Client onboarding is repetitive. Every new SEO client asks some version of the same questions: how long until we see results, why does my ranking move every week, what does this metric mean, should we use AI for content. A lot of that gets answered live, on calls, over and over.
A custom GPT is an obvious fit for that. Below is how I'd structure a set for SEO onboarding, what each would contain, and how to keep them from confidently making things up in front of a client, which most guides skip.
The concept
A custom GPT is ChatGPT with specific instructions and files attached.
You can:
- Write instructions (a prompt that defines how the GPT behaves)
- Upload files (methodology docs, FAQs, anonymized examples)
- Control who can access it
The set I'd build
GPT 1: SEO basics for new clients
Instructions: "You are an SEO expert. Explain SEO concepts to a smart but non-technical person. Use examples from the client's industry. If the answer isn't in your files, say so and suggest asking the account lead."
Files: the agency's SEO methodology, a glossary, a common-questions document.
GPT 2: Content strategy
Instructions: "You are a content strategist. Help clients understand how content gets planned, briefed, reviewed and published, and why."
Files: content planning templates, the editorial standards, a description of the review process.
GPT 3: Reporting and metrics explained
Instructions: "You explain SEO metrics and what they do and don't mean. Use only the anonymized examples in your files. Never cite another client's numbers."
Files: metric definitions, a sample report with dummy data, common misreadings (impressions vs. clicks, average position, why month lengths matter).
GPT 4: Link building
Instructions: "You explain how modern link building works, how placements are vetted, and why every placement needs to be defensible."
Files: the vetting criteria, the anchor-text approach, examples of placements that would be rejected and why.
GPT 5: Client-specific
Instructions: "You are [Client]'s SEO reference. Answer questions about their strategy using only their documents."
Files: their strategy summary, competitor overview, content calendar and past reports. One GPT per client, never shared across clients.
What a good answer looks like
Take the most common question: "Why does my ranking for [keyword] go up and down every week?"
A good GPT answer: "Rankings fluctuate. Google shows different results to different people in different places, and rank trackers approximate your position from a sample. Week-to-week swings are usually noise. Real gains or losses tend to show up as a trend over several weeks, and clicks in Search Console are a better signal than a single tracked position."
That's generic and correct, which is what you want from a GPT. The dangerous answers are the specific ones. Ask "should I hire a writer or use AI?" and an under-grounded GPT will happily invent a cost per post and an expected ranking position for each option. It sounds authoritative and it's made up.
Grounding rules
I spend a lot of my time on systems where every claim has to trace back to a source. The content engine I designed at work keeps a claim ledger, so every factual statement is tied to a source before a human sees the draft. Applied to client GPTs, that means:
- Answer from the files or say "I don't know." Put it in the instructions, and test that it actually does it.
- No invented numbers. If a figure isn't in a file, the GPT shouldn't produce one. Prices, timelines and ranking predictions are where hallucinations do the most damage.
- Never mix clients. Shared GPTs get anonymized examples only. Client-specific GPTs get one client's data.
- Test it adversarially before a client sees it. Ask the questions you'd least like it to improvise on, and see whether it holds the line.
What GPTs do well
- Repeat explanations: if the question is covered in the documents, a well-instructed GPT handles it consistently
- Scaling the depth: simple questions get simple answers, complex ones get detail
- Being available: clients can ask at 10pm without waiting for a call
What they do poorly
- Novel questions: if it isn't documented, the GPT may hallucinate or go generic
- Strategic decisions: "what should we do with our budget?" needs someone who understands the business
- Diagnosis: if rankings drop or traffic falls, the GPT can't investigate. That needs a person with access to the data.
So treat a GPT as a filter: it handles the repeat questions that don't need judgment and routes the rest to a person.
How to build one
1. Notice which question you answer for every new client 2. Write the answer down properly, as a short methodology document 3. Create a GPT with that document and strict grounding instructions 4. Test it yourself, especially on questions it shouldn't answer 5. Share it with one client and read the conversations 6. Only then add the next one
Building custom GPTs requires a paid ChatGPT plan. Check current plan terms for who can use a GPT you share.
Should you do this
If you're explaining the same things to every client, it's worth trying. Document your methodology, ground the GPT in it, and test it hard before anyone outside your team sees it.
Most of the value is in the documentation. A good methodology document makes onboarding better with or without a GPT on top of it.