ChatGPT Drove Just 0.62% of Google Organic Sessions, Analyst Finds

By Tom Whitfield

ChatGPT sessions amounted to 0.62% of Google organic sessions across a sample of 100 sites, according to analysis by Marie Haynes, reported in the essay "In GEO is a distraction." Her conversion data also favors Google on total volume. The numbers describe only the properties examined. The gap they reveal matters more than the totals.

The attribution problem starts with a familiar journey. A buyer asks an AI assistant which supplier fits a problem, discusses alternatives, then opens Google and searches for a company name. The buyer visits and purchases. Google receives the credit. The assistant may never have sent a visit at all.

A report about arrivals cannot settle what influenced the people arriving.

What does a Google visit actually mean?

The essay's author argues that someone searching for a company could be starting an investigation — or following a recommendation from an earlier conversation with an assistant. A buyer might use an assistant to understand a problem before knowing which product category they need.

The answer could influence the decision without sending anyone to a website. An assistant could introduce a company, or describe a rival so poorly the buyer rules it out.

This justifies investigation. It does not produce a revenue estimate. The author states plainly: the attribution gap "doesn't give us a revenue estimate, however plausible the journey sounds. I'd want to see that evidence before putting a revenue number in the proposal."

Why parametric memory defines GEO

The author rejects new labels for work SEO already covers, and notes Google itself says established SEO practices remain relevant for its AI search features, with no special optimization required. Calling familiar SEO work GEO gives it a new label without changing the work.

The distinction that makes GEO useful is parametric memory: the knowledge a model has learned through training, held in its numerical weights. Influencing what a model learns about a brand is the separate job the author means by GEO. SEO addresses retrieval — making sources discoverable, accessible and suitable.

The working shorthand: AI visibility = SEO + GEO. Instructions and user context also shape responses, so the shorthand does not describe everything.

Can learned associations steer the search?

Evidence exists. In "What Drives AI Citations?," researchers compared brands recommended without search against brands in system-generated queries. In the OpenAI API configuration tested, brands recommended without search appeared in queries in 67% of cases, against 11% otherwise.

Google describes query fan-out as issuing related searches across subtopics and data sources. A model's learned associations may influence where those searches start. Memory could give a brand a head start before any visit appears in analytics.

The same study observed the system picking up niche-vendor names from an initial round of search results.

What training experiments show

The Outlandish paper demonstrates one mechanism. Researchers inserted texts into controlled training runs and found learning an association could raise a keyword's probability elsewhere — "spillover priming." Training on text produced substantially more of this effect than placing text in a prompt.

Repetition improves recitation, but reciting a passage does not prove the model can use the facts in a different question. In biography experiments on smaller models trained from scratch, changing how the same facts were presented improved factual answering.

Open-weight models make this testable. Tools such as Pythia provide models saved at different training stages. The caveat is firm: findings from open-weight models cannot be assumed to carry over to closed production systems from OpenAI, Anthropic or Google. Outcomes are statistical, and a few encouraging responses prove nothing.

The author works at geoSurge, which has run this research for two years — part of why the possibility of influencing learned associations is taken seriously.

When content is the product

Nati Elimelech posted publicly on LinkedIn about recipe-site economics: Search traffic was rising after a refactor, yet the business model was in question even while SEO worked. For a service business, a recommendation still leads to a hire. A recipe consumed elsewhere loses both the visit and the chance to build a reader.

The author expects websites to evolve toward fewer pages, with more business transacted through API endpoints and MCP connections. If visits remain the default KPI, marketers will judge interactions by a visit that never needed to happen.

A publisher could remain useful as a source while losing the visit that funded the work. Blocking alone does not solve that business problem.

What a mention is worth

A mention could be a recommendation, a warning or a correction. A dashboard could color all three green. A warning is a strange thing to count as a win.

Buyer research and controlled comparisons can test whether an intervention improves outcomes. Neither rising branded searches nor a nicer answer proves incremental revenue. In the opening journey, the supplier earns the sale and Google helps the buyer arrive — but the assistant decided whose name went into the search box.

Google gets the thank-you. The question for marketers is who made the introduction, and whether the attribution stack will ever be built to record it.