How to build an AI agent for an SEO specialist

If you run an SEO or digital agency, you already know where the month goes: rank summaries, technical flags, meta tags, briefs — work that has to happen for every client, every cycle, and that nobody on the team wants to do twice. Hiring for it means a junior at 840,000 ₽ a year and three months before they're useful. It's 2026 and everyone's already tried pointing an AI at it; for some agencies the output is now part of the deliverable, and for others it's generic filler the client spots immediately. This article is the setup that puts you in the first group — one role, start to finish, using an SEO specialist as the worked example.

What this agent should actually do

The first mistake is scoping it as "does SEO." An agent needs a task it can finish and hand back, not an open-ended mandate. A working scope for this role: weekly rank tracking summaries, flagging technical issues from Google Search Console, drafting title tags and meta descriptions at scale, turning a content brief into a first-draft outline, and checking whether AI crawlers can actually reach key pages.

What it should not do: talk to the client directly, make strategic calls about which keywords to prioritize, or sign off on a technical migration. Those stay with a human — the agent handles the recurring, well-defined slice of the work.

The role: name, task, knowledge base

Creating the role itself is three clicks in the panel: a name ("SEO Specialist"), a task description, and a knowledge base. The task description is where most people undersell the setup — "handle SEO tasks" produces a generic agent. Write it the way you'd brief a new hire on their first day: what they own, what format the output should be in, and who reviews it before it goes out.

What to put in the knowledge base

• Your agency's technical SEO checklist — the actual one you use, not a generic one
• Client report templates, including the tone and level of detail clients expect
• A short list of "good brief" examples for copywriters, so the agent's output matches your bar
• Per-client notes: CMS, niche, known quirks (a client on a headless CMS needs different technical checks than one on WordPress)

An empty knowledge base is the single most common reason a new role feels "generic AI" instead of "our SEO process." The model is the same one everyone else is using — the knowledge base is what makes the output yours.

You don't need to write any of this from scratch. The material already exists in your agency: past client reports, the checklist someone maintains in a shared doc, the briefs that already got approved. Strip them of client-identifying details where needed and paste them into one plain Markdown file — one file, not ten. That single file is the document the role's retrieval searches over, and one file holding ten real examples gives that search more to draw on per query than ten separate files would.

The system instruction

Keep it specific about output format and boundaries, not just topic:

You are the SEO Specialist role for [agency name]. You handle recurring
technical and on-page SEO tasks for our client portfolio.

For every task:
- State which client and domain you're working on before anything else.
- Use the knowledge base named in the task as the source for our
  process, report formats, and that client's specifics.
- Pull data from connected tools (Search Console, rank tracker) rather
  than estimating — if a number isn't available from a tool, say so
  instead of guessing.
- Output format matches the client report template in the knowledge
  base unless the task says otherwise.
- Flag anything that needs a strategic decision (budget, prioritization,
  scope change) instead of deciding it yourself — hand it back with the
  relevant data attached.

Tools and integrations

Unlike a plain chatbot, this role doesn't need any of its data tools set up by hand. Search Console data, Yandex Metrica, PageSpeed, backlink/Domain Rating data via Ahrefs (Rapid), Bitrix24, website parsing, Google search, and an MCP server for real browser automation (Playwright/Chromium — the agent can actually load a page and look at it, not just fetch raw HTML) all ship with Orakul and are already available to every agent, this one included. There's no separate setup step where you "add the SEO integrations" — you write the task, and the agent already has the tools to do it.

The bot-access check is a good example: it's the same no-LLM check (robots.txt plus a live fetch per crawler) running behind the free tool on this site, available to this role without any extra configuration.

Schedule and delivery format

Recurring tasks should run on a schedule, not wait for someone to remember to ask. A weekly technical flag summary and a monthly client report covers most agencies' needs. Delivery matters as much as content: a finished PDF or spreadsheet in Telegram gets read; a wall of chat text gets skimmed and forgotten.

Testing and tuning

Run the role in parallel with a human for the first two weeks before trusting it unsupervised. Every time the output misses something a human would have caught, that's a knowledge-base gap, not a reason to abandon the role — add the missing context and re-run the same task.

The knowledge base is frozen at launch, unless you keep feeding it

This is the part people skip, and it's the one that matters most. Setting up the role this way gives you an agent whose knowledge is current as of the moment you built it — the underlying model's training data, plus whatever you loaded into the knowledge base at setup. It does not learn about a new Google update, a changed SERP feature, or a shift in AI-search behavior on its own. Nothing about "agent" implies "keeps up with the field automatically."

If you want a specialist that actually grows instead of one that's permanently stuck on its launch date, you need a standing process to feed it new material. A concrete version of that process: monitor a source your team already trusts — Search Engine Journal is a reasonable one for SEO — translate the pieces that actually matter into plain Markdown files, and load those into the role's knowledge base on a regular cadence. That's it; it's not more complicated than that, it just has to actually happen on a schedule instead of "whenever someone remembers."

Common mistakes

• Vague task descriptions. "Do SEO" produces generic output; "produce the weekly technical flag report using the template in the knowledge base" produces a usable one.
• Empty knowledge base. The model has no idea what "your" process looks like unless you tell it.
• Assuming the tools still need setting up. Search Console, Metrica, site parsing and the rest already ship with the role — the actual gap is usually in the knowledge base, not the toolset.
• Treating the launch-day knowledge base as permanent. Without a standing process to add new material, the agent's understanding of the field stops advancing the day you finished setting it up.
• Expecting strategic judgment from a routine role. Keep prioritization and client-facing decisions with a human; the agent's job is the recurring execution underneath them.

Knowledge base and skills aren't the same thing

Everything above is about the knowledge base — what the agent knows: your checklist, your report formats, your per-client notes. A skill is a different layer: it's what the agent knows how to do — a specialized capability that changes the technique behind the output, not just the reference material it draws on. Knowledge answers "what does our process look like." A skill answers "how do you actually execute this class of work at a specialist's level."

Skills are an Agency-plan capability on Orakul, on top of the per-agent tools every role already gets. A concrete example from an adjacent discipline: the conversion copywriter skill on mcpmarket.com adds structured conversion-copywriting technique on top of whatever process and facts the knowledge base already supplies. The knowledge base makes the role work the way your agency works; a skill raises the technical ceiling of what it produces.

Максим Сафьянов

I build Оракул: a self-hosted AI-agent orchestrator for digital and SEO agencies — unlimited roles, your own server, source code handed over at setup.