2026 State of AI Discovery Report

ChatGPT Commands 92% of AI Referral Traffic: Analysis of 6.77M LLM Sessions. Learn how the AI discovery market is consolidating, which challengers are gaining ground, and what it means for SEO and marketing strategy.

Previsible analyzed 166 GA4 properties from November 2024 through May 2026, spanning websites across SaaS, e-commerce, finance, legal, health, insurance, education, publishing, and ticketing. This analysis reflects LLM search behavior and product change over time.

In sum, we evaluated 6.77 million LLM-driven sessions and uncovered 9.9x growth in 19 months: 92.4% of it comes from ChatGPT.

This is the largest AI traffic study we’ve ever published, measuring referral traffic from standalone LLM platforms.  The data tells a consolidation story. ChatGPT is transforming AI discovery.  The challengers the industry bet on 12 months ago aren’t the ones gaining ground.

Important context: AI discovery happening inside Google’s own results, including AI Overviews, almost certainly represents a larger volume of AI-driven traffic than all standalone LLM platforms combined. We exclude AI Overviews from this study because they operate on a fundamentally different measurement paradigm. They don’t generate trackable referral sessions the same way standalone platforms do. This report measures standalone LLM referral traffic specifically, and ChatGPT’s dominance applies within that frame.

Key Findings

• ChatGPT commands 92.4% of all trackable LLM referral traffic and is still gaining relative share.
• Claude grew 64x over the tracked timeframe and overtook Perplexity in March 2026.
• Gemini is the quiet #2, growing 3.2x since November 2024 with little volatility.
• Perplexity peaked in March 2025 and has fallen 61% since. Copilot collapsed 96% from its peak.
• Monthly LLM sessions hit 644,478 in May 2026 – a 9.9x increase from November 2024.

AI Traffic Is Accelerating

In mid-2025, AI traffic seemed to be peaking for some sectors like SaaS. Turns out it was only a pause on the way to new highs.

From November 2024 to May 2026, total monthly AI-referred sessions grew 9.9x.

The November 2025 dip requires context. Monthly sessions fell 50% in a single month, driven almost entirely by a drop in ChatGPT referrals (448,412 in October to 213,345 in November). Other platforms remained stable. 

This was likely a model-related change. We’ve seen that seemingly modest tweaks can massively impact referral traffic as documented by Search Engine Land last fall.  Many sites saw a 50% decrease in traffic because ChatGPT preferred Wikipedia and Reddit.

Whether this reflects solely an OpenAI product change or a seasonal user behavior shift isn’t clear without more detail from OpenAI. Sessions recovered to 442,609 in December 2025.

The message is clear: search is becoming more like the LLM chat experience and LLMs are increasingly capturing more searches.  Based on conversations we had with engineers at Google I/O, it’s full speed ahead to an AI search future.  The mix of that traffic, how it refers to publisher sites and where to invest are what matter now.

Platform Market Share: A Consolidation Story

The last time we looked at the data in December 2025, ChatGPT was at roughly 84% followed by Perplexity (8.9%), Gemini (4.5%), Copilot (2.1%) and Claude (0.6%). Since then, Perplexity and Copilot have dropped dramatically and ChatGPT has taken more share of AI traffic. This report measures standalone LLM referral traffic specifically. AI Overviews are excluded.

This average over 19 months smooths out some ups and downs to create some digestible market insights.  However, what it misses are ups and downs are entire product life cycles representing billions of investment trial and error over a rocky 19 months. What we’ll look at next is the month to month trajectory which tells a more important story.

ChatGPT: 92.4% Share and Still Pulling Away

47,606 sessions in November 2024 became 610,910 by May 2026. That’s a 12.8x growth over 19 months with no sign of slowing. GPT is the market leader across most industries and page types, it is the only LLM sending meaningful referral traffic.

ChatGPT is the AI discovery surface that matters right now. Optimizing for “AI visibility” without prioritizing ChatGPT is optimizing for an abstraction.

Claude Grew 64x and Overtook Perplexity in March

Claude took longer to allow web search, releasing in March 2025. That may be part of the reason why, according to Cloudflare, it is 30,000 times harder to get a visit from Claude versus traditional Google Search.

