Graphite retracts its widely cited claim that AI usage equals 56% of search volume

Graphite has retracted the headline comparison from its own March report, the one that told software marketers AI assistant usage is already 56% the size of global search. In an updated study published November 9, the SEO and AEO agency says it no longer has a reliable way to make that comparison at all, because its original numbers undercounted Google's own search traffic.

The retraction matters because that 56%-of-search figure, and its US counterpart of 34%, became a stock slide in pitches for GEO budget, in content strategy decks, and in trade coverage arguing that AI had already closed most of the gap with search. Graphite's own research page now carries an "Updated research" banner directing readers to the new study, "The State of Gen AI," co-authored by CEO Ethan Smith and senior data scientist José Luis Paredes.

The figure spread fast because it filled a gap. Most AI-visibility data comes from single-platform dashboards or self-reported numbers from the AI companies themselves, so a cross-platform comparison with a specific percentage attached was unusually quotable. Trade press, GEO vendors and marketing teams picked it up without independently checking Graphite's underlying session definitions, and it circulated for eight months before the retraction.

Why the comparison broke

The March study pulled AI session counts from six assistants and set them against search engine sessions drawn mostly from desktop and mobile web. It left out mobile app sessions for search engines, including Google's own Android app, which Graphite now estimates logs about 11.8 billion monthly sessions in the US alone. That app covers far more than search: image recognition, translation and Google Discover all run through it, and Graphite says it has no reliable way to isolate how many of those sessions are search.

"We can no longer reliably compare the relative size of AI for search vs. search engines," Graphite wrote in the new report.

The effect runs the opposite direction from what GEO advocates might expect. Because the Android app's search sessions were left out of the March denominator, Graphite says its own study likely undercounted Google, which means AI's relative share was probably overstated, not understated, in the figure marketers have been repeating for eight months.

What the new numbers actually show

Dropping the search comparison doesn't mean Graphite found less AI usage. The six largest assistants, ChatGPT, Gemini, Claude, Grok, Perplexity and DeepSeek, drove an estimated 76 billion monthly sessions worldwide in September, up 37% year over year. But the growth rate has cooled: Graphite puts the current pace at roughly 7% worldwide on a rolling three-month basis, a sharp slowdown from the 396% growth the market saw in the year to September 2025.

Referral traffic to third-party sites tells a smaller-scale story than session counts. Graphite's analysis of roughly 43,000 US domains found AI still accounts for less than 1% of average website traffic, even as referrals grew 92% worldwide and 75% in the US. ChatGPT holds about 81% of worldwide AI session share and 60% of referral share; Claude's worldwide sessions grew 1,901% off a small base, while Perplexity has fallen 27% worldwide from its October 2025 peak, a decline Graphite attributes partly to copyright lawsuits and native OS search integrations crowding it out.

What this means for your GEO numbers

Every software marketer who has cited the 56% figure, or built an internal dashboard on a similar cross-platform comparison, now has a data point that its own source no longer stands behind. Search Engine Journal's reporting notes a related wrinkle: OpenAI has separately said ChatGPT handles roughly a fifth as many search-like queries as Google, without disclosing Google's own figure, and counts a "search" differently than Graphite does. None of the vendors measuring this space agree on a denominator, which is exactly why a single retracted comparison can't be patched with a different single source.

The exposure isn't limited to anyone who quoted the number directly. Product marketing teams building AI-visibility talking points into battlecards, growth teams sizing GEO budget against paid search budget, and comms teams using the stat as a press hook are all carrying a number its source has disowned. The fix isn't swapping in a different vendor's percentage; it's being explicit, internally and externally, about what a session-based estimate can and can't tell you.

• Pull the 56%/34% AI-vs-search figure from any live deck, blog post or sales narrative that cites it
• Treat vendor session and traffic reports as directional trendlines, not precise market-share numbers you can defend in a board meeting
• When you quote a benchmark externally, carry its stated caveats with it, not just the headline number
• Lean on your own first-party data, Search Console's AI impressions, your analytics' AI-referral segment, for any number you present as decision-grade