Tracking What Grok Says About Your Brand

Grok is xAI's assistant. The evidence base this library is written from contains no xAI documentation. That means we can't tell you, as of October 2026:

• which user agent Grok uses to crawl or fetch pages
• whether a robots.txt rule controls Grok's search, its training, or both
• when Grok searches the web and when it answers from memory
• whether Grok exposes the full list of pages it read, or only the ones it cites

Compare that with the vendors that do publish this. OpenAI's crawler page names GPTBot (training), OAI-SearchBot (ChatGPT search) and ChatGPT-User (user actions), and states "Each setting is independent of the others." Anthropic and Perplexity document their own bots, and Meta names Meta-WebIndexer. If you track those engines too, see how ChatGPT brand monitoring works and tracking Claude's answers.

For Grok, don't guess a crawler token or a setting. Check xAI's own current documentation before you write any rule, and treat everything else on this page as observation.

Every AI answer comes from one of two places. Either the model searched the web for this query, or it answered from what it learned in training, frozen at a cutoff date. On every platform our evidence covers, that choice is made per query and is often the model's call.

This matters for tracking. A wrong price in a searched answer points to a page the engine read. A wrong price in an unsearched answer points to old training data, and nothing you publish this quarter will reach it.

Since Grok's retrieval behavior is undocumented in our evidence, the only signal you have is visible: did the answer show sources or not. Log that for every answer. Don't label an answer "grounded" or "ungrounded" beyond what the screen shows you.

Keep this simple enough that you'll actually run it every month.

1. Write 10 to 20 buyer questions. Use the questions a prospect asks before buying: "best [category] for [use case]", "is [your brand] worth it", "[your brand] pricing". Decide whether to include competitor names; tracking prompts that name competitors covers the tradeoff.
2. Freeze the wording. Paste each prompt exactly the same way every run. Small wording changes produce different answers.
3. Control the session. Use a fresh, logged-out or new session where you can, with no custom instructions. Saved context changes answers on other platforms, so assume it can here.
4. Run on a fixed schedule. Monthly is enough for most brands. Add a run after any big launch or pricing change.
5. Log every answer, including failures. An error or refusal is a row, not a skip.

What to record per answer

Presence counts are easy and tell you little on their own. The more useful number is accuracy.

• Count errors as a rate. Divide answers with at least one wrong fact by answers that returned. Established factuality metrics such as FActScore normalise by a count of observations; none simply sums penalties.
• Exclude failed responses from the denominator. An answer that errored out is not a clean answer. Vectara's hallucination leaderboard keeps refusals out for the same reason.
• Keep the worst error separate. One answer saying you were acquired, when you weren't, matters more than ten wrong headcounts. A pure average hides it as your log grows.
• Separate "wrong" from "can't verify". If Grok states a number you never published, that's unverified, not proven false.

There is one outside data point. The Tow Center at Columbia Journalism Review tested AI search tools on news citations and reported Grok 3 answered 94% of queries incorrectly. That was news, not brand facts, on an older model, so treat it as a reason to check, not a prediction for your brand.

When the log shows a wrong fact, the next step is finding where it came from. A tool that alerts you when AI says something false automates this for the engines it covers.

The levers with evidence behind them are access levers. They remove blocks; they don't guarantee citations.

• Put key facts in the raw HTML. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Whether Grok's fetcher renders JavaScript is not documented in our evidence. Server-rendering your pricing and product facts covers the crawlers that are documented, and costs nothing for the rest.
• Check your firewall, not just robots.txt. A CDN or firewall can block a crawler that robots.txt allows. Look in your server logs for requests identifying as xAI or Grok, and see whether they get a 200 or a 403.
• State facts plainly. A specific price or one-line definition on a fetchable page gives an engine a sentence it can quote.
• Look at the cited URLs in your log. When an engine cites other sites instead of yours, being named accurately on those sites is associated with appearing in the answer.

What the evidence does not support: schema, llms.txt, or "AI-friendly" rewriting as ways to get cited. A critical survey of 45 GEO studies found no technique with a stable, cross-platform effect on discoverability.

Be clear about the limits before you report numbers upward.

• It's a sample. Your prompts are a proxy. Engines rewrite queries internally, and you never see those rewritten queries.
• It's not a rank. No engine in our evidence publishes position data for AI answers, so "position 2 in Grok" can't be checked against anything.
• Cited is not the same as used. Research on how models cite found citation output can be changed without changing the sources.
• Third-party pages can mislead it. A 2026 study of LLM search agents found that manipulated web content can sway what agents endorse once it's in the results. Read whether a competitor can manipulate AI answers and scan cited pages for hidden instructions. Never plant such content yourself.

99Visibility's paid plans audit ChatGPT, Google AI Overviews, Gemini, Perplexity and Claude, not Grok, so the manual log is the method here. For the engines a tool covers, see free AI brand monitoring across several engines, and track Meta AI answers the same way if your audience uses it.

Which robots.txt token controls Grok? Our evidence base contains no xAI crawler documentation, so we can't name one. Don't copy a token from a blog post without checking it against xAI's own current pages. Meanwhile, check your server logs for requests that identify as xAI or Grok, and confirm your firewall isn't returning 403s to them. A robots.txt allow means nothing if the CDN blocks the request.

How often should I run my Grok prompt log? Monthly is enough for most brands, with an extra run after a launch, rebrand or pricing change. What matters more than frequency is consistency: the same prompts, worded identically, in a fresh session each time. Answers vary between runs, so one bad answer is a signal to recheck, not a trend. Look for errors that repeat across several runs.

Does Grok search the web for every question? As of October 2026, that isn't documented in the evidence base behind this library. On the platforms that are documented, searching is decided per query, often by the model itself. So log whether each Grok answer shows sources. An answer with no sources may come from training data, which nothing you publish now can change quickly.

Can social listening tools track Grok answers? Social listening tools watch public posts and mentions. An AI answer is generated privately for one user and isn't published anywhere, so these tools generally don't see it. To know what Grok says, you have to ask it yourself or use a tool that queries it. Our page on [social listening and AI answers](/questions/social-listening-ai-answers) explains the gap.

What should I do when Grok gets a fact wrong? First, confirm the correct fact is stated plainly on a page of your site that loads without JavaScript. Second, check whether your own pages contradict each other, for example a different price on the pricing page and the footer. Third, look at the URLs Grok cited and correct any third-party page you can. Then rerun the prompt next cycle. Nobody can promise the answer will change.