Custom SEO agent: what it is and how to build one
Short answer A custom SEO agent is a chatbot you set up once with your own instructions, reference files and a fixed report format, so every page review follows the same checklist. Build it as a ChatGPT project or a Claude project, not a custom GPT: OpenAI retires custom GPTs on December 11, 2026. Check its facts before you act.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callWhat a custom SEO agent is
It is a chatbot you set up once so it reviews pages the same way every time. Three parts make it "custom":
| Part | What it is | Example |
|---|---|---|
| Instructions | The job, the checklist and the report format, written once | "Review one service page. Score 6 checks. List the top 3 fixes." |
| Reference files | What it should know about you and about Google's rules | Your service list, towns served, Google's guidance pages saved as PDFs |
| Output format | The same table every time | Check, score, evidence, fix |
Built this way, it is not autonomous. It doesn't log into your site, watch rankings or fix anything. You paste a page, it reports, you decide. That is still useful: a consistent second reader for every page.
A true agent, one that runs on a schedule, pulls data and makes changes, needs code or an automation tool connected to data sources. That is a separate, bigger project (covered below).
Build it as a project, not a custom GPT
Don't start a custom GPT now. OpenAI announced the end of them:
"Custom GPTs are scheduled to retire on Dec 11, 2026." (OpenAI Help Center: Custom GPT retirement and migration FAQ)
The same page points people who used a GPT for ongoing work to projects: "Projects keep related chats, files, and instructions together." OpenAI is moving GPT creators to plugins, where a GPT's instructions become a "skill". For one owner reviewing their own pages, a project does the job with less setup.
In ChatGPT (per OpenAI Help Center: Projects in ChatGPT):
- Select New project in the sidebar and name it, for example "SEO page review".
- Open the project menu (•••), then Project settings, and paste your instructions (next section).
- Upload your reference files. "The number of files you can upload depends on your plan."
- Start each review as a new chat inside the project.
OpenAI says project instructions "apply only within that project and override your global custom instructions", so your SEO rules won't leak into other chats. Projects are available "subject to plan availability and workspace settings".
In Claude, projects work the same way: a knowledge base plus project instructions. Claude's help center says: "Projects are available to all users, including those with free Claude accounts. Free users can create a maximum of five projects."
What to put in the instructions
Write them as a short job description. Six parts cover it:
- Role and reader. "You review web pages for a [plumbing company in Calgary]. The reader is the owner, not a marketer. Plain words."
- Input. "I will paste the page text, its title tag and its meta description. Review only what I paste. If something is missing, say so; don't guess."
- Checks. The list in the next section, each tied to a Google page in your files.
- Scoring. "Score each check 1 to 5 and quote the words on the page that justify the score."
- Format. "One table: check, score, evidence, fix. Then the 3 fixes that matter most, in order."
- Limits. "Don't invent search volumes, rankings or statistics. If a fix needs data you don't have, say what data to get."
Point 6 matters most. Google warns about exactly this:
"Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data. Because of this, generative AI outputs may contain inaccuracies (also known as hallucinations)." (Google Search Central: Guidance on using generative AI content)
What should the agent check?
Base every check on what Google publishes, not on SEO folklore. Many ready-made SEO prompts carry rules Google has said aren't real. Upload these Google pages to the project and tell the agent to follow them over its own habits:
| Check | What Google says | Tell the agent |
|---|---|---|
| Does the page answer the visitor's question? | Google's helpful-content page asks whether readers "need to search again to get better information from other sources" | Name what a customer still has to ask after reading |
| Experience and trust (E-E-A-T) | "While E-E-A-T itself isn't a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful." (Creating helpful content) | Look for who did the work, real photos, licenses, reviews, contact details |
| Length | Google asks if you write to a word count because you heard it has a preferred one: "(No, we don't.)" | Never flag a page for word count alone |
| Title tag | No length limit; long titles are "truncated in Google Search results as needed, typically to fit the device width" (Influencing your title links) | Judge whether it is clear and specific, not whether it hits a character count |
| Meta description | "There's no limit on how long a meta description can be" (Control your snippets) | Check that it summarizes the page; long ones get cut, not penalized |
| AI-written text | Google says to "manually factcheck and review all AI-generated content", including titles and meta descriptions (Guidance on generative AI content) | Flag claims that need a source or an owner's check |
Add your own business files too: services with prices if you publish them, towns you serve, your licenses, and a copy of your top pages and search queries from Search Console. With that, it can say "people search for your service in Airdrie, and this page never mentions Airdrie" instead of generic advice.
