Outbound referral traffic from ChatGPT to the rest of the web grew 206% in 2025. That is from Semrush’s April 2026 analysis of more than a billion rows of U.S. clickstream data (Semrush, 2026), and it matches what Adobe measured from the retail side a year earlier: traffic to U.S. retail sites from generative AI sources jumped 1,200% between July 2024 and February 2025 (Adobe, 2025).
So generative engine optimization, GEO, is now on almost every discovery call I take. Clients want to know how their brand shows up when a machine writes the answer. The honest reply is longer than the pitch decks suggest.
The traffic is real, small, and brutally concentrated
Growth rates like 206% and 1,200% describe a small base growing fast. On the accounts I can see, assistants still refer a fraction of what organic search delivers. Treat the percentages as a direction, not a stampede.
The direction is consistent, though. Adobe’s same dataset, drawn from over a trillion visits to U.S. retail sites, had already shown generative AI traffic up 1,300% during the 2024 holiday season before the February figure landed. Semrush’s 206% figure, measured a year further on, says the curve did not flatten. And the people arriving from an assistant have been told exactly which page answers their question, which makes this qualified traffic in a way a generic search click never was.
The distribution should temper enthusiasm further. In the same Semrush dataset, roughly 30% of all ChatGPT outbound clicks land on just ten domains, and Google alone collects 21.6% of them. A long tail exists, but the head is very heavy, and most brands are competing for scraps of it.
Meanwhile the classic search click keeps getting rarer, which is exactly why citations matter at all.
When an AI Overview appeared on a Google results page, users clicked a traditional result link on just 8% of visits, versus 15% when no summary was shown. Only 1% clicked a source link inside the summary itself (Pew Research Center, 2025).
Put those findings together and the strategy writes itself: fewer clicks are leaving the answer layer, so being the source the answer is built from is the new shelf position. Whether a click follows is a separate fight.
One measurement caveat from practice: a chunk of assistant traffic arrives in analytics with no referrer and gets logged as direct. Tag what you can, segment the known AI referrers, and accept that the true number is somewhat higher than the report shows.
What verifiably moves citations
The only controlled experiment of real note remains the GEO paper from Princeton and IIT Delhi researchers, presented at KDD 2024. Across a benchmark of roughly 10,000 queries, adding statistics, quotations and source citations to a page increased its visibility in generative engine answers by up to 40% (Aggarwal et al., 2024).
Read that finding carefully, because it is quietly conservative. The tactics that won are the credibility signals a good editor would demand anyway: numbers with named sources, quotable claims, clear attribution, fluent writing. Keyword stuffing performed poorly in the same tests. The machines prefer pages that read like evidence.
Translated into the work I actually do for clients:
- Be the primary source of a number. Original data gets cited; your paraphrase of someone else’s data does not.
- Keep entity facts boringly consistent. Same brand name, same founder, same locations, same product claims across the site, LinkedIn, directories and press. Assistants cross-check, and inconsistency reads as unreliability.
- Structured data and a clean heading hierarchy help retrieval. Cheap insurance, not magic.
- Comparison pages earn citations because assistants love a ready-made ranking. If you refuse to compare yourself to alternatives, a competitor’s comparison becomes the answer.
A war story from this spring. A B2B services client asked why one competitor kept surfacing in ChatGPT answers for their category. The reason was unglamorous: that competitor had published a pricing survey with actual numbers in it, and every assistant I tested leaned on it as the citable source in the niche. Nothing about their product was better. They had simply given the machines something to quote.
My advice to that client was not a GEO retainer. It was to publish the benchmark their industry was missing, with real methodology attached, and let the assistants find it.
An honest word on how noisy this still is
I review AI adtech tooling as an independent technical reviewer, so I say this with some sympathy for the people building in the space: most GEO measurement today is prompt sampling dressed up as rank tracking. The same question asked twice returns different sources. Model updates reshuffle citations overnight, and nobody publishes a changelog.
Anyone selling a deterministic GEO score is early at best. What works instead is modest: keep a fixed basket of real buyer questions, sample them monthly across ChatGPT, Perplexity and Google’s AI results, log which sources get cited, and watch the trend. That is a spreadsheet discipline, not a platform purchase.
Budget accordingly. I would not defund working paid channels to chase citations in 2026, and I would not publish another word of generic content that neither ranks nor gets quoted. The work that earns citations, meaning original research, digital PR, consistent entity data and a technically clean site, is the same work that has been winning organic search for a decade. GEO is less a new discipline than a new scoreboard on an old one.
The machines reward exactly one thing reliably: being genuinely worth citing. That has the convenient side effect of being good marketing anyway.