17.7 billion AI agent requests in a single quarter. A 45% jump from Q1. And more than half came from Meta - a company that sent almost no visitors back in return.
This is the traffic layer most website owners aren’t looking at.
Meta’s Paradox: Biggest Crawler, Lowest Return
DataDome - an infrastructure protection company analyzing 5 trillion signals daily across 400+ enterprise clients - released its Q2 2026 AI Traffic Report with a finding that should change how publishers think about AI access.
Meta accounted for 9.1 billion requests out of 17.7 billion total AI agent requests in Q2 2026.
The company runs two distinct crawlers. Meta-ExternalAgent, used for model training, grew 74% QoQ to 5.3 billion requests. Meta-WebIndexer, used for real-time indexing for Meta AI, grew 163% QoQ to 3.75 billion requests. For the first time in June, WebIndexer surpassed ExternalAgent in monthly volume (DataDome, 2026).
The catch: Meta sends “almost nobody back in return.” This is the core publisher complaint crystallized in data - AI companies extracting content without returning referral traffic or measurable value.
Who Actually Sends Visitors: ChatGPT Leads, Claude Fastest Growing
While Meta dominates volume, OpenAI and Anthropic are the agents actually delivering visitors to your site.
ChatGPT maintains overwhelming referral dominance at 80-88% of all monthly AI-driven referral traffic, with +17% QoQ growth - consistent and reliable (DataDome Q2, 2026).
Claude recorded the sharpest growth: referral visits up +111% to 876,000 visits in Q2 - nearly doubling from 415,000 in Q1. Perplexity grew +37%.
Grok from X/Twitter? Down -74% to just 24,000 referral visits. A dramatic collapse despite X being a billion-user platform.
The pattern is clear: crawl volume and referral value move in opposite directions. The agent hitting your site most often is not the agent sending you the most business.
MCP Traffic: The New Signal Nobody Is Monitoring
The most significant emerging finding in the report: MCP (Model Context Protocol) traffic has appeared as a distinct and rapidly growing category.
From late April 2026, MCP traffic escalated from negligible to approximately 500,000 requests per day, with clear usage cycles - peaks and troughs aligned with working hours. These are not passive crawls. MCP calls like initialize, tools/list, and prompts/list reveal agent intentions before action - an AI is actively interrogating your site’s capabilities (DataDome, 2026).
If your site exposes structured data, APIs, or services, MCP traffic signals that AI agents are interacting with your infrastructure - not just reading your HTML.
What This Means Beyond the US Enterprise Market
By Q4 2025, 1 in 31 website visits came from AI crawlers - up from 1 in 200 just one year earlier (No Hacks, 2026). This ratio is accelerating with no signs of slowing.
For markets in Southeast Asia, the gap is stark. Nearly 90% of companies in the region plan to experiment with AI agents in 2026 (EDB Singapore, 2026). Yet awareness of AI traffic management - let alone active agent policies - is minimal.
In Vietnam specifically, the AI Law effective March 2026 establishes governance frameworks for AI deployment. But web infrastructure governance around incoming AI agents? That conversation hasn’t started.
Only 54% of DataDome enterprise customers have implemented agent trust policies - tools that classify which AI agents can do what on their infrastructure (DataDome, 2026). In US/EU enterprise contexts, this is still a minority. In Southeast Asia, it’s essentially zero.
The bandwidth problem is real and immediate. AI crawlers consume server resources like human visitors but generate no conversions. For sites on unoptimized hosting, this directly increases operational costs without any benefit.
Three Steps That Actually Matter Now
Measure first. Enable GA4’s AI Assistant channel filter under Traffic Acquisition. Review server logs for user-agent strings. You need a baseline before you can make decisions.
Classify by intent, not volume. Not all AI crawlers are the same. ChatGPT and Claude crawl to understand and cite content - valuable to you. Meta crawls to train models and index for its own AI products - valuable to Meta. Treat them differently.
Optimize for the agents that send traffic. With ChatGPT holding 80-88% of AI referral share and Claude growing fastest, these are the two “readers” worth writing for. Structure your content to be citable. Ensure your key pages are accessible to these crawlers. The MCP signal means the infrastructure layer matters too - not just the text.
The question isn’t whether AI agents are hitting your website. They are. The question is whether you know which ones, what they’re doing, and whether any of them are actually working for you.
For most organizations - including most in Vietnam and Southeast Asia - the honest answer right now is no.
NateCue's Take
Most website owners are measuring traffic with Google Analytics and calling it visibility. This report reveals a parallel layer that GA4 doesn't show - AI agents, crawlers, MCP requests - and it's larger than most teams expect. The strategic question isn't whether to block AI crawlers. It's whether your content is being read and cited by the agents that actually send visitors back. Right now, that means optimizing for ChatGPT and Claude - not Meta, which dominates crawl volume but returns nearly zero referral value. For markets like Vietnam and Southeast Asia, this gap is even larger. AI traffic management is essentially a concept imported from US/EU enterprise - but the crawlers don't check borders before consuming your server resources. The 54% of DataDome enterprise clients who have agent trust policies figured this out. Most publishers haven't.