Keyword Research With Muse: Give Meta's Agent Real Data

Keyword research with Muse means Meta’s Muse, the personal agent Meta launched on September 8, 2026, not MarketMuse, which is a different product. Muse gets its own computer in Meta’s cloud, with its own browser, and it keeps working after you close the app.

That browser is the trap. Ask Muse which keywords to target and it can open Google, read the top ten, and hand you a tidy table. Ask for search volumes and you will get numbers. There is nowhere on a Google results page it could have read them from.

Muse can plan, browse and run the work while you are somewhere else. What it cannot see is search volume, difficulty, or a clean read of who ranks in Google’s top 10 for the country you sell in. That half needs a live data source. The Muse setup page connects one with a single prompt and a short guided sign-in, and if you run a business on Muse’s new small business connectors, SEO for Muse small business is the shorter path.

What Muse can and cannot do for keyword research

Split the job in two and the whole topic gets simpler.

What Muse is good at: Taking a vague topic and turning it into themes. Grouping a long list by the job the searcher is doing. Reading the pages that rank and saying what kind of page wins. Writing the brief. And, the part that matters most here, doing all of that while your phone is in your pocket and messaging you when it is done.

What it cannot do: Tell you that a term gets 320 searches a month in the US. Tell you how hard it is to rank for. Tell you which of your own pages already sit on page two for it. None of that lives in a model, and none of it is printed on a web page Muse can open.

That split is not a Muse problem. Keyword research with Claude hits exactly the same wall, and so does every model you could put in Muse’s seat. The model is the analyst. The numbers have to come from somewhere else.

What makes Muse different is the computer. A chat model with no browser at least has no way to pretend it looked. Muse does.

Why Muse’s browser makes keyword research worse

Here is the take, and it runs against how Meta demos the product. For keyword research, Muse’s browser makes it worse, not better.

A rendered Google results page looks like data. It has positions, titles, domains, a clean order. It is not data. Google’s own help page on why search results differ between people lists time, location, language and personalization as reasons two searches for the same phrase return different pages. The page Muse reads is one version, for one machine, at one moment. Unless you check, you do not know which country Google thought it was serving.

And even a perfect copy of that page tells you nothing about demand. Ten blue links do not say whether ten people or ten thousand typed the query.

The real damage is to confidence. A model that has just read ten genuine titles has something that feels like evidence. When it then writes “about 2,400 searches a month” next to them, the made-up number borrows the credibility of the real titles beside it. You would catch a bare guess. You will not catch a guess sitting in a table of facts.

There is a rules problem too. Google’s spam policies describe “scraping results for rank-checking purposes” without permission as machine-generated traffic. An agent opening Google every Monday to see where you rank sits close enough to that description that I would not build a weekly task on it.

The fix is not to take the browser away. It is to tell Muse, in every prompt below, that numbers come from the SEO tools and nowhere else. A live data tool returns volume and difficulty for a country you name, and returns them as data the next step can sort.

Muse’s connectors cover your business, not your search data

On September 29, Meta added Muse for Small Business. According to Unite.AI’s report, the new connectors are Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, alongside the user’s Facebook and Instagram business accounts. Help Net Security adds Meta’s line that nothing is “published, sent or spent” without your approval.

That is a strong list for running a shop. Muse can read your Shopify sales, your Stripe revenue and your Klaviyo campaigns. Not one of those apps knows what anyone typed into Google last month.

So a small business owner can ask Muse “which products should I write about?” and get an answer built from sales data and a browsed results page. The sales half is real. The search half is the browser problem from the section above.

Meta has opened connector building to outside developers, and TechCrunch reported more than 1,500 applications in less than a week. We submitted ContextBolt SEO to that directory. Open Settings, then Connectors, and look for it. If it shows up there, connect from there. If it does not, use the Custom Connector route below.

Connect a live source with one prompt

Muse has no MCP settings box. Meta’s help center says to ask Muse to create a Custom Connector for any service outside the list, and Muse guides you through it. Start a ContextBolt SEO trial, then send this.

Copy this prompt

Create a Custom Connector for ContextBolt SEO. It is a remote MCP server
over streamable HTTP at https://seo.contextbolt.app/mcp. It uses OAuth
sign-in, so open the sign-in page and I will enter my email code. Once
it is connected, list the SEO tools it gives you.

When Muse opens the sign-in page, check the address is on seo.contextbolt.app, enter your SEO account email and the code, and allow the connection. Meta does not review custom connectors, and its help page says so, so the domain check is yours to do.

Then send a new message asking what SEO tools it has. You want keyword research, keyword difficulty and SERP overview at minimum. If the list is empty, the sign-in did not finish. The Muse setup page has the personal URL fallback for that case, and SEO with Muse is the long walkthrough.

One more step pays for itself. Ask Muse to connect Google Search Console and approve read-only access on Google’s screen. Search Console tools spend no credits, and your own impressions are the best keyword source you have. The Muse Search Console setup covers that step on its own.

From one seed to a shortlist in four credits

This is the core run. It goes from a single phrase to one recommended target, and it stops before any writing happens.

Copy this prompt

Before you start, ask me for my site, who buys from it, and one seed
phrase people type into Google. Then run keyword research once on that
exact seed, 50 ideas, United States. Use the SEO tools for every number.
Do not open Google in your browser and do not estimate any figure.
Group the ideas by the job the searcher is doing, then pick the three
groups that fit my buyers. Pull the live top 10 for the best term in
each group and tell me who ranks and whether anything there is weak.
Recommend one target with your reason. Stop there. Do not draft.

