Is AI agent traffic killing the internet?

Are AI agents really taking over internet traffic?

AI agents are generating billions of web requests, but they still account for a small share of internet traffic. The bigger story is how quickly that share is growing and what websites get in