docs still live at docs.x.ai under the xAI API name, and some pages already credit “SpaceXSI.” None of the renaming changed how Grok searches — but search for all three names when you look for documentation.

Paid placement isn’t a route. The S-1 says SpaceX does “not currently sell or offer advertisers the ability to place ads on the Grok API,” and describes ads in the stand-alone Grok product only as something there “might be” an opportunity for. Until that changes, a citation in Grok is earned through the two halves above.

1. 01Heard

   Allowing GrokBot in robots.txt gets you into Grok.

   Checked

   xAI documents no such token, and a February 2026 controlled test never saw it in a request. Grok’s fetches arrive as desktop Chrome, so what matters is that your default rules and your WAF let them through.

2. 02Heard

   A CDN's block-AI-bots switch keeps Grok out.

   Checked

   Those switches match declared bots, and Grok’s fetcher “does not self-identify at all” (Cloudflare). Only behavioural rules — challenges, datacenter blocks, rate limits — catch it, and they catch real people too.

3. 03Heard

   Grok mostly cites X posts.

   Checked

   It depends on the question. Ahrefs’ June 2026 data put x.com at 1.4% of top-50 citations, behind Reddit, YouTube and Facebook; Presenc AI’s panel put X posts near 45%, heaviest on current events.

4. 04Heard

   Ranking on Google means ranking in Grok.

   Checked

   xAI names no index provider for its web search. Google rankings may help indirectly by making you easy to find and discuss, but there’s no documented link.

5. 05Heard

   GA4's AI Assistant channel shows your Grok traffic.

   Checked

   Only visits that carry a referrer. grok.com’s citation chips use noreferrer, so those clicks arrive as Direct.

6. 06Heard

   Grokipedia is the way into Grok.

   Checked

   xAI doesn’t document Grok using it, and it isn’t among the 50 domains Grok cites most. Its documented citations are in ChatGPT and Google’s AI answers.