Recently I joined the SEO in 2025 podcast series hosted by David Bain, where we discussed what I believe is the most practical shift SEO professionals can make right now. The conversation centred on moving beyond one-off prompts and into building trained custom GPTs for clients — a change that saves significant research time and produces far more relevant, on-brand output. Below are the key insights from that interview, turned into a practical guide you can start applying immediately.
Watch the video: SEO in 2025 — Victoria Olsina on Custom GPTs for Clients
Why Prompts Alone Are No Longer Enough for SEO Agencies
Most SEOs are still working the same way: write a prompt, get a result, tweak it, repeat. The problem is that every new piece of content requires the same upfront research effort. You’re essentially rebuilding context from scratch each time.
A trained custom GPT changes that. Once you load a client’s data into the model — branding documents, messaging guides, their website, blog posts — the GPT carries that context permanently. You stop re-explaining the client and start producing output that already sounds like them.
The Research Problem Custom GPTs Actually Solve
One of the most underrated benefits is how quickly a trained GPT can answer questions about a client that the client themselves struggle to answer clearly.
Ask most B2B clients to list all their product features and benefits, and you’ll get a vague overview at best. Ask a GPT trained on their full website and documentation, and you get a structured, detailed answer in seconds. This is especially valuable in complex verticals like Web3, blockchain, and developer tooling — areas where the technical depth makes traditional research slow and error-prone.
A GPT trained on a client’s full content library effectively becomes a subject matter expert on that client. The SEO professional’s role shifts from researcher to strategist.
How to Build a Custom GPT That Actually Delivers
The quality of the output depends entirely on the quality of the input. This is the single most important principle. A GPT trained on one podcast interview will produce noticeably weaker output than one trained on a full website, multiple blog posts, brand documents, and historical content.
Step 1: Data Collection
Gather everything available about the client:
- Website content (all key pages)
- Blog posts and articles
- Branding and messaging documents
- Past campaign assets
- Any interviews, podcasts, or talks
The more comprehensive the input, the more reliable the output. Volume and variety both matter.
Step 2: Training the Model
Write a clear system prompt that defines what the GPT is supposed to do, how it should communicate, and what tone and style it should follow. Once the prompt is written, run it back through the model itself — ask the LLM what the prompt is missing and how it could be improved. This self-review step consistently surfaces gaps you’d otherwise miss, from language preferences (UK vs American English) to tone rules and CTA specifications.
Step 3: Customisation and Fine-Tuning
After the GPT is in active use, feedback will come in from content writers, marketers, or the client directly. The changes are usually small — sentence case preferences, word substitutions, updated CTAs or slogans. Feed these corrections back into the training material and update the instructions accordingly. Approved, finalised content should loop back into the knowledge base to reinforce what good output looks like.
Step 4: Integration with Existing Tools
Custom GPTs can be connected to automation platforms like Make.com or Zapier to slot into existing workflows. OpenAI Assistants — the API-accessible version of custom GPTs — support this kind of integration directly, making it possible to trigger content generation from other systems, pass output downstream, or build fully automated pipelines.
Step 5: Continuous Improvement
A custom GPT is not a set-and-forget tool. Improvements come from three sources: your own observations of output quality, periodic prompts back to the model asking how the instructions could be strengthened, and structured client feedback. Product name changes, rebrandings, new features — all of these should be incorporated into the instructions as they happen rather than left to accumulate.
What SEOs Should Stop Doing to Free Up Time for This
The honest answer is that long, exhaustive technical audits are a poor use of SEO time in most cases. An audit with 200 line items in a spreadsheet rarely gets fully implemented. The effort of producing it often outweighs the value delivered.
Keeping technical audits to 5-10 prioritised, actionable items forces better decision-making and increases the likelihood that the work actually gets done. The time saved can be redirected towards building and maintaining the GPT infrastructure that compounds in value over time.
Putting It Into Practice: What the Stack Looks Like
| Component | Tool | Purpose |
|---|---|---|
| Data gathering | Website scraper / manual export | Build the training corpus |
| GPT creation | ChatGPT Plus custom GPTs | Client-specific AI assistant |
| Prompt improvement | ChatGPT or Claude | Self-review and refinement loop |
| Automation | Make.com or Zapier | Workflow integration |
| API access | OpenAI Assistants | Programmatic use |
| Content review | Human writer or marketer | Final quality check |
Conclusion
Moving from prompts to trained custom GPTs is one of the highest-leverage changes an SEO professional or agency can make right now. The research bottleneck disappears. The output becomes more consistent and on-brand. And once the system is built, it gets better with every iteration rather than starting from zero each time.
Start with one client. Gather everything. Build the GPT. Let the model tell you what the prompt is missing. Then watch how much faster the work moves.
Here’s every single interview of the Majestic podcast series:
- SEO in 2023: The importance of using hybrid models for AI
- SEO in 2024: Build your own prompt libraries for your clients
- SEO in 2025: Improve the output from AI by creating custom GPTs for your clients
- SEO in 2026: Customise AI to automate your SEO processes
Frequently Asked Questions
What is a custom GPT and how does it differ from a standard ChatGPT prompt?
A custom GPT is a version of ChatGPT trained on specific documents, instructions, and examples you provide. Unlike a standard prompt, it retains your uploaded knowledge base across all conversations, meaning it already understands your client, their tone, and their subject matter without needing to be re-briefed each time. Custom GPTs are available to ChatGPT Plus subscribers and can be shared directly with clients or team members.
How much data does a custom GPT need to produce useful SEO content?
More data consistently produces better output. A GPT trained on a full website, blog archive, and brand documents will outperform one trained on a single document. For complex or technical clients — particularly in Web3, SaaS, or B2B sectors — comprehensive data collection is the single biggest factor in output quality. Even a GPT trained on one interview can be useful as a starting point; the key is to keep adding material as it becomes available.
Can custom GPTs be connected to content management systems?
Yes. OpenAI Assistants — the API-accessible equivalent of custom GPTs — can be connected to tools like Make.com and Zapier, which in turn can push content to WordPress, Notion, Google Docs, or most CMS platforms with an API. This makes it possible to build end-to-end content workflows where the GPT generates a draft that is automatically routed for human review and then published. Full automation is possible but a human review step is strongly recommended for client-facing content.
How do you improve a custom GPT over time?
Improvement comes from three sources: your own observations of where output quality falls short, periodic self-review by asking the model what the current instructions are missing, and structured feedback from the client or content team. Finalised, approved content should be fed back into the knowledge base to reinforce correct output. Any changes to branding, CTAs, product names, or tone preferences should be updated in the system prompt immediately rather than managed through individual prompts.
Is a custom GPT useful for smaller SEO clients, not just large ones?
Yes — the value scales regardless of client size. The core benefit is eliminating repetitive research time, which is relevant whether the client is a one-person business or a large organisation. For smaller clients, even a GPT trained on their website and a few key documents significantly speeds up content production. The setup time is low, and the ongoing time savings compound quickly, making it worthwhile for any client with recurring content needs.
What should SEOs stop doing to make time for building custom GPTs?
The most practical change is reducing the scope of technical audits. A 200-item audit spreadsheet is rarely implemented in full — research consistently shows that shorter, prioritised action lists drive better implementation rates. Keeping audits to 5-10 items frees up significant time without reducing the quality of technical recommendations. Other repeatable tasks like keyword research, content brief writing, and on-page optimisation are also strong candidates for GPT-assisted workflows, further compounding the time savings.
Need help turning ideas like this into working systems?
This is the kind of AI marketing work we do with teams that want repeatable workflows, not one-off experiments.
https://victoriaolsina.com/services/ai-marketing/