Here is why each line is there.

Ask for the business first: A keyword can be easy and worthless. Muse cannot tell which unless it knows who buys from you and what a sale looks like. If Muse already has your Shopify or Stripe connected, it can read some of this itself, which is the one place the business connectors genuinely help keyword research.

One seed, run once: The keyword research tool returns up to 50 ideas per call, each with volume, difficulty and intent. It costs one credit. Running it on six variants of the same idea spends six credits and gives you six tables to reconcile. One good seed beats six.

A phrase, not a name: Seeds are things people type. A brand, product name or domain usually comes back empty. An empty lookup costs nothing, but it tells you nothing either. To research a competitor, ask for the keywords their domain ranks for instead.

A country, not a city: Keyword data comes by country, with the United States as the default. If you sell in one town, read the numbers as national demand and judge local fit yourself.

The live top 10, three times: Each SERP overview is one credit and returns the ranking pages with position, domain, title and URL, for the country you named. That is the clean version of what the browser was pretending to do. Add the one research call and the whole run is about four credits.

Stop before drafting: A target that still looks good tomorrow is worth a page. One that does not was cheap to throw away. Drafts are the expensive part, in your time if not in credits.

Every result also saves to your SEO Dashboard on its own, so the run you did on Tuesday is there next week without scrolling a Muse transcript.

A weekly shortlist Muse runs on its own

This is where Muse earns its place over a chat you have to remember to open. Meta’s help page on reminders and scheduled tasks says Muse can create recurring tasks that repeat daily, weekly or on a custom interval, and that they continue until you cancel them. You describe the task and the timing in plain words.

The best weekly keyword job does not start from a seed at all. It starts from your own Search Console, because the terms you already rank for just off page one are cheaper wins than anything new.

Copy this prompt

Every Monday at 9:00 AM, pull my Search Console quick wins for the
last 28 days. Take the five queries with the most impressions and
check keyword difficulty for each with the SEO tools. Message me a
shortlist of three: the query, my position, impressions, difficulty,
and the page that should target it. Do not edit or publish anything.

The quick wins tool finds queries sitting just off page one, around positions 5 to 20, with real impressions over the last 28 days. It is free, because it reads your own data. The five difficulty checks are one credit each, so the whole task is five credits a week, roughly 20 to 25 a month out of 1,000.

Three habits keep it that cheap.

Weekly, never hourly: A weekly read is enough for this job. A task that fires every hour spends the month before you notice.

Check the Upcoming tab: Meta lists every scheduled reminder and task there, and you can ask Muse what is scheduled at any time. Recurring tasks keep going until you cancel them, so cancel the ones you stopped reading.

Keep the last step yours: The prompt says do not edit or publish. Keep that line. Meta’s own pitch is that Muse advances the work on its own, and a plan to rank for something will happily include publishing. Search feedback takes months. A wrong title tag live for a quarter costs more than the hour you saved.

What makes a Muse shortlist survive

Muse will hand you thirty targets if you ask. Three is a better answer, and these four questions get you there.

Does the top 10 leave room? Not “is the difficulty low” but whether anything on that page is beatable. A forum thread, a five-year-old post or a thin listicle is beatable. Four established publishers and a help center page are not, this quarter.

Would the person searching buy anything? Informational terms build an audience. Commercial terms build revenue. You need both, but pick on purpose each time.

Do you have something real to say? If your page would be a summary of the ten already ranking, it competes with the pages it summarizes and loses.

Can you name the page it becomes? If you cannot describe the finished page in one sentence, the target is a topic, and topics turn into unfocused posts.

Then ask Muse the question it will not volunteer. “What is the strongest reason not to target this?” Asked “is this a good keyword?”, every agent leans toward yes.

Where it still falls short

The numbers are estimates: Volume and difficulty come from a large keyword index, not from Google’s own ledger. They are decision-useful and directionally right, as in every SEO platform. Your real clicks live in your Search Console, which is why the weekly task starts there.

It needs your context every time: Muse does not know your margins, your season or the page you shipped last week unless you tell it or connect the app that does. Thin context gives a generic shortlist.

A target is not a working page: Once a term makes the list, the page it lands on still has to load fast and be indexable. Site audit with Muse is the next run, on the same connection.

Muse is not Muse Code: Meta’s terminal coding agent is a separate product with its own setup. Everything here is the personal agent on the web, iOS and Android.

Running it with ContextBolt SEO

ContextBolt SEO is ours, so weigh the bias. It is a hosted SEO data server your agent connects to over MCP. Because it lives at a public HTTPS address, Muse’s cloud computer can reach it, where a server running on your laptop never could.

Once connected, Muse can ask for keyword research, keyword difficulty, the live top 10, a domain’s traffic, the keywords a competitor ranks for, backlinks, page and site audits, AI visibility, and your own Search Console. You ask in plain words and Muse picks the tool. Every result saves to your SEO Dashboard automatically, dated, so a task that runs every Monday builds a record you can read back.

The trial is 7 days with 100 credits, which is roughly 25 seed-to-shortlist runs. After that it is $35 a month for 1,000 credits. Checking your balance is free, and so is everything that reads your own Search Console.

Is Muse the right agent for this?

For keyword research as a one-off, any agent with a live data source does the job, and Claude does it well. The model matters less than whether it can see real numbers.

Where Muse pulls ahead is the second week. A shortlist that arrives every Monday, built from your own Search Console, checked against real difficulty, sitting next to your sales data, is a habit most small sites never manage to keep by hand.

Just do not let it browse its way to the numbers. Give it the data, and keep the publish button for yourself.