SEJ writer says GEO's real job is what a model has learned

A buyer asks ChatGPT what supplier fits a problem. ChatGPT recommends your firm. The buyer talks it through, opens Google, searches your company name, and buys. Your analytics give Google the credit. That is the scene Pedro Dias uses to open an essay in Search Engine Journal this week. His argument is that GEO's real work is upstream of any visit.

Dias is a writer at SEJ and works at geoSurge, a firm doing research on AI search. He wrote the piece as a reply to Marie Haynes. Her own SEJ piece last week argued that GEO is a distraction. Her numbers are quoted in his piece. His own figures come from a published paper.

The number from the paper

Dias cites a paper called What Drives AI Citations. The researchers compared two conditions. First, they asked a model to recommend brands with no search available. Second, they let the system generate its own queries. The setup used the OpenAI API. Brands recommended without search then appeared inside the generated queries 67 percent of the time. Brands not in the recommendation set appeared 11 percent of the time.

The point Dias makes with that gap is simple. What the model has already learned about a brand shapes the queries it sends to Search. A brand that lives in the model's memory gets a head start. One that does not can be missed before the first search even runs.

What parametric memory is

Dias defines two things a reader needs. Parametric memory is what the model learned during training. It sits in the weights. Retrieval is what the model fetches at answer time. He calls the mix AI visibility, and his shorthand is SEO plus GEO. SEO works on the pages that get retrieved. GEO works on what the model learned.

He points at a second paper, Outlandish, as evidence that training changes associations. Putting a text into training raised the keyword's probability in related answers. Putting the same text in the prompt had a smaller effect. He is careful about the limit. Those tests used open-weight models under controlled conditions. He writes that whether the same effect holds inside OpenAI, Anthropic or Google's production models needs separate evidence.

What this changes for a brand team

Dias says he does not reject Marie Haynes's numbers. He says a referral figure describes who sent the arriving visit. It is silent on who shaped the earlier decision. If an assistant introduces a shortlist and the buyer searches your name on Google, your analytics will record Google. The assistant leaves no trace in that report.

His test for a brand team is to study the learning. The suggestion is to put research resources into how a model is taught. Open-weight model research gives you a way to run controlled comparisons on how a brand is taught. He warns that counting an appearance and explaining it need different evidence. A tool that reports a mention does not tell you what the model said. The model may have recommended, warned about, or corrected a reader on your firm.

Source

Google Gets The Thank-You, Pedro Dias, Search Engine Journal.