Keyword research with Grok got a twist on October 7, 2026. Elon Musk posted that Grok Bot, xAI’s always-on agent, will now use whichever back-end model is most likely to do a task best, and he named Claude Opus 5.5 among them. So the bot you ask for keyword ideas next week may not be running Grok at all.
That sounds like it should fix keyword research. It does not. Open the Grok app, ask Grok 4.7 for the monthly search volume of “home espresso machine”, and you get a number. Ask again in a fresh chat and you may get a different one. Neither came from a search database, because the model does not have one.
This guide covers both ways people do keyword research with Grok. The Grok 4.7 chat, and Grok Bot with its routines. What each one is good at, the step that gives it real numbers, prompts you can copy, and a routine that hands you a keyword shortlist every Monday.
- Grok cannot see search volume, difficulty or Google’s top 10. Any keyword number it gives you without a live source is generated, not looked up.
- Grok is good at the judgment half. Expanding a seed, clustering by intent, reading a SERP, and finding the words people use on X.
- A smarter model does not fix the numbers. Even if Grok Bot hands the task to Claude Opus 5.5, an invented figure only gets more persuasive. The data is now the bottleneck.
- Connect the data in Grok Bot chat. ContextBolt SEO is a hosted SEO data server your agent connects to. One public URL gives Grok Bot keyword volume, difficulty, the live top 10 and your own Search Console. Free for 7 days, then $35 a month.
- Put the research on a routine, keep the decision. A weekly shortlist costs about four credits.
What Grok can and cannot do for keyword research
Split the job in two and the confusion goes away.
The judgment half is Grok’s. Turning one seed into themes. Grouping 200 terms by the job the searcher is doing. Reading a page of results and telling you whether a listicle or a tool page wins it. Noticing that three of your planned posts chase the same intent. That is reasoning over text, and Grok 4.7 is good at it.
The data half is not. How many people search a phrase a month. How hard it is to rank. Who sits in Google’s top 10 today. What your own site already gets impressions for. None of that is in the model. xAI’s model page puts Grok 4.7’s knowledge cutoff at May 2026, and even before that date the training data never held a keyword database.
Grok does have two live search tools, and people mistake them for keyword data. X search returns posts and user profiles from X. Web search reads pages. Neither returns a search volume. A phrase can be everywhere on X this week and get 30 Google searches a month.
The fastest proof is the two-chat test from our Grok for SEO guide. Ask for the volume of three keywords, open a new chat, ask again. If the numbers moved, they were written, not retrieved.
The fix is to give Grok a live source, and the place that works is Grok Bot. If you already know you want this running every week, the Grok Bot SEO routines page has the ready-made set. The connection itself takes two minutes on the Grok Bot setup page.
Grok chat or Grok Bot: two different jobs
“Grok” means two products here, and they take data in different ways.
| Question | Grok chat (Grok 4.7) | Grok Bot |
|---|---|---|
| Where it runs | The Grok app and grok.com | Its own cloud computer, shipped with Cursor |
| Model | Grok 4.7, released September 21, 2026 | Announced October 7, 2026: the best back-end model per task |
| Live keyword data | Paste it in. Not a custom MCP host | Add a public MCP URL in chat |
| Runs on a schedule | No | Yes, with routines |
| Best for | X listening, clustering, reading a SERP you paste | Research that keeps going after you close the laptop |
Grok 4.7 is the current chat model. xAI’s Grok 4.7 announcement pitches it at long, hard tasks, and its 500,000-token window holds a full Search Console export without trimming. It is a strong analyst. It still analyzes only what you hand it.
Grok Bot is the one that can fetch. It keeps a computer running in the cloud, takes an MCP server when you ask in chat, and runs routines on a schedule. That makes it the right home for Grok Bot keyword research that does not need you sitting there.
There is a third Grok, the terminal agent. If you work in a repo, Grok Build adds the same server with one command. This guide sticks to the two above.
Today’s news: a smarter model is a better guesser
Here is what Musk posted on October 7, 2026.
“Important note regarding Grok @Bot: Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.”
