How we run our own SEO and content with an AI agent: what it does, the guardrails and what still needs a human
We run an AI agent, which we call Autopilot, on our own site every other day. It reads a brief built from Search Console, GA4, Bing, HubSpot and an AI-written priority list, fixes crawl and link errors, rewrites titles, meta descriptions and FAQs on pages that are close to ranking, and publishes one new article for a proven search gap. It may merge only a short list of low-risk changes after the build and our QA checks pass, everything else goes to a pull request for a person to approve, and every action is logged. Strategy, approvals, outreach and sales follow-up stay with people.
Key takeaways
Why we put an agent on our own site first
Clients ask us whether AI agents can run parts of SEO and content operations. We did not want to answer from vendor demos, so we built one for planconversion.com and gave it real responsibility: it changes the live site without waiting for us, inside rules we wrote.
The goal is not more pages. It is a site that stays technically clean and keeps improving the pages and topics our buyers actually search for, with a record of what changed and why.
What the agent reads before it acts
Every run starts from one brief, served by a private admin endpoint. It is the same view we use ourselves each morning.
- Search Console: queries and pages over 28 days, with rising and falling queries, queries ranking just off page one, and pages with a click-through gap.
- GA4: sessions, sources and conversion events, including visits from AI assistants.
- Bing Webmaster Tools: crawl errors, crawl stats and top queries, refreshed daily.
- HubSpot: recent leads and deals, and which landing pages produce booked meetings and pipeline.
- A priority list written each morning by Claude from all of the above, the "Do next" list the agent works through in order.
What it does on each run
- Needs attention first: broken links, crawl errors and dead URLs, using our server logs for the exact URLs when a crawler report only gives counts.
- Optimize up to five existing pages that already rank near page one: title, meta description, FAQ answers for the exact queries people use, and internal links.
- Publish one new article for a proven gap, only if it passes our content gate: a named buyer persona can act on it, it maps to a service we sell, real search or social demand exists, it carries our point of view, it does not duplicate an existing page, and every claim is verified from a primary source.
- Verify: build the site, run automated QA, then check every changed URL live for status, canonical and title, and revert at once if anything fails.
- Log the run: every action, how each priority was handled, and anything that needs a person.
Get an AI visibility read across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews.
The guardrails that make unattended runs safe
The agent may merge only a written allow-list of changes: crawl and link fixes, redirects to a clear successor, titles and meta descriptions, FAQs, internal links and the run’s new article. New landing pages, design changes, scheduled jobs and anything about how the agent itself works go to a pull request and wait for approval. The rule that let it act on the priority list at all was itself a pull request a person approved.
Some actions are off limits whatever the data says. It never emails anyone, never changes CRM records, never invents statistics, clients or results, and never quotes pipeline numbers on the public site.
| Change | Agent merges it | Goes to a pull request |
|---|---|---|
| Crawl, link and redirect fixes | Yes | |
| Titles, meta descriptions, FAQs, internal links | Yes | |
| One new article that passes the content gate | Yes | |
| New landing pages or tools | Yes | |
| Design, scheduled jobs, the agent’s own rules | Yes | |
| Emails, CRM changes, ad account changes | Never | Never |
What it changed in its first runs
The first two runs took place on November 21 and 22, 2026. Each is recorded in a pull request on our site repository and in the admin log.
- Fixed a broken internal link on a French ABM article that pointed to a page that does not exist.
- Optimized ten pages: five in the ABM and 6sense cluster, our ABM service page, the Demandbase vs 6sense comparison, two ChatGPT Business Hub articles and an Ottawa GEO page. Changes were titles, meta descriptions, FAQs and internal links, each tied to a specific query that already ranked.
- Published two articles for gaps it found in Search Console: how to run one-to-many ABM, and switching from 6sense to Demandbase.
- Found from server logs that search crawlers were requesting English paths of French-only articles, added redirects for five of them, and marked two dead image files as gone.
We will judge these changes on click-through rate and position for the queries named in each pull request, about four weeks after each run, and on the meetings and pipeline the pages produce. It is too early to report results, so we do not.
What still needs a person
The agent is good at the repetitive, data-heavy work that slips when a team is busy. It does not replace the work that builds trust off the site: client reviews, partner listings, guest articles and relationships with editors. It also does not answer leads. We pair it with a separate alert that tells us when a strong-fit lead has waited too long for a reply, because faster follow-up does more for pipeline than another optimized title.
Strategy stays human too. We decide which services and buyers matter, approve anything structural, and review the log. The agent makes that review fast because every action is written down.
How to start with an agent on your own site
- Build the brief first: one place that combines search, analytics, crawl and CRM data. An agent without CRM outcomes optimizes for traffic.
- Write the allow-list and the never list before the first run.
- Require a build, automated QA and a live check for every change, with automatic revert.
- Start with fixes and page optimization, add new content only after a few clean runs.
- Log every action and review the log weekly for the first month.
Frequently asked questions
Can an AI agent run SEO without human review?
For a short list of low-risk changes, yes, if every change is built, checked and verified live with automatic revert. Structural changes, new page types and anything about the agent’s own rules should still go through human approval.
What data does an SEO agent need?
Search Console, analytics, crawl reports from Google and Bing, and CRM outcomes such as meetings and pipeline by landing page. Without CRM data the agent can only optimize for traffic.
How often should an SEO agent run?
We run ours every other day. That is frequent enough to fix crawl errors quickly and act on new query data, and slow enough for search engines to show the effect of the last changes.
Does an AI agent replace an SEO team?
No. It takes over repetitive analysis, fixes and on-page work. People still set strategy, approve structural changes, build off-site authority and turn visibility into pipeline.
Which AI model does your agent use?
It runs on Claude, as a scheduled Claude Code task with access to our site repository, analytics and CRM through connectors, inside the rules described in this article.
Related service: Generative engine optimization. See also pricing and client results.
We run enterprise B2B demand generation for teams in Montreal, Toronto, Ottawa and the United States.