This traffic is precious and growing. In November of 2024, we saw only 133 sessions.  Last month in May, it had grown by 64x to 8,528.  While mostly flat through 2025, hovering between 1,700 and 2,000 monthly sessions, traffic jumped in January to 2,340, then 3,514 in February and 9,501 by March. A 4x acceleration in two months.

It’s clear that Anthropic’s second-mover advantage is paying off.  They focused on product and sticking to their principles and it’s paying off.

In March 2026, Claude surpassed Perplexity in monthly referral sessions for the first time and has stayed ahead.

Claude’s growth correlates with its expansion into agentic coding tools, professional workflows, and enterprise adoption. The enterprise advantage that the industry expected Copilot would win may be materializing for Claude instead, likely because Claude’s integration points (Claude Code and Claude Cowork) are increasing rapidly in popularity.

Claude is no longer a rounding error. If your audience includes technical buyers, developers, or professional services, Claude visibility is becoming a real factor, and the window for early positioning is now.

Gemini Grew 3.2x With No Volatility

Starting at 5,598 sessions in November 2024 and ending at 18,119 in May 2026, Gemini has consistently grown steadily to take over the #2 spot.

Gemini’s integration into Google Workspace and Android means its actual discovery footprint likely extends beyond what referral tracking captures. What we measure here may undercount Gemini’s real influence on how users find and evaluate brands.

According to Google’s VP of Search Liz Reid, “Google Search is AI Search.”  Gemini and Google search are collapsing into a similar experience where you can start in the search bar, transition into Gemini conversation, interact with inventory from your website, and use their Universal Cart Protocol to buy.  If that model scales, referral traffic becomes secondary to inventory integration. That said, closed ecosystems have been attempted many times and rarely succeed long-term. Whether Google can sustain this model without alienating merchants and publishers remains an open question.

Gemini isn’t flashy but it’s durable. Its steady trajectory and Workspace/Android integration suggest the referral numbers undercount its real discovery footprint. It’s the model to watch as Google continues to push new features like Spark and Search Agents that are designed to capture more modalities. 

Declining Platforms: Perplexity and Copilot

Perplexity showed early promise of being a lasting LLM for search discoverability.  Based on what we’re seeing, Perplexity peaked at 17,507 monthly sessions in March 2025, and by May 2026, traffic dropped to only 6,788.  That’s a 61% decline.

Like its competitors, Perplexity is shifting toward retaining users within the Perplexity experience – browser, agentic experiences. In these modes, they don’t need to send you traffic; you are getting the traffic as a browser or as an agent.  You may see more Perplexity traffic coming by way of GA4’s new AI-Assistant Medium (see our free Data Studio Report to track for your site).

Copilot’s decline is steeper. It peaked at 8,651 sessions in August 2025, then declined to 339 by May 2026. A 96% drop. In our prior report, we highlighted Copilot’s rapid growth as evidence that workplace-embedded AI discovery was arriving through Microsoft Office. That growth proved to be a short spike, not a trajectory. Copilot adoption is seeing even more pressure now that the pricing model is changing.  Microsoft is even looking at switching to Deepseek as a potential cost saving measure.

Neither Perplexity nor Copilot is a growth bet for traffic acquisition. Perplexity’s value may persist in specific niches (see model personalities below), but broad optimization for either has diminishing returns.

Industry Penetration: Where AI Traffic Concentrates

Traffic from LLMs is not evenly distributed across industries.  We wanted to better understand this by looking at the percentage of a vertical’s total traffic coming from LLMs.

Across the dataset, LLM traffic represents a fraction of total sessions. Drilling down into this penetration view, it helps us to make decisions about when and where to invest.  For example, there has been a contraction in LLM traffic in the health space and a consolidation around branded content.  That might lead to building out authority signals like PR and focusing on consistency signals to LLMs.  Let’s take a deeper look at each industry:

The fastest-growing verticals started from the smallest bases. E-commerce grew 37x from near zero. Users are now asking LLMs for product recommendations and landing on product pages with purchase intent already formed. Insurance grew 18.9x to 1.51% penetration, the highest current rate of any vertical. Education grew 5.4x, consistent with LLMs’ strength in synthesizing multi-step learning queries.

