Agentic SEO: What It Is and How to Run It in 2026

Updated October 28, 2026, by the Rillow team. We make Rillow, an SEO data server for AI agents, and we planned this post with it. There are no affiliate links here.

Agentic SEO is SEO work done by an AI agent that takes several steps on its own. The agent looks up keyword data, reads who ranks in Google today, decides which page to write, drafts it, checks it and later tracks how it ranks, while a person sets the goal and approves what goes live. You can run it in 2026 with an agent you may already use, such as Claude Code, Codex, Cursor or Gemini CLI, once you connect it to real SEO data. This guide explains the loop, what the agent needs, what to keep for yourself, and shows a real run: the research behind this post.

What is agentic SEO?

In AI-assisted SEO, you ask a chatbot for one thing, such as a title or an outline, and you carry the answer to the next tool yourself. In agentic SEO, the agent carries the work from step to step. It calls a keyword tool, reads the numbers, calls a results tool for the keywords that look promising, compares them with your site, and comes back with a plan and a draft. Each step depends on what the last one found, which is the part a fixed workflow in Zapier or n8n can't do.

SEO.com defines it as a strategy where "AI agents plan and execute SEO tasks on their own". Siteimprove stresses that the agents run continuously, catching ranking drops and broken links before a person would notice them. Both pages say people still set the direction and approve changes.

Interest in the term is new and growing. On October 28, 2026, DataForSEO counted 170 searches a month for "agentic seo" in the US, 271% more than a year before, with a keyword difficulty of 0.

The other meaning of agentic SEO

The top Google result for "agentic seo" is Addy Osmani's agentic-seo project on GitHub, which uses the term for a different job. It audits documentation and websites for "Agentic Engine Optimization": whether AI coding agents can find, parse and use your docs. If you publish developer docs, that check is worth running. The rest of this guide is about the first meaning, agents doing your SEO.

How an agentic SEO loop works

An agentic SEO loop repeats the same seven steps for every page. Here is what the agent does in each one, and where its information comes from.

1. Read what it knows about the site. The agent starts from a saved description of the product, the audience, the voice and past decisions, so it doesn't ask you the same questions every session.
2. Research keywords. It looks up search volume, difficulty and trend for ideas around your product, from a keyword database rather than from memory.
3. Read Google's current top 10. For each promising keyword, it checks which kinds of pages rank (tool lists, guides, forums, vendor pages) and how strong those sites are, to judge whether your site can compete.
4. Write a brief and a draft. The brief lists what every ranking page covers and what none covers well. The draft adds what only you know about your product.
5. Run checks and fix them. Keyword in the title and first paragraph, title length, heading order, internal links, image alt text and stock AI phrasing.
6. Wait for a person. Someone reads the page and publishes it.
7. Track and refresh. The agent checks the page's Google position every day, and when it slips, it reads the new top 10 and updates the page.

Steps 2, 3 and 7 need data an agent doesn't have on its own. Claude, GPT and Gemini have no search volumes, no live rankings and no access to your Search Console. Without a data source they can still answer with a number that looks precise, so the first job in any agentic SEO setup is connecting one.

What an agent needs for SEO

MCP, the Model Context Protocol, is how the agent calls outside tools in the middle of its work. Our list of the best SEO MCP servers compares 12 data sources by what they give an agent and what they cost. If you only need one thing to start, connect a data source. Instructions and memory make the results more consistent, but they can't replace numbers the agent doesn't have.

A real agentic SEO run: how we planned this post

On October 28, 2026, we asked Claude Code, connected to Rillow's MCP server, to find two or three good pages for rillow.ai to write next. Here is what happened, step by step.