The post is on X, and RuntimeWire’s write-up is careful about what it leaves open. It does not say which task goes to which model. It does not say whether you can pick one, or see which one ran.
So nobody outside SpaceX knows if a keyword question goes to Grok 4.7 or to Opus 5.5. For keyword research it barely matters, and that is the uncomfortable part.
A better model is better at everything that makes a fake number believable. Tighter reasoning about intent. A cleaner table. A plausible difficulty score next to a plausible volume, with a sentence about seasonality underneath. If a keyword task does land on Opus 5.5, an unconnected bot will produce a more convincing invented figure than it did last week. Not a truer one.
Claude hits exactly this wall, which is why our keyword research with Claude guide opens on the same split. The model was never the bottleneck for the numbers. Now that the best models can sit behind one bot, the data clearly is. Spend your attention on what the bot can see, not on which brain it borrowed.
Grok 4.7 keyword research in the chat
The chat cannot fetch keyword data, but there are three jobs it does well. Give it the material and keep it off the numbers.
Listen to X before you pick a seed. This is where Grok has an edge, because it reads X directly. It finds how buyers phrase a problem this week, which is often not how you phrase it on your page.
Copy this prompt
Search X for posts from the last 30 days where people describe
problems with home espresso machines. Group what they say into
themes. For each theme, quote the exact phrases they use.
Do not estimate search volume or difficulty for anything.The last line matters. Without it, Grok tends to decorate the themes with numbers.
Cluster a list you already pulled. Export your queries from Google Search Console and paste the file in. Those are real impressions for your own site, so the numbers are true before Grok touches them.
Copy this prompt
This is my Search Console query export for the last 3 months.
Group the queries by the page that should answer them. Flag any
group where two of my pages compete for the same intent. List the
queries where I rank between 8 and 20 with at least 100 impressions.
Use only the numbers in the file.Read a SERP you paste. Copy the titles and snippets of the top 10 for a term and ask what kind of page wins it and what all ten are missing. Grok 4.7 is genuinely good at this, as long as the ten results came from Google and not from its memory.
The limit is the loop. Every new keyword means leaving the chat to get its numbers. That is the gap keyword research with Grok Bot closes.
Connect live keyword data to Grok Bot
MCP, the Model Context Protocol, is the standard that lets an agent call an outside tool in the middle of an answer. Connect an SEO MCP server and “what is the volume” stops being something the model writes. It becomes something a tool returns.
Grok Bot has no settings form for this. You tell it in chat. Start a ContextBolt SEO trial or use your subscription, then send this.
Copy this prompt
Add this MCP server: https://seo.contextbolt.app/mcpWhen the bot offers a sign-in link, open it. Sign in with your ContextBolt SEO account email, enter the code we email you, and choose Allow connection. If your version of the bot cannot finish sign-in, the Grok Bot setup page has the fallback.
Then check it worked in a new message.
What SEO tools do you have? List them.You want keyword research, keyword difficulty, SERP overview and domain overview in that list. One rule trips people up. Grok Bot’s computer is in the cloud, so a server running on your own laptop is invisible to it. ContextBolt SEO is already hosted on public HTTPS, which is the shape the bot can reach.
Connect Search Console next. Ask the bot to connect it and approve read-only access on Google’s screen. Reading your own Search Console data spends no credits.
Grok Bot keyword research, step by step
Once the tools are there, the workflow is short. The discipline is in where you stop.
1. Give it the business first. Your site, your country, who buys, and what a conversion is. A keyword can be easy to rank for and worthless to you. The bot cannot tell which unless you say.
2. Expand once, from one seed. One expansion returns up to 50 ideas with volume, difficulty, intent and cost per click. Six expansions give you six tables to reconcile.
Copy this prompt
I sell refurbished espresso machines to home buyers in the US.
Run keyword research on "home espresso machine" for the United
States. Group the ideas by the job the searcher is doing. Mark
the ones with clear buying intent and difficulty under 30.
