GPT-4 and SEO: looking back at how the work changed
In SEO work, GPT-4 takes over routine work: preparing data, grouping queries, drafting. Checking and deciding stays with the editors. A view from October 2023.
GPT-4 takes busywork off SEO teams: preparing data, grouping search queries, drafting. The decision stays with the editors, and they check every result before it moves on. That's where I stand in October 2023, a little over half a year after GPT-4 came out, from the seat of an agency that is just starting its shift to AI.
GPT-4 is a language model from OpenAI. It writes text and analyzes text, and you can use it in the chat through ChatGPT Plus or through an API. The task should decide which model handles it. For simple rewording, GPT-3.5 is often enough. For analysis over several steps, GPT-4 is more reliable.
Analysis can be prepared
GPT-4 can explain findings from a crawl and suggest groups of search queries. Before that comes a separate step: preparing the data. The model can only handle a limited amount of text per request, the so-called context window. A full crawl export of a larger website won't fit, and neither will a large export from Google Search Console. So the data gets cleaned, cut down and sorted before it goes into the model. That prep work shapes what comes out at the end.
What comes out is a working hypothesis. Whether a priority is right only shows once you check it against the website and the business goal behind it. Until then, even a plausible suggestion is not yet a finding.
Editing remains its own process
Every draft needs a brief, a source check and sign-off from someone who knows the subject. GPT-4 can hallucinate, which means it makes up sources, numbers and facts that sound plausible. For SEO content that calls for a simple rule: every source and every number gets checked against the original before sign-off.
In February 2023 Google made clear that it rates the quality of content, not whether AI helped write it. For search, that settles the question of AI or no AI and leaves the question of quality. How many words a model produced says little about that. What counts is whether the text answers a concrete search question. Whoever passes on a text has to read and check it first.
One page or several
An example: GPT-4 sorts a long list of search queries by topic. Two groups look alike but reflect different reader questions. One wants a term explained, the other is looking for a provider. The model sees similar words; the editors see the intent behind them.
So the editors check the intent behind each group. If the same reader question sits behind both, one page is enough. If there are two questions, it takes two pages. Before that, they look at what the website already covers. A new page is only worth it if it adds something of its own next to the existing ones, with its own take and checked facts.
That is also what E-E-A-T aims at, Google's standard for experience, expertise, authoritativeness and trustworthiness. Reworded generic content doesn't meet it, whether a person or a model wrote it. Grouping saves time. The decision stays with the editors.
The model sees similar words; the editors see the intent behind them.
What GPT-4 means for SEO work
For SUMAX this is a task for the whole company. This year we started moving how we work, our roles, services and rules onto AI together with the team, and I'm at the beginning. The foundation for everyone comes first: users, permissions and access, with security built in from the start. After that come tools for individual tasks, built with the people who use them.
The editorial team's process knowledge belongs in one place, not in heads and Excel lists. Only a workflow that is written down can be checked and improved with AI support. In SEO work, that gives the editors time for the decisions a model can't make.