Last month we looked at our Search Console queries and found something odd: people were searching for "keyword research mcp" and "seo tools mcp" — and finding almost nothing worth clicking. Real demand, thin results.
So here's the walkthrough we wish existed when we started. The same steps work with any MCP server that exposes keyword data.
What MCP changes about keyword research
Without MCP: you think of a topic in your chat, switch to a keyword tool tab, export a CSV, paste it back, and lose the thread. With MCP: your AI client calls the keyword tool directly and you stay in one conversation. The research stops being a detour and becomes part of the thinking.
A concrete 15-minute workflow
Seed list from the customer's own words. Paste the last ten client emails into the chat and ask for the recurring phrases buyers actually use — not the industry jargon. This is your raw seed list, 15-25 phrases.
Expand and score in one pass. Ask your client to pull search volume and difficulty for the seed list plus two levels of expansion. Read the numbers in the chat, not in a spreadsheet. Kill anything with difficulty above your site's realistic range — for a small team site that's usually anything above 40.
Cluster by intent, not by topic. "b2b marketing automation" and "best b2b marketing automation platform" are different intents — comparison versus category. Ask the chat to group keywords by what the searcher is trying to decide. One page per decision, not per keyword.
Check the AI layer. Ask two or three AI assistants who they'd recommend in your category. The queries they cite are effectively tomorrow's keyword list — that's where buying research is moving.
What the numbers should warn you about
We ran this on our own site this week. Three months of data: 3,770 impressions, 10 clicks, average position 59.6. Last 28 days: 25 impressions. The lesson wasn't "write more content" — it was that one broad head term we were drifting toward was already lost (311 impressions, zero clicks). The terms we could actually win were the specific, tool-adjacent ones with a fraction of the volume. Small teams win long-tail decisions, not categories.
Common failure modes
- Scoring keywords before reading the customer's language. The data tool doesn't know your buyer says "research hours" when the industry says "market intelligence."
- Clustering by similarity instead of intent. Similar words, different decisions.
- Ignoring the AI-citation layer entirely. It's the only channel growing while classic organic flattens.
If you maintain an MCP server yourself, one practical note: health-check uptime matters more than you think. Directory crawlers test endpoints continuously, and a server that fails auth checks gets marked degraded in listings within days.
Disclosure: we build a marketing-data MCP server (SupaMarketers Data) and run supamarketers.com. This article is our honest workflow — the same steps work with any equivalent tool.
Top comments (0)