What is agentic SEO? One name for two jobs

Agentic SEO is two jobs under one name. The first is using AI agents to do SEO work. The second is making a site readable and citable by agents that browse and cite on a user's behalf. Ahrefs, SEO.com and Siteimprove define only the first.

The two jobs need different controls and different owners. I keep them apart, and the rest of this piece does the same.

What is agentic SEO?

Agentic SEO is the agent as your staff, or the agent as your visitor. The three definitions below all describe agents doing SEO work for the site owner. They differ on how much autonomy the agent gets and on the goal, and none of them covers agents as visitors to the site.

One name, two jobs

Ahrefs defines the first job in its May 2026 agentic SEO guide ↗︎: "Agentic SEO means applying AI agents to SEO workflows so they can act, adapt, and recover on your behalf". The agent works for you, inside your tools.

The second job is the reverse. web.dev's guide to agent-friendly websites ↗︎ says a site now has a new type of visitor: an autonomous system that interprets input, plans, and acts for a user. Here the agent works for someone else, and your site is the thing it reads.

Where the definitions disagree

SEO.com's definition ↗︎ reads: "Agentic SEO is a new approach to search engine optimization where AI agents plan and execute SEO tasks on their own." The Siteimprove ↗︎ version reads: "Agentic SEO is the use of proactive AI agents to continuously optimize digital content for discoverability." Siteimprove's agents are the site owner's tools, and it points their output at discoverability in AI-driven search. That blurs job one's method with job two's goal. Its page has the agents "plug directly into CMS platforms" to monitor technical health and fix errors, but the aim is being found, which is a different job from being read by an agent.

SEO.com's definition stresses autonomy, yet the same page says it is hard "to rely on it to be fully autonomous" and that "it still requires some level of human oversight to keep it in line". So its definition and its fine print pull apart. Ahrefs says the work is not fully autonomous and needs a person in the loop. That is a human in the loop: someone who approves what matters before it ships. Check who is accountable before you buy anything labeled agentic.

A useful test is whether the product chooses its own next step or follows a fixed script. Ahrefs draws the line: a workflow tool such as n8n or Zapier runs a pipeline someone builds and maintains, and an agent is given an outcome and adapts. The post that sorts AI SEO tools by job applies the same test to the tools themselves.

Sort any phrase with three questions:

1. Who is the agent working for: you, or a user who has never heard of you?
2. Whose site does it touch: yours, or everyone's?
3. Who signs off on the result?

"AI SEO agent", "SEO agent" and "AI agents for SEO" almost always answer job one. "Agent-friendly website" answers job two. Shopping agents, covered below, are job two.

How do AI agents do SEO work?

An AI SEO agent is a language model with tools that runs a loop on an SEO task: read data, decide the next step, act, check the result. Ahrefs' guide ↗︎ describes the connector as MCP, the Model Context Protocol, the standard plug that links an agent to data and actions such as your CMS.

What follows is a permission sheet for one task: a weekly scan for pages losing traffic. It is a template assembled from the sources, not a record of any client setup.

1. Reads. The agent may read Search Console, analytics and crawl exports without asking.
2. Writes. None to the CMS, redirects or robots.txt (the file at a site's root that tells crawlers which pages they may fetch) without approval.
3. Plan first. It shows the plan before any multi-step run.
4. Failure. It retries a failed tool call once, then reports the error.
5. Limits. It stops and asks past 30 minutes or 1,000 API calls.
6. Output. A ranked refresh queue, one sentence of reasoning per URL, and a flag on anything it is unsure about.

What the agent may do and its permissions

Rules 2 through 5 come almost word for word from the setup prompt in the same guide. That prompt makes production tools read-only unless the user approves a write action. It also sets the 30-minute and 1,000-call limits in rule 5.

Claude Code's permission modes ↗︎ show the same idea in a real product. In Manual mode (config value default), reads run without asking and edits stop for approval. "Plan mode tells Claude to research and propose changes without making them." On current versions, interactive sessions now start in auto mode, which runs everything with background safety checks. So the mode has to be set on purpose. Choose the mode before the agent touches a production system, not after.

How it fails and recovers

Ahrefs lists three failure patterns. An agent can make the same inference errors as the model underneath. A 500,000-row crawl can make it skip rows or stall. And long hands-off runs break more often than short ones.

Hence rule 4. Ahrefs' prompt tells the agent to retry a failed tool once, then surface the error instead of guessing. The run limits in rule 5 exist for the same reason: a stalled or looping agent should stop and ask.

Test the sheet on a copy of the data. Break a connection on purpose and see whether the agent reports the failure or invents a result.

Where a person decides

The agent proposes. A named person approves. Ahrefs says anything client-facing needs a human.

Put the decision point after the ranked queue and before any write. One person owns the queue, reads each reason, and approves or rejects each fix. Without a name on that step, the permission sheet is decoration.

WorkHacker's managed SEO automation page covers the run-for-you version of this setup. For the list of tasks worth automating at all, see what SEO automation is, and what to automate.

