SEO AI agents: what an agent can do for a small site, and where it goes wrong

A chat assistant answers a question. An agent takes a goal and works towards it in steps, calling tools as it goes. For SEO those tools are usually four: a way to read search data, a way to read and edit your site's files, a browser or search tool to look at the results page, and a way to run commands such as a build.

The coding agent you already build with is an SEO agent the moment you give it those tools and a clear job. You do not need a separate product. What makes it useful is not the model: it is the data you hand it and the rules you set.

• Data: a Search Console export, a connection to Search Console, or a report someone wrote about your site.
• Files: the repository, so it can change a title, a heading, a paragraph or a link and show you the diff.
• The results page: what already ranks for a search, and what people ask around it.
• Rules: one page per topic, no invented numbers, small changes you can measure.

Run the agent through one loop at a time, for one page or one search. Each pass ends with a change you can date and judge four weeks later.

1. 1

   Research real demand, in the searcher's words

   Ask the agent to collect what people actually ask about your problem: the questions shown on the results page, forum threads, reviews of the tools you replace. It returns the phrasings, not a keyword list. Example: "export invoices to accounting without a plugin", not "invoice export tool".

2. 2

   Decide the page type for each search

   Look at what ranks now. If the top results are step-by-step guides, you need a guide. If they are comparison pages, a feature page will not rank. The agent can read the results page and say which type it is; you confirm.

3. 3

   Check that the page does not already exist

   Before writing anything, the agent searches your own site for a page on the same topic. If one exists, it improves that page rather than creating a second one that competes with it.

4. 4

   Write or improve one page

   The agent drafts the change in your repository: the answer in the first paragraph, the steps, the example, a clear title and description. You read the diff and correct anything it could not know.

5. 5

   Publish and note the date

   Ship the change on its own, not bundled with ten others. Keep a line in a changelog file: the date, the page, the search it targets, what changed.

6. 6

   Read Search Console after four weeks

   Search Console data lags about two to three days. Compare the 28 days after the change with the 28 days before, on that page's queries only.

7. 7

   Improve what is close before writing more

   Pages sitting just off page one, and pages seen often but rarely clicked, come first. A new page starts from nothing; a page at position 11 needs a better answer and a few links.

The last step is where most of the gain is on a small site, and where agents are least likely to go on their own. Say it in the task, every time.

An agent does what it is asked, quickly and at volume. That is exactly what makes these four failures easy.

• Hundreds of pages from a keyword list. Google's spam policies, since the March 2024 update, define scaled content abuse as generating many pages primarily to manipulate rankings and not to help users, whether they are produced by automation, humans or both. A page per city or per keyword variant, with the same text and a word swapped, is the classic case.
• Invented figures and features. Asked for a comparison, a model will fill gaps with plausible numbers, prices and capabilities. On your own site that becomes a false claim under your name.
• Rewriting pages that were working. "Optimise this page" can turn a page that ranked into one that does not. The agent has no memory of why the old wording won unless you give it the page's queries.
• Chasing positions instead of clicks. A page moving from 30 to 18 is still on page two. Judge by clicks on the page's queries, not by a position chart.

For the long version of the policy, read the guide on AI content and Google and the glossary entry on scaled content abuse.

Split each step of the loop by how much damage a wrong answer does. Automate what is cheap to check. Review what reaches the public. Keep the decisions that need knowledge only you have.

An agent without your figures guesses. There are three ways to give it the figures, from simplest to most connected.

• An export. Download the Queries and Pages tables from the Search Console performance report as CSV and drop them in the repository or the chat. Simple, dated, and the agent cannot wander beyond it. The guide on analysing Search Console with AI has the prompts that suit each table.
• A connection through the Model Context Protocol. MCP is an open protocol, introduced by Anthropic in November 2024, for connecting assistants to tools and data sources. With a Search Console server connected, the agent can ask for a page's queries itself when it needs them.
• A written report. A document that already states the findings with their figures, which the agent reads and works through. Useful when you want the diagnosis done once and the agent to do the edits.

Whichever you choose, tell the agent where its figures come from and forbid any others. The exports have limits of their own: the performance report UI caps the rows in each table and never gives page and query together in one table.

Put the rules where the agent reads them at the start of every session: CLAUDE.md for Claude Code, or your tool's equivalent. Rules in a file hold across sessions; rules said once in chat are forgotten.

## SEO rules for this repository

Data
- Use only figures from the files in /seo-data (Search Console exports, dated in the file name).
- Never state a search volume, position, click count, price or feature that is not in those files or in this repository.
- If a figure is missing, say so and stop. Do not estimate.

Pages
- One page per topic. Before creating a page, search /content and /app for an existing page on the same topic. If one exists, improve it instead.
- Never generate pages from a keyword list or a template with one word swapped.
- Do not rewrite a page that gets clicks unless the task names that page and its queries.
- The answer to the search goes in the first paragraph.

Changes
- One page per change. Keep each diff small enough to judge on its own.
- Never change robots.txt, redirects, canonical tags or noindex without asking.
- After each change, append one line to /seo-data/CHANGELOG.md:
  YYYY-MM-DD | /path | target search | what changed

Measurement
- Judge a change by clicks on the page's queries, 28 days after against the 28 days before.
- Improve pages close to page one before proposing new pages.

Here is a task that fits the loop. The site is yourproduct.com, a tool that sends invoices; the export shows the page /guides/recurring-invoices at an average position of 9 for "how to set up recurring invoices", with many impressions and few clicks.

Task: improve /guides/recurring-invoices for the search "how to set up recurring invoices".

1. Read /seo-data/queries-pages-2026-09.csv and list the queries this page is shown for, with clicks, impressions and position. Use only those figures.
2. Look at the current results page for the search. Name the page type that ranks and the questions shown there.
3. Read the page. Tell me what the first paragraph answers and what it does not.
4. Propose a new title (about 60 characters), a new description (about 155 characters) and a first paragraph that answers the search directly.
5. Add the missing steps or questions as sections, using only features that exist in this repository.
6. Show me the diff. Do not publish.
7. When I approve, add the line to /seo-data/CHANGELOG.md.

Before, the page opened with "Billing made simple for growing teams." After, it opens with "To set up recurring invoices, open a client, choose Repeat, pick the interval and the first date." That is the kind of change an agent does well and you can judge in four weeks.