AI traffic is exploding, but the sudden surges you’re experiencing on your site may not be delivering the value you think. Crawl volume and referral value are diverging drastically to create an illusion of growth, which can lead businesses dangerously astray.

According to a recent report by bot and fraud protection platform DataDome, 17.7 billion AI agent requests were processed in the second quarter of 2026, representing a quarter-over-quarter increase of a massive 45%. It’s the traffic delusion in action: the traffic spike is real, but the value is not.

Automated Agents: The Stats

When it comes to AI agents, Meta holds the crown as the web’s biggest bandwidth hog, and in June 2026, Meta‑WebIndexer overtook Meta-ExternalAgent for the first time. Why does this matter for your business? Because although these automated agents consume masses of bandwidth, they send almost no visitors. As such, they artificially inflate analytics and insight dashboards while contributing zero value.

Moving on to ChatGPT, we find a very different story. Although its crawl volume decreased 6% quarter-on-quarter, its referral value went through the roof in the same period. ChatGPT is currently driving around 80% of all monthly referral traffic, with human click-throughs seeing significant growth. The bottom line is that ChatGPT might be hitting up your servers less often but, unlike Meta, is sending you almost all the real customers coming from AI tools and platforms.

Other AI agents are also seeing significant shifts, such as Claude, whose referral traffic has increased by 111%, and Perplexity, which enjoyed a 37% increase. Grok, meanwhile, has seen its referral traffic bomb, decreasing by 74%.

MCP Traffic as the New Risk Surface

Model Context Protocol (MCP) traffic is now measurable and, as a new risk surface, is rising fast, with peaks approaching 500,000 a day and clear usage cycles emerging. For businesses, this is yet another risk layer that requires monitoring and potentially defending against.

MCP is how agents access APIs, tools, and structured data and standardizes how agents request actions. As a result, agents have become more capable - but also riskier. By expanding the risk cycle, MCPs create new opportunities for inventorying, scraping, automated exploitation, credential stuffing, basic logic abuse, and more. To make things even more difficult, MCP traffic is still unpredictable and immature, with early implementations varying wildly in behavior. For example, some agents misdeclare intent or over-request tools, and it’s this unpredictability that makes MCP traffic such a risky surface.

The Cost of Phantom Traffic

AI agent surges aren’t just misleading; they’re expensive. A wave of AI traffic hitting your site could trigger an autoscaling event, which means you’ll see your compute and bandwidth spend skyrocket, while high-frequency crawlers inflate your operational cost. These AI agents aren’t ever going to buy anything or engage with your site meaningfully, but will run up your bills.

The financial impact goes well beyond infrastructure. Phantom traffic also affects CDN costs, database load, and application performance. Plus, because an AI traffic surge is usually unpredictable, businesses can’t plan for it, meaning your team ends up scaling reactively, which is typically more expensive than preemptive planning.

The Rise of Spoofed Agents

With AI agents becoming more valuable, attackers are increasingly spoofing their identities to bypass bot detection. These spoofed agents are now one of the fastest-growing automated threats faced by businesses, with fraudsters able to impersonate legitimate agents like ChatGPT and Claude. In this way, attackers hide behind trusted agent identities to perpetrate credential stuffing, scraping, and inventorying.

The problem is clear: user-agent strings are ridiculously easy to fake. Anyone can claim to be ‘ClaudeBot’ or ‘ChatGPT-User’ with a simple header change. Because many businesses whitelist these agents without verification (perhaps yours is one of them), attackers get instant access to areas they shouldn’t. Without using behavioral analysis, it’s impossible to tell the difference between a legitimate agent and an AI imposter up to no good.

Agent Trust is the New Must-Have Capability

There’s a reason that 54% of DataDome’s customers have already adopted agent trust, with the rest sure to follow in the coming months and years. Agent trust policies allow businesses to:

  • Allow agents that send real customers.
  • Block non-value agents that drain resources and infrastructure.
  • Detects spoofed user-agent strings.
  • Monitor MCP traffic before abuse proliferates.

These capabilities are the reason Forrester awarded DataDome the Leader accolade in the Bot Management and Agent Trust Management category. Agent trust is about evaluating agent behavior rather than just blocking bots, and is a fundamental move away from a binary block/allow approach to one of continuous trust scoring. They may be autonomous, but agents are treated, with this method, as digital actors with intent, patterns, and value.

Essentially, agent trust is the only way to prevent falling victim to the AI traffic delusion, because it allows businesses to see traffic spikes for what they are. In the near future, agent identity will likely become a standard, with agents carrying metadata concerning their purpose, origins, and capabilities. This metadata will be treated as a digital password that trust systems will validate.

How DataDome Can Help Pop the AI Traffic Delusion

Most analytics and bot attack mitigation solutions lump all AI agents together, which creates the illusion of growth, but DataDome gives you a clear, true picture. This means you can see which agents send real customers, which drain bandwidth without adding value, which distort analytics, and which are spoofed or malicious.

DataDome has the industry’s most advanced agent classification engine. It uses a range of factors to differentiate a malicious scraper from a helpful ChatGPT-type bot, including:

  • Behavioral biometrics
  • Request pattern analytics
  • MCP metadata
  • Device fingerprinting
  • Intent modeling

The solution is also the first platform to treat MCP traffic as a risk vector. While most other bot mitigation vendors pretty much ignore MCP traffic, DataDome not only measures it, but tracks usage cycles, detects anomalies, and alerts you to early signs of abuse. Crucially, DataDome incorporates MCP metadata into trust scoring, critical given that MCP agents are becoming ever more automated and capable.

Seeing the Truth Behind the AI Traffic Mirage

The surge in AI traffic has created a new type of digital mirage: traffic that looks impressive but represents little to no commercial value. Businesses may be seeing record-breaking volumes of requests, yet the customers behind these numbers are a shimmering illusion. The AI traffic delusion is dangerously misleading, actively shaping wrong decisions, skewing KPIs, and draining infrastructure budgets.

As AI agents evolve (and they’re doing so fast), the gap between volume and value is expected to widen. Some agents, like ChatGPT and Claude, truly drive beneficial referral traffic and discovery. Others, however, like Meta-WebIndexer, eat up bandwidth without sending a single human to your site. With MCP-based automation also quickly rising, the risk surface is expanding in ways many businesses aren’t equipped to monitor, let alone tackle.

Because of all these factors, visibility, classification, and trust matter more than ever. Solutions like DataDome cut through the noise to separate real customers from phantom traffic and illusory growth signals. With AI traffic accelerating, businesses need clarity and the ability to see through the automated mirage, and they need it now.