Agentic SEO is search work done by an AI agent that plans the steps, pulls your data, makes or drafts the changes and checks the result, stopping for a person wherever you have told it to. You give it a goal, not a prompt.
The same name also covers a second job: making a site that AI agents, browsing for a buyer, can read and act on. The first changes how SEO gets done. The second changes who the site is built for.
What agentic SEO means in practice
An SEO agent works in a loop. It takes a goal, such as finding out why a group of pages lost clicks, decides what to check, runs the checks with the tools it has been given, reads the results and picks the next step until the job is done.
The loop is what makes it agentic, not the model. The model behind a chatbot becomes an agent once it can call tools, read what comes back and choose what to do next without a person typing each step.
Take one task, explaining a fall in organic clicks. The four ways of doing it differ mainly in who decides and who checks.
| Way of working | Who picks the steps | Who checks the output | When something looks wrong |
|---|---|---|---|
| Manual | You | You | You dig further |
| AI-assisted | You, one prompt at a time | You | You ask again |
| Automated workflow | Whoever built it, in advance | Nobody, unless a rule fails | It carries on, or stops |
| Agentic | The agent, toward your goal | The agent, then you | It investigates, within its permissions |
The last row is the useful one and the risky one. An agent that can investigate on its own can also be wrong on its own, faster than anyone can review, which is why what it may do alone matters more than which model runs it.
What an agent needs before it touches a site
An agent needs four things to do SEO work, and none of them is a subscription to an SEO tool.
- An environment that can act. A coding agent can read files, run commands and edit a site’s source, which a chat window cannot. Claude Code and OpenAI’s Codex are the best known, and one leans toward a session you steer, the other toward handing work off.
- Connections to your data. Agents reach other services through the Model Context Protocol, an open standard with over 10,000 published servers when the Linux Foundation took it on in December 2025. Connect Search Console, Bing Webmaster Tools, analytics and a browser first.
- A rulebook. A plain text file the agent reads at the start of every session: what the site is, how work gets checked, and what it must never do, such as publish a figure nobody has read on its source. AGENTS.md, a shared format for it, was in use in more than 60,000 projects by then.
- Skills and a schedule. An agentic SEO skill is a written procedure the agent loads when a task needs it, such as your checks before publishing. A schedule runs it unasked; in Claude Code, a recurring task started in a session expires after seven days, so use a cloud or desktop scheduled task.
Workflows worth handing to an agent
The best first jobs are ones a person does weekly, does badly when busy, and can check in minutes. Each one below runs on free data: Search Console, Bing Webmaster Tools, server logs and the site’s own source.
Pages losing clicks
Each week the agent compares the last 28 days in Search Console with the 28 before, lists the pages and queries that fell, and checks the usual causes: fewer impressions, a lower position, another page of yours competing, or a change to the page itself.
It returns a short list with the likely cause beside each page. Deciding what to rewrite stays with a person.
Written as a skill, the whole procedure fits on one screen. A trimmed example, saved as SKILL.md in a folder of its own:
---
name: weekly-click-check
description: Find pages that lost
organic clicks this week, and the
likely cause. Use for the Monday
Search Console check.
---
- From Search Console, pull clicks,
impressions and position by page
and query: the last 28 days and
the 28 before.
- Keep pages that lost at least 20%
of their clicks and 20 clicks.
- For each, check impressions,
position, which of our pages now
ranks for its top queries, and its
edit history.
- Write one line per page: the page,
the loss, the likely cause.
- Change nothing. Save the report as
reports/clicks-[date].md.A check after every release
Releases break SEO quietly. A template drops the canonical tag, a staging noindex reaches the live site, a redirect loops. After each deploy, the agent fetches the key pages, compares titles, canonicals, robots directives, status codes and structured data with the last good copy, and flags every difference.
Internal links and orphaned pages
The agent reads every page, maps which ones link where, and lists the pages nothing links to and the pairs that should link but do not. It can draft the sentences, but a link earns its place on a phrase that teaches something, so a person reads each one.
Pages competing for one query
When two of your pages trade places for the same query in Search Console, Google is unsure which one answers it. The agent finds those pairs and proposes which page to keep, merge or redirect. The call stays with a person, because a merge is hard to undo.
Figures that have aged
Statistics go stale faster than pages. The agent lists every figure on the site with its source and date, flags those older than your limit and finds the newer edition, which a person then reads on the source page before anything changes.
AI answers that cite you
An agent can put a fixed set of buyer questions to AI engines every day and record who gets cited. Because AI citations shift from one run to the next even when nothing has changed, it should report a rate across many runs, never one answer.