Use the tool's numbers only. Do not draft anything.3. Read the live top 10 before you commit. A difficulty score is a shortcut. The SERP is the evidence. Forum threads, a stale post or a thin listicle can be beaten. Four big publishers and a manufacturer page probably cannot this quarter.
Copy this prompt
For the three best targets from that list, pull the live Google
top 10. For each, tell me what kind of page ranks, who owns the
top three spots, and whether anything there is weak enough to
displace. Recommend one target and say why. Then wait for me.4. Shortlist, then stop. The bot will happily draft. Do not let it yet. A target that still looks good tomorrow is worth a post. One that does not was cheap to drop.
Every result also saves to your SEO Dashboard on its own, so Monday’s shortlist is still there on Thursday without scrolling a transcript.
A weekly keyword shortlist routine
Here is where Grok Bot earns its keep. Cursor’s help page puts it plainly: a routine tells a Bot when to run a workflow, on a schedule or after an event, and it runs in the cloud while you are elsewhere.
Copy this routine
Set up a routine to run every Monday at 8:00 AM.
1. Pull my Search Console quick wins: queries in positions 5 to 20
with real impressions in the last 28 days.
2. Run keyword research once on "home espresso machine" for the
United States.
3. Pick the three ideas with the best mix of volume, low difficulty
and buying intent. Skip anything my quick wins already cover.
4. Pull the live Google top 10 for those three only.
5. Message me one recommended target, two backups, and the single
quick win most worth a rewrite. Do not draft or publish.Change the seed to yours before you send it.
On the phone, open the Bot’s profile, tap Add routine under Routines, and Grok Bot starts a message that begins “Set up a routine to”. Finish it with the steps above. On desktop, click the Bot’s name, then Tasks, and ask in chat. The routine then shows under Routines, where you can Pause, Test, Edit or delete it. The schedule follows the time zone in Settings, under Bot.
Press Test once before you trust it. It runs real work straight away, so you see the tool calls happen and the message it sends you.
What it costs. Quick wins come from your own Search Console, so they are free. Keyword research is one credit. Three SERP pulls are three more. That is about four credits a week, roughly 16 a month out of 1,000. The free trial comes with 100 credits, so you can watch the routine run before you pay anything. A lookup that comes back with no data costs nothing.
Run the same seed every week and it still costs its credit each time, but the answer gets more useful. When something moved, a repeat SERP pull leads with it, such as a new #1 result.
More routines built on the same tools are on the Grok Bot SEO routines page.
Where it still falls short
Connected data is not used data. The model still chooses when to call a tool. Ask a vague question and you can get a vague answer built from nothing. Say “use the tool’s numbers” in the prompt, and check the reply has figures, not adjectives.
A real table looks messy. Awkward numbers like 1,300 and 90. A missing difficulty on a long-tail row. A flat “no data”. A generated table is tidier than that, every cell filled, every figure round. If it looks too complete, ask which tool it called.
The numbers are estimates. Volume and difficulty come from a large keyword index. They are decision-useful and directionally right, not Google’s own ledger. Your Search Console figures are the real ones, for your own site.
Countries, not cities. Keyword research takes a country, United States by default. If you sell in one city, use the country data for direction and your own Search Console for the local truth.
Loud on X is not demand. Grok’s X read is the best place to find language. It is not a reason to target a term. Check the volume before you write.
Publish stays with you. A routine can research and shortlist all year. Search results take months to move, so a page the bot published unread can cost a quarter before you notice. The SEO with Grok Bot guide makes the same call.
Grok or Claude for keyword research?
For the numbers, it does not matter. An unconnected Grok and an unconnected Claude are both guessing, and as of today Grok Bot may be running Claude under the hood anyway.
What Grok brings is narrower and real. The X read, which helps you find the words buyers use before you pick a seed. And Grok Bot, which keeps a routine running after you shut the laptop. If you do your research in one sitting at a desk, the Claude workflow is the same idea with a shorter setup.
Pick the assistant you already work in. Then connect ContextBolt SEO, because the half that decides whether a post can rank is the half no model brings with it. Start the free trial, send the bot one line, and ask it for this week’s shortlist.
