From Search Results to Query Fan-out: The Reality of Google’s AI Expansion

Google has moved AI Overviews from a niche experiment to a core search pillar, appearing in 43% of US searches as of 2026. That’s a massive leap from 15% just a year ago. But for those of us building on the web, the headline isn't just the "AI-fication" of the SERP; it’s the underlying shift in how queries are processed. We are moving away from "finding a page" toward "synthesizing a concept," and that requires a fundamental change in how we structure our data.

The Mechanics of Query Fan-out

To handle complex, multi-part prompts, Google is leaning into "query fan-out." Instead of trying to find one perfect match for a messy human question, the system breaks the query into subtopics and runs multiple searches simultaneously. It then synthesizes those pieces into a single answer. This is a massive signal for anyone building content systems: "keyword matching" is officially being replaced by "concept decomposition." The system isn't looking for a single picture on the box anymore; it's looking for all the individual pieces of the puzzle across your entire domain.

The Deep-Link Referral Gap

The data on how LLMs cite content reveals a specific, high-stakes pattern. While ChatGPT’s share of responses with web citations jumped fivefold last year (from 1.3% to 6.8%), about 65% of those cited URLs are buried two or three folders below the main domain. Meanwhile, 58.8% of AI referral traffic still hits homepages.

This creates a massive "referral gap." The AI is successfully navigating deep into your site architecture to find the "meat," but the traffic isn't always following it back to the surface. If your deep-content pages are siloed without a clear, logical path for a crawler to follow, you’re losing the chance to be part of that "fan-out" process. You need to build paths, not just pages.

Engineering for a Gated Data Era

The real story here is the looming economic shift in data access. With generative AI site visits up 70% and tools like Cloudflare’s "Pay per Crawl" emerging, we are entering a gated era of data. For publishers and engineers, the challenge is no longer just "getting found"—it’s about creating a structure that is discoverable enough for an AI to decompose into subtopics, yet valuable enough to potentially justify a paywall.

The gap between a flashy AI demo and a production environment where you maintain traffic while managing "crawl costs" is where the next year of engineering will happen. It’s time to stop optimizing for humans alone and start architecting for the machines that are now reading your work.