Search Platforms and LLMs Enter Symbiotic Phase as 45 Billion Monthly Sessions Reshape Discovery

Large language models and traditional search engines are settling into an unexpected partnership rather than a zero-sum competition, according to new industry data showing generative AI platforms now account for an estimated 45 billion monthly sessions globally while simultaneously funneling users back to conventional search.

Research from Similarweb indicates that platforms like ChatGPT, Perplexity, and Claude have become significant sources of referral traffic to search engines, a dynamic that challenges earlier predictions of wholesale displacement. The pattern suggests users employ LLMs for initial exploration before turning to search for deeper investigation or transactional intent.

"LLMs complement rather than replace traditional search, at least for now," said Melanie Meikle, group partner for paid search at Dentsu, speaking to the evolving relationship between the two discovery modes. The convergence is forcing marketers to abandon siloed channel strategies in favor of integrated approaches that span both ecosystems.

Google's first-quarter earnings underscored the commercial stakes, with the company reporting that AI solutions were the largest contributor to Cloud growth and that its open models had been downloaded over 500 million times. The performance arrived one day after a Wall Street Journal report claimed OpenAI had missed revenue and user targets due to ChatGPT growth slowing as Google's Gemini captured market share.

(The competitive dynamic extends beyond consumer-facing tools: Google's TPU chips are gaining traction with clients including Meta, positioning the company as both an AI model provider and infrastructure rival to Nvidia.)

Image-generation capabilities are proving particularly effective at driving app adoption, according to data from Appfigures. Google's Gemini saw 22 million additional downloads in the 28 days following its Nano Banana image model release—more than four times its baseline. ChatGPT added 12 million installs after introducing its GPT-4o image model, roughly 4.5 times more than it saw from text-focused model upgrades.

The shift toward hybrid architectures is also accelerating beneath the surface. Alexander Amini of Liquid AI noted that top models, including those with trillion-parameter scale from Alibaba and Qwen, are moving toward combinations of transformer and alternative architectures rather than relying solely on the transformer design that has dominated since 2017. The technical evolution reflects growing recognition that different tasks may require different computational approaches.

Meanwhile, evidence of real-world LLM efficacy in high-stakes domains remains sparse. A randomized trial cited in Nature Health found that patients assisted by LLMs for medical scenarios performed poorly, prompting researchers to recommend systematic human user testing before public healthcare deployments. Exceptions include a Nature Health study of LLM use in 16 Kenyan primary care clinics and a single-center ophthalmology trial showing higher diagnostic rates with AI assistance.

Anthhropic's decision to withhold its Mythos model—built for defensive cybersecurity research but deemed too risky for public release—illustrates the governance tensions accompanying capability advances. The move reflects a broader industry debate over whether to prioritize open access or controlled distribution as models gain offensive potential.