The agent read rillow.ai's memory and plan in Rillow first, so it knew our audience (founders who build with Claude Code, Codex and Cursor) and the 12 pages we had already published. Then it tried seeds around that audience, such as "claude code seo", "lovable seo", "ai visibility tracker" and "serp checker". This is part of what find_keywords returned for one of them:

{"seed":"agentic seo","country":"United States",
 "keywords":[{"keyword":"agentic seo","volume":170,"difficulty":0,
   "intent":"commercial","top10_domain_rating":45,
   "top10_referring_domains":5,"yearly_change":"+271%"}]}

Its first proposal was a free AI Overview checker, a post on Claude SEO skills and a Semrush pricing page. We asked for searches closer to 1,000 a month, and it came back with "semrush pricing" (1,900 a month), "ahrefs pricing" (1,000) and "ahrefs vs semrush" (880 in each word order), all with page-one results it judged winnable.

The numbers were right, and the plan was still wrong for us. People comparing Semrush and Ahrefs prices are shopping for an SEO suite, and Rillow is a data source for agents. So we told the agent we compete in agentic SEO and AI SEO. It looked up "agentic seo", "ai seo" and "ai agent seo", read Google's top 10 for the best ones, and found what you see above: 170 searches, difficulty 0, a page one of GitHub, Reddit, Siteimprove, WordLift, SEO.com, Search Atlas and a Medium post, and a median length of 1,871 words among the pages it could measure. It also pointed out that "ai seo" gets 6,600 searches a month but that Semrush, Neil Patel and Nightwatch hold its first page, so that one is a long-term bet.

The whole research round used about 4 Rillow credits, or about $0.40 of data. The agent did every lookup and comparison. The decision that mattered, which market to write for, came from a person, and that is the pattern in the next section.

What to keep for a person

An agent can do most SEO tasks, and a few are expensive to get wrong. Keep these four.

1. Choosing which searches to go after. The run above shows why: data ranks keywords by volume and difficulty, and only you know which searchers can become your customers.
2. Reading each page before it goes live. Google's guidance on generative AI content says "it is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." It also says that generating many pages without adding value for users may break its spam policy on scaled content.
3. Site-wide technical changes. One wrong canonical tag, noindex or robots.txt rule can drop many pages from Google at once. Let the agent find the problem and prepare the fix, and read the diff yourself.
4. Outreach. An agent can find sites that list tools like yours and draft a note, but you decide what to send and to whom.

SEO.com lists the same risks from the other side: agents with live access "can push unintended live changes, create duplicate content", produce schema errors and break links, so it recommends an approval step before changes go live.

How to start agentic SEO with your own agent

These steps use Claude Code. Codex, Cursor and Gemini CLI work the same way, and our setup page has the commands for each.

1. Connect a data source. For Rillow, run this in your terminal, then type /mcp inside Claude Code and sign in:

claude mcp add --transport http rillow https://rillow.ai/mcp

2. Add SEO instructions. Our free Rillow SEO skill installs as a Claude Code plugin, with guides for a new site, audits, keywords, content, AI search and measuring results:

/plugin marketplace add davit-khaburdzania/rillow-seo-skill
/plugin install rillow-seo@rillow

3. Ask for one job at a time. These three prompts cover the loop:

Read what Rillow knows about my site, then find 5 keywords my site can realistically win this quarter. For each, show volume, difficulty, who ranks today and why we could compete. Don't add anything to the plan until I pick.

Write the post for the first not-started item in my plan. Start from the brief, add details only we know about the product, and save it to Rillow. Fix every check marked fix. Don't publish.

Check how my published posts are doing. For any post that fell 3 or more places from its best, read the new top 10 and tell me what to change on the page.

4. Schedule the check. Claude Code runs a prompt on a schedule with /schedule, so the last prompt can run every Monday and send you what moved.

Our SEO automation guide goes task by task through what to schedule and what to keep, and the Claude SEO guide covers which Claude plan you need.

Hosted agentic SEO platforms or your own agent

Some products run the agent for you. Visibility.so describes "autonomous AI SEO agents that research, write, optimize, and build links", and SEO.ai calls itself an AI SEO agent that creates content, publishes it to your website and builds backlinks. Hosted agents suit teams that want the work done without opening a terminal.

Running your own agent suits people who already build with Claude Code, Codex or Cursor. You see each step and each number, the agent edits your site's code directly, every change comes back as a diff you can read, and you pay for the data your agent looks up instead of a seat. You also carry the setup, and you need to check the work yourself.

Frequently asked questions