ChatGPT Drove 0.62% of Google Organic Sessions, and the Rest Went Uncredited

ChatGPT sessions amounted to 0.62% of Google organic sessions across a sample of 100 sites, according to data reported by Marie Haynes and discussed by Pedro Dias in "GEO Is a Distraction." The figure anchors a wider argument: analytics credit the last visit, not the conversation that produced it. Dias, who works at geoSurge, a company researching this field for roughly two years, challenges what referral numbers can and cannot establish.

What does the 0.62% figure actually prove?

Haynes measured ChatGPT against Google across 100 sites. Her conversion analysis also favors Google on total volume. Dias calls these useful observations about the properties she examined. He states plainly that he has not inspected the underlying GA4 data.

The gap matters. A buyer can ask an assistant to shortlist suppliers, discuss alternatives, then open Google and search for one company by name. They visit. They buy. Google receives the credit. The assistant never sent a visit.

Dias writes: "A report about arrivals cannot, by itself, settle what influenced the people arriving." He agrees with Haynes on two points: transactions can happen without a site visit, and weak evidence should not ground a new discipline. He disagrees on how far referral figures carry us.

Takeaway: arrival data measures arrivals, not influence.

Where does GEO separate from SEO?

Dias has been skeptical of new initialisms for work SEO already covers. His working distinction is parametric memory: the knowledge and associations a model learned during training, stored in its weights. Influencing what a model learns about a brand, and how it applies that knowledge, is the job he calls GEO.

Retrieval is different. In retrieval-augmented generation, the model consults external material while answering, alongside what it already knows. Dias offers an analogy: a specialist colleague who brings learned knowledge and can also open a reference folder. Changing the folder differs from changing what the colleague knows.

He notes a caveat. A model's weights do not work like human memory, and its answer does not identify which contribution came from which source. Google itself says established SEO practices remain relevant for its AI search features and that no special optimization is required.

His shorthand: AI visibility = SEO + GEO, with instructions and user context also shaping responses.

Can memory shape the search itself?

A study cited by Dias, "What Drives AI Citations?," compared brands recommended without search against brands appearing in the system's generated queries. In the OpenAI API configuration tested, brands recommended without search appeared in generated queries in 67% of cases, against 11% otherwise.

Dias flags the limit. The comparison shows what the model recommends under those conditions, not whether changing training would change the searches. That would require a different experiment.

The same study observed the system picking up niche-vendor names from an initial round of search results. Instructions and retrieved material can also shape subsequent queries. His interest is whether learned associations change the searches a system makes, before any visit appears in analytics.

Can you study the learning directly?

Dias points to controlled experiments. In one paper, researchers inserted texts into training runs and found that learning an association raised a keyword's probability in other contexts, especially within the same theme. They called this spillover priming. Putting text in a prompt produced substantially less of this effect than training on it.

He adds a caution: a keyword appearing in an irrelevant answer could make that answer worse. He would read the answer before celebrating the mention.

Repetition raises another question. It can improve a model's ability to recite a passage, but recitation does not prove the model can use the facts in a different question. In biography experiments, researchers trained smaller models from scratch and found that changing how facts were presented during training improved factual answering.

Open-weight models make this testable. Researchers can download models, change training material, and compare versions. Resources such as Pythia provide models saved at different training stages, with code and data.

Dias lists the limits:

• A business publishing pages does not control whether a provider trains on them.
• Seeing a page crawled or cited does not establish that the model learned something useful.
• An uncited reply alone does not establish a change in weights.
• Findings from open-weight models may not carry over to closed production systems from OpenAI, Anthropic or Google.
• Outcomes are statistical; a few encouraging responses prove little.

He also discloses his commercial stake. He works at geoSurge, which has run this research for about two years. That experience, he writes, is part of why he takes the possibility of influencing learned associations seriously.

Takeaway: the experiments establish effects under test conditions, not a publishing recipe for recommendations.

What happens when the content is the product?

Dias cites a public LinkedIn post by Nati Elimelech on recipe-site economics. Search traffic was rising after a successful refactor, yet Elimelech questioned the business model while the SEO was working.

For a service business, an assistant's recommendation can still lead to a hire. A recipe can be fully consumed inside an answer. The publisher pays for production and loses both the visit and the chance to build a returning reader.

Dias would not dismiss blocking as pointless, but doubts blocking alone solves the longer-term problem. He expects many sites to evolve toward fewer pages, with more business exposed through API endpoints and MCP connections that let assistants use data and services directly. If visits remain the default KPI, those interactions would be judged by a visit that never needed to happen.

His hard question: what gives someone a continuing reason to depend on the original source after consuming one answer?

What should a mention be worth?

A dashboard can color every brand mention green. Dias notes a mention could be a recommendation, a warning or a correction. A warning about your company is a strange entry for the wins column.

Buyer research, self-reported discovery and controlled comparisons can investigate journeys that referral reports miss. Neither a rise in branded searches nor a nicer answer, taken alone, proves incremental revenue.

His closing frame stays sharp. In the journey he describes, the supplier earns the sale and Google helps the buyer arrive, but the assistant decided whose name entered the search box. Google gets the thank-you. Dias wants to know who made the introduction.

Monitor next: whether controlled experiments on learned associations translate into measurable changes in generated queries, and whether MCP-style access gives publishers any business case beyond the visit.