How do you become discoverable to agents that browse and cite?

Being found by agents is the second job, and the name for that work is GEO, generative engine optimization: work aimed at improving visibility in AI search experiences. GEO, AEO (answer engine optimization) and AI search optimization name the same discipline, and I use GEO from here on. Agentic SEO is the wider term, because it adds job one. The changes come in three groups: bot access, page structure, and what to skip.

Let the right bots in

Each AI company runs more than one bot, and the bots do different jobs. Blocking the wrong one removes you from answers or leaves training open. As of October 2026, these are the bots the vendors document:

OpenAI's crawler overview ↗︎ says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Anthropic's help article on its crawlers ↗︎ warns that disabling Claude-SearchBot or Claude-User may reduce your visibility in user search results.

Robots.txt is not a lock, and behavior differs by operator. For Perplexity-User, Perplexity's crawler page ↗︎ says "Since a user requested the fetch, this fetcher generally ignores robots.txt rules." Anthropic's help page says its bots honor robots.txt. If a page must stay private, use real access control.

Audit in order:

1. Open your robots.txt and list every AI user agent it mentions.
2. Match each to the table. Decide per bot: search, user fetch, or training.
3. Check your firewall. Perplexity notes a web application firewall may need to allow its bots explicitly.

Make the page usable by an agent

Agents see a page as screenshots, raw HTML, or the accessibility tree, which is the browser's semantic summary of roles, names and states. Most combine these. The web.dev guide ↗︎ says "Agents that take screenshots will likely be confused if your website layout is constantly shifting."

The same guide lists the fixes:

1. Use <button> and <a> elements, not styled <div> or <span> elements.
2. Link every <label> to its input with the for attribute.
3. Remove transparent overlays that cover interactive elements.
4. Keep actionable elements visible and larger than 8 square pixels.
5. Keep action locations stable across pages.
6. Audit the accessibility tree in Chrome DevTools.

web.dev also describes WebMCP, a proposed web standard where a site defines tools that tell agents what it can do. The guide warns it is in active development and may change. Treat it as something to read, not to ship. WorkHacker's AI SEO services page lists the site-side work.

What Google says you do not need

Google's AI optimization guide ↗︎ lists AEO and GEO as common terms and calls the work SEO from Google Search's perspective. Its mythbusting section says "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search". The same section says Google Search ignores llms.txt files, and that making one neither helps nor harms your visibility there. An llms.txt file is an optional text file a site publishes for AI systems to read.

It also says you do not need to chunk content, rewrite it just for AI systems, or chase inauthentic mentions. Spend that time on crawlable original content instead. Structured data is a different case: it is not required for generative AI search, and no special schema.org markup is needed. Google still advises keeping it, because it helps with eligibility for rich results on Google Search.

The limit: this is Google's position about Google Search. Other systems may read llms.txt, and Google says that is fine.

What about AI agents and product shopping?

Shopping agents are job two with a checkout attached. Shopping protocols are outside what this piece can source, so it makes no claim about them. The web.dev guidance ↗︎ carries over directly, and it names a catalog example.

What changes for a catalog

The guide's example of a layout that confuses agents is an Add to cart button in a different place for each product category. A catalog is exactly where that happens.

Check three things on a product template:

1. The Add to cart control is a native <button> in the same place on every category.
2. Variant pickers such as size and color use native controls with linked labels.
3. The purchase step asks for confirmation. web.dev notes that WebMCP tools can request user interaction before a sensitive action such as a purchase.

Google's AI optimization guide ↗︎ adds that Merchant Center feeds and Google Business Profiles help products and services appear in AI responses. Feeds and clean product pages do different jobs, and both count.

What to do this week

1. Decide which job you are buying or building. Write it in one line: agent as staff, or agent as visitor.
2. Write the permission sheet for one agent task. Default to read-only, require approval for writes, and set the retry-once rule and a run limit.
3. Name the person who approves anything client-facing, and put the name on the sheet.
4. Check robots.txt against each AI bot in the table, and check your firewall for the same bots.
5. Test your top product or lead-form page for stable layout and native <button> and <a> elements.

The agents will change. The permission sheet and the semantic HTML will not.

FAQ

Is agentic SEO the same as GEO or AI SEO?

Not quite. GEO and AI search optimization cover job two, being found and cited by AI systems. Agentic SEO also covers job one, using agents to do SEO work. Google's AI optimization guide says AEO and GEO are still SEO from Google Search's perspective.

Do AI agents replace an SEO team?

The sources do not say so. Ahrefs says agentic SEO is not fully autonomous and needs a human in the loop for anything that matters, especially client-facing work. SEO.com's definition stresses autonomy, but its page also warns against over-reliance. Both Ahrefs and SEO.com keep a person in charge.

Do I need llms.txt?

Not for Google Search. Google's guide says you do not need new machine-readable files, AI text files, markup, or Markdown to appear there, and that Google Search ignores llms.txt. Other services may read the file, so publishing one is optional.