SMB and finance tell a different story: moderate growth off of higher starting points. Finance started at 0.56% and reached 1.19%. SMB started at 0.4% and reached 1.71%. These verticals adopted AI discovery earlier and are now in steady accumulation rather than explosive acceleration.

Health is the only vertical where AI penetration declined, from 0.23% to 0.17%. E-commerce AI traffic is almost entirely ChatGPT; no other model sends meaningful referral volume to e-commerce properties.

---

Page Type by Vertical: AI Discovery Intent Varies More Than You Think

Where LLM traffic lands reveals what users are asking AI to help them do. But “where it lands” depends entirely on the vertical. The aggregate view tells a tidy story about a shift from information to commerce. The vertical-level view tells ten different stories. That’s the one that matters.

• SaaS: Search pages dominate (34.6%). ChatGPT trusts the domain but can’t pick the right page, so it sends the user to your internal search result page. Sites with strong search UX convert these sessions. Sites without one lose high-intent visitors at the front door.
  
• Publisher: News pages capture 54% of LLM referrals. But against 120+ million organic sessions, penetration is 0.08%. Publishers produce the content LLMs train on and cite, but capture almost none of the resulting traffic.
  
• E-commerce: Product pages are the primary landing surface. AI traffic is almost entirely ChatGPT, and users arrive with purchase intent already formed. Structured, comparable product data is an AI discoverability requirement, not just a conversion optimization play.  Learn more about the best way to segment a catalog for AI platforms.
  
• Education: Course pages are the entry point (52%). Users ask LLMs “where can I learn X” and land directly on course content, bypassing blogs, guides, and marketing pages entirely.
  
• Financial Services (Finance + Insurance): Blog content leads, but location and conversion pages reveal intent. While blogs capture the largest share of LLM traffic, location pages represent bottom-of-funnel, offline-conversion intent arriving through an AI channel. Conversion pages (enrollment, signup, product entry points) show some of the highest per-page AI density in the dataset, with LLM-to-total ratios above 2%.
  
• Health: About pages serve as signals (42.1%). When users ask an LLM a health question and get directed to a site, their first action is evaluating the source. About pages serve as a credibility checkpoint.  More from John Vantine at GoodRX on AI-driven strategies.
  
• Legal: The most evenly distributed spread. Blog (28.2%), about (12.3%), contact (11.7%), and location (10.3%) each take meaningful share. LLMs route users through the full evaluation arc:  practice areas, credentials, contact.
  
• Ticketing: Search (20.8%) and home (18.5%) split the lead. Low overall penetration and sending traffic to the home page suggest LLMs handle basic event discovery rather than deep evaluation.

Model Personalities: How Each Platform Sends Traffic Differently

The aggregate market share data obscures meaningful behavioral differences between platforms. Each LLM routes traffic to different page types, revealing distinct “personalities” in how they handle discovery. These differences have practical implications for where you invest.

Search-Pattern Models (ChatGPT, Gemini): Domain Trust, Page-Level Uncertainty

ChatGPT and Gemini have similar patterns around how they refer traffic.  They trust domains broadly but can struggle with page-level specificity. ChatGPT sends 28.8% of its traffic to internal search pages. It names the brand but often can’t name the page. Gemini does the same, routing users to search infrastructure and topical hubs rather than guessing at a specific URL. This may be the most actionable finding in the study. Unlike market share percentages, which shift with every model update, the pattern of LLMs trusting domains but struggling with page-level specificity appears structural. It persists across verticals and time periods in our data, suggesting it reflects how retrieval-augmented generation works rather than a temporary product quirk.. Both try to meet the intent but sometimes approximate the best page to link to.  They acknowledge domain authority and let the user navigate from there. ChatGPT’s second-largest concentration is product pages, consistent with its dominance on commercial queries. Gemini’s pattern is steadier and less commerce-oriented.

If this is the case, you need to think of your internal search experience as an acquisition tool. The model did the hard work of selecting your domain. Your search UX determines whether that visit converts, and if you answer your own customer’s questions or competitors do.  Navigation matters just as much. Clear site architecture, logical menus, and prominent internal linking help both users and LLMs find the right page.  It also points to a need to utilize your documentation and forum content. Here are some examples we’ve seen:

• Documentation Search – How do webhooks work in HubSpot?
• Community / Forum Search – Has anyone solved Salesforce duplicate lead issues?
• Release Notes / Changelog Search – Did Asana recently add AI features?
• Product Search – I need a non-AI stock image of fruit on a table.

