The 40% GEO uplift did not replicate

Almost every GEO pitch deck in circulation traces back to one 2024 paper reporting up to a 40% visibility uplift from adding citations, quotations and statistics to content. It is a good paper. It has also been substantially qualified by everything that came after it, and the qualification has not reached the decks.

What the follow-up work found

A 2025 benchmark tested 54 combinations of GEO method and domain. Three were significantly positive. More awkwardly, the gains it did find were largely zero-sum: when everyone applies the technique, the advantage disappears, because the technique was never creating value — it was redistributing attention among documents.

No content-side technique has yet demonstrated a stable causal effect on organic discoverability that holds under broad adoption. That is a genuinely hard thing to sell, and pretending otherwise is how this discipline earns the reputation SEO spent fifteen years escaping.

What did survive, with strong evidence

• Retrieval position and relevance dominate. Whether a document gets cited is mostly about whether it gets retrieved. Everything downstream is second-order.
• Off-site entity strength beats backlinks by roughly 3×. Brand mentions across the web correlate with AI visibility at about 0.66; backlinks at about 0.22.
• Around 84% of AI citations are earned or third-party media. Directories, review sites, comparison content, trade press.
• Freshness earns roughly 3.2× the citations. Recently updated content is retrieved disproportionately.
• No major AI crawler executes JavaScript. Content that only exists after hydration does not exist.
• `llms.txt` is effectively unused. Hundreds of requests across half a billion bot events, and Google states it ignores the file.

What that implies for an agency's roadmap

The high-confidence work is unglamorous and mostly not on the client's own website: get crawlable, get server-rendered, get into the third-party sources the engines actually read, keep things current, and build entity presence. The low-confidence work is the content tinkering that fits neatly into a monthly retainer.

And measure it properly

If the effects are this contested, the measurement has to be good enough to detect one. A before-and-after on single-number scores from three runs cannot distinguish a real intervention from a Tuesday. Bands, run counts and fortnightly windows are not statistical pedantry here — they are the only way to know whether anything you did worked.