A few weeks ago we shipped the MCP connector, which lets your own AI assistant read your live SEO data: rankings, competitors, Search Console history, AI visibility, the security and readiness scans. It changed the kind of question you could ask. It did not, on its own, change the quality of the answers as much as we hoped.
Here is what we kept seeing. Ask a connected assistant “why is traffic down” and it looks at the first chart it finds and offers the first explanation that fits. Usually that is “a Google update”, because that answer needs no further work, and it is wrong about half the time. The assistant had every number it needed. What it lacked was the method: what to rule out first, what counts as proof, when to admit it is not sure.
So we wrote the method down
An AI skill for SEO is a short instruction file an assistant reads when a particular kind of question comes up. It says which tools to call and in what order, what evidence settles the question, what to check before blaming anything, and how confident the conclusion is allowed to be. The traffic drop skill, for instance, makes the assistant confirm the drop is real before explaining it. A site tracking a handful of keywords has an average position that swings when one keyword moves, and a single day’s dip can recover the next morning. Then it rules Google in or out against the confirmed update timeline, separates a ranking loss from a click loss at unchanged rankings, names the domain that took the slot, and checks whether your own deploy or a blocked crawler lines up with the date. Only then does it answer.
Every skill ends with one of three words. Confirmed means the data shows the mechanism directly. Likely means the timing and pattern fit and nothing contradicts. Possible means it is a hypothesis the stored data cannot settle, and the skill says what would. A skill that cannot answer says so rather than guessing. That last rule turned out to matter more than any other.
What the fourteen cover
Two are for getting going: a setup skill that writes your business context, seeds keywords from your Search Console history, pins the competitors your results actually show and adds the first buyer questions for AI visibility, and a weekly brief across every site you track. Six are about rankings and traffic: the traffic drop, keyword opportunities within reach of page one, a review of one page, pages slowly decaying, whether a release on a known date helped or hurt, and what to publish now for the next seasonal peak. Three are about AI visibility: where assistants mention you and where they do not, why one stopped citing you, and which competitors get recommended instead. Two cover domain health: the certificate, email authentication and header audit, and broken backlinks turned into a redirect list. The last one turns any finding into a report you can hand to a client or a board, with the data dates and a confidence legend on it.
They are all on the skills page, each with the question it answers, something you might actually type, and the tools it reaches for. Every card links to the skill file itself, so you can read exactly what your assistant will be told to do.
The setup skill is the one that writes
Most of the fourteen only read. The setup skill is the one that adds things, and it is the reason the connector is no longer read-only. It can add keywords in the country you sell in, add the questions you want tracked in AI assistants, and pin a competitor. A few of the others offer the same when they find something worth tracking, under the same rules. Every one of those costs you something, keywords are fetched daily and AI questions cost per run, so the rules around it are strict. It shows you the list against your plan’s remaining allowance and adds nothing until you confirm the exact items. A batch that would exceed the allowance is refused before anything is written, in full, not partly. And it pins a competitor only when the stored results show that domain actually ranking for your keywords; the refusal comes back with better candidates instead. Nothing on your website is ever touched. The connector has no access to it.
Free, including for client work
The skills are on GitHub under a plain-English licence. Use them, change them, share them, run them for your clients. The one thing you may not do is sell them or charge for access to them. Reading your data through them needs an account with the connector enabled, which is £5 per user per month on every plan, and reading as often as you like costs nothing extra because every skill works from snapshots already collected.
They are written in the Claude skill format, so Claude Code picks one up from a folder without being told. Each is also plain instructions in a single file, so you can paste the body into a project, a custom GPT or another agent’s system prompt and it behaves the same way. If you ask a question that none of them covers, open an issue and tell us what it would take. A skill has to be answerable from the tools the connector exposes today, which is the rule that keeps them honest.