Content-Selection Models (Perplexity, Claude): Page-Level Precision, Long-Form Bias

Perplexity and Claude pick specific pages. Perplexity over-indexes on blog and long-form content, with 13% of its classified traffic landing on blog pages, as opposed to 8.7% in the overall mix. Perplexity treats blog posts, guides, and editorial content as primary sources in ways the search-pattern models don’t. Claude skews even further toward educational content like guides, courses, and research-oriented pages. Claude users engage with longer-form content at higher rates than any other platform’s referrals.

For publishers and content-forward brands, these models deliver smaller volume but higher-quality referrals to the content they’ve invested most in producing. If your strategy depends on long-form content driving qualified traffic, Perplexity and Claude visibility matters disproportionately to their market share.

What SEO and Marketing Teams Should Do Now

These recommendations are grounded in data.

1. Optimize for ChatGPT first.

As the number one referrer of AI search, build your AI visibility strategy around ChatGPT. Expand to other platforms when their volume justifies the investment.

2. Monitor Claude.

Claude overtook Perplexity in March. If your audience includes developers, technical buyers, or professional services, Claude visibility is material now, not speculative. Early positioning creates compounding advantages.

3. Treat product pages as AI entry points.

Product pages now capture 43% of all e-commerce LLM traffic. Users arrive with purchase intent formed. Structured, comparable product data across PDPs is no longer just a conversion rate optimization play. It is an AI discoverability requirement. If your product data isn’t clean, scannable, and comparison-ready, you’re invisible to the fastest-growing discovery channel.  Apple is a great example of great UX with LLM-friendly content.

4. Make pricing transparent and easily digestible for AI systems.

The case for transparent pricing is no longer just “buyers like it.” It is now an AI visibility issue. Gartner says most B2B buyers prefer rep-free buying experiences; G2 is actively moving toward exposing software price ranges because buyers need them to compare and build internal cases; and Google’s structured data guidance shows that pricing is a machine-readable surface, not just a sales artifact. “Contact us for pricing” may protect sales process control, but it gives AI systems very little to summarize, compare, or recommend.

5. Prioritize search experiences.

AI-referred traffic lands on internal search results pages roughly 25% of the time across all industries. Your internal search experience is now an acquisition tool, not just a navigation feature. If users land on your search page from an LLM and get poor results, you’ve lost a highly engaged user that an AI specifically sent to you.

6. Track AI traffic by page type, not site-wide.

Your overall AI penetration rate may be 0.25%. Your pricing page might be 3x that. Your product pages might be 5x. Your site-wide average hides where AI traffic actually concentrates. Measure by page type.  You can use this free Looker Report that tracks LLM traffic and AI Assistants.

---

Our Next Insights

Volume and penetration tell you where LLMs send users. They don’t tell you what those users are worth.

Conversion rate by LLM platform is the single most valuable unanswered question in AI discovery. Which platforms send users who buy? Which send users who bounce? Does a ChatGPT referral convert better or worse than a Google click? Does a Perplexity referral to a blog post lead to pipeline at the same rate as a ChatGPT referral to a product page?

We built this dataset to answer those specific questions. If the last 19 months are any indication, the answers will change faster than most teams are ready for.

Track LLM updates at previsible.com/ai-seo-benchmark

Methodology

This is the third Previsible AI Traffic Study. The dataset spans November 2024 through May 2026 (19 months) across 166 GA4 properties in SaaS, e-commerce, finance, legal, health, insurance, education, publishing, events, and SMB.

The property pool is subject to change between studies. Absolute figures should not be compared directly across Previsible reports without accounting for differences in sample composition and size. All 166 properties in this study are present for the full 19-month observation window, so trajectory analysis within this report reflects behavioral change, not sample expansion.

AI penetration rate is calculated as: (LLM Sessions / Total Sessions) x 100. It can be measured site-wide, by industry, or by page type. Sessions follow the GA4 default definition (30-minute inactivity timeout). LLM sources tracked include chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai, grok.com, you.com, and meta.ai, along with associated subdomain and app-referral variants captured in GA4.