What to look for in an SEO MCP server

Where the analysis runs

A server that hands back raw rows asks the model to do statistics over a truncated sample, which is exactly the situation where a confident wrong answer is likely. A server that computes decay, cannibalization, CTR curves and rank movement before the model sees them leaves it doing what it is good at, which is interpretation.

A checklist worth running

These are the questions that separate servers in practice, and every one of them has a checkable answer in the documentation.

• First-party data from your own accounts, or third-party estimates
• Transport and protocol version, and whether it needs a local process
• Authentication: revocable keys, scopes, and whether a read key can be talked into writing
• Documented rate limits, and what the server returns when you cross one
• A default date range that accounts for reporting lag, so the model does not need today's date
• Whether mutations are audit-logged

Tool count is a poor proxy

Selection accuracy falls as a catalogue grows, because the model has more near-synonyms to choose between. A focused set with clear titles and honest descriptions is easier to use than a catalogue that lists everything the vendor has ever built.

How AutoPSEO answers it

Sixteen tools over Streamable HTTP at protocol version 2025-06-18, analysis computed server-side against stored daily rows, bearer keys that are read-only by default, 60 requests per minute per key with an HTTP 429 and a Retry-After header beyond it, and a REST surface at /v1 with an OpenAPI 3.1 document for anything that does not speak MCP.

Test before you standardise

Connect a candidate, ask the same three questions you ask every week, and check two things: that the numbers reconcile with the Search Console UI once you allow for the date window, and that the answer states which periods it used.