Where Google draws the line for agents
Google’s rules judge the result, not what produced it. The one agents hit first is scaled content abuse: many pages made to move rankings rather than help people. Its spam policies, updated in August 2026, give many AI-made pages that add no value for users as an example.
A page for every variation of a question costs an agent nothing. Google’s guide to optimizing for generative AI features says such pages, including ones aimed at the related searches its AI runs, break that policy when made to sway rankings or AI answers.
Since 15 May 2026, Google’s documentation has said its spam policies apply to its generative AI answers, such as AI Overviews and AI Mode, as well as to ordinary results. The same guide says optimizing for those AI features is “still SEO”.
Google’s guidance on generative AI content, updated in October 2026, adds the rule agents need most: a person should fact-check everything AI writes before it is published, including titles, meta descriptions, structured data and alt text.
Metadata is where agents work fastest. A hundred rewritten titles are a hundred claims nobody has read yet.
How much to let an agent do alone
Set an agent’s permissions by how easily each action can be undone, not by how sure it sounds. A wrong report costs a few minutes. A wrong redirect, a dropped page or a published claim costs traffic and trust, and both are slow to win back.
| What it may do | Examples | Who signs off |
|---|---|---|
| Read and report | Pages losing clicks, release checks, citation logs | Nobody; a person reads the report |
| Draft for review | Link sentences, refreshed figures, new titles | A person, change by change |
| Act after approval | Merges, redirects, template edits | A person, after reading the plan |
| Act alone, with a log | Repointing a link to a page that moved | A person, reading the log weekly |
Then make the hard limits mechanical. Claude Code’s documentation says its hooks give “deterministic control”: they run at fixed points every time, rather than when the model chooses to. A hook can block a deploy until the release check has passed.
Start every new workflow on the first row, and let it earn the next with a few weeks of correct reports.
The other agentic SEO: sites agents can use
The second meaning, also called agentic search optimization or SEO for AI agents, is about visitors rather than workers. A browser agent opens your site on someone’s behalf, reads it and acts: it compares prices, fills in a form, books a call.
Google named its own in March 2026. Its crawler documentation lists Google-Agent, used by agents on Google’s infrastructure to navigate the web and take actions when a user asks. Because a person triggered the visit, Google says fetchers like this generally ignore robots.txt.
Google is also testing Web Bot Auth, documented in May 2026, so that a site can verify such a visit really comes from Google.
How agents read a page matters as much as whether they arrive. Google’s web.dev guide to agent-friendly websites, updated in April 2026, says they use screenshots, raw HTML and the accessibility tree, the browser’s summary of every button, link and field.
Its advice is plain web practice: real buttons and links instead of styled boxes, and layouts that do not shift from page to page. The guide says everything it suggests also makes a site better for people.
Lighthouse now tests for this. Its Agentic Browsing category, new in Chrome 150, checks among other things that every interactive element has a name. Chrome’s June 2026 announcement calls it “informational and unbenchmarked”: checks passed, not a score out of 100.
It also checks for an llms.txt file, though Google Search ignores llms.txt and says keeping one neither helps nor harms.
For a B2B site, the visitors that matter are buyers. Forrester’s 2027 predictions, published on 21 October 2026, expect AI agents to influence 75% of B2B purchase decisions, and two-thirds of providers to lose deals because buyers cannot verify their claims.
Both are forecasts, and Google’s own guide files agentic experiences under things to explore if they suit your business and you have time. The work costs little either way: pricing, integrations and proof on plain pages, where a buyer’s AI tools can read and compare them.
Where agentic SEO goes wrong
- Publishing what it invented. Agents produce plausible figures and sources with the same confidence as real ones. Nothing an agent states as fact goes live until a person has read it on the source.
- Volume because it is cheap. A page for every town, keyword or competitor costs an agent minutes, and it is the pattern Google’s spam policies describe.
- Marking its own homework. An agent that checks a page it wrote will usually pass it. The check should be something it did not write: a hook, a script or a person.
- Reports nobody reads. A loop that lands in an unread folder every morning is worse than none, because everyone assumes someone is watching.
- Too many workflows at once. Start five in a week and you cannot tell which one is wrong.
How to start agentic SEO this week
Pick one workflow, usually the weekly check on pages losing clicks, and run it read-only for a fortnight beside the person who does the job now.
Compare them each week: what the agent found that the person missed, what it got wrong, and how long each took. Once its reports have been right for a month, let it draft, then give it the next workflow.
Log every change an agent makes, with the date and the reason. When rankings move, that log is the first place to look.
● About the author
Hamza Hashim is the marketing strategist at Katama. He owns the channel mix and the order it gets built in: what ships first, what waits for evidence, and what gets stopped when the evidence says stop. Meet the team.