Where a DIY SEO agent goes wrong
- Made-up numbers. Asked for keyword volumes or competitor rankings, a chatbot may produce a confident table with no data behind it. Take numbers only from a real data source, such as Search Console.
- Reading the wrong page. If it fetches a URL itself, it may see a cached or partial version. Pasting the text you see is the reliable input.
- Scores look precise but are opinions. A "7 out of 10" doesn't predict rankings. Use scores to compare your own pages, not as a forecast.
- Its rules age. It knows only what's in its training data and your files. When Google changes a guideline, replace the file.
- It drifts. Re-run one already-reviewed page each month. If the report changes a lot, tighten the instructions.
When you need a real automated agent
If you want it to run every week on 200 pages, pull rankings and write reports with no one pasting anything, a chat project isn't enough. You need code or an automation tool, a model's API, and data sources. Two things surprise owners:
- Data access costs money and changes. Google's own search API is closing: "The Custom Search JSON API is closed to new customers," and existing users "have until January 1, 2027 to transition to an alternative solution." (Custom Search JSON API overview) Paid SEO tools charge for their data access too.
- Someone has to maintain it. Connections break, models change, and the rules in its files need updating.
For most small businesses, the project version covers the reviewing. The building and fixing is where the time goes.
Example
A roofing company with 12 service pages
The owner sets up a ChatGPT project with the instructions above, the four Google pages and a list of towns served. Each review takes about 5 minutes per page. The agent flags that 8 pages never say who does the work or show a license number. The owner adds crew photos and the license to each page in one afternoon.
A bookkeeper in Toronto checking new blog posts
She uses a free Claude project as a pre-publish check. The agent's table flags a post with an unsourced statistic and a title that says "Tips" without saying for whom. She fixes both before posting. It never touches her site; she stays the editor.
A clinic that wanted full automation
A two-location clinic plans a scheduled agent connected to rank data and an AI model's API. The quote covers API fees, data fees and about 10 hours a month to maintain it. They keep the project version for page reviews and pay for the data side only when they need it.
FAQ
Can I still create a custom GPT for SEO?
Don't. OpenAI's FAQ says custom GPTs retire on December 11, 2026, and their pages become inaccessible then. Build in a project instead.
Do I need to code to build a custom SEO agent?
No. A ChatGPT or Claude project needs only written instructions and uploaded files. You need code only to run it on a schedule with live data.
Can a custom SEO agent improve my rankings by itself?
No. It finds problems and suggests fixes; the changes you make decide the result. Google ranks the page, not the review.
Is it safe to upload my business files?
Upload only what you'd share with a contractor. Leave out passwords, client records and financial data. Check your plan's data settings first.
Should it use Google's E-E-A-T as a score?
Use E-E-A-T as a checklist of trust signals, not a score Google calculates. Google says E-E-A-T "itself isn't a specific ranking factor".
Different for you?
- You'd rather do the review by hand first: How to do an SEO audit yourself.
- You want to know what Google means by trust and experience: E-E-A-T for a small business website.
- Your agent will also write pages, not just review them: AI content that Google likes.
- You need the search data to feed it: Google Search Console for small business.
- You'd rather not build or maintain anything: AI SEO agent alternatives or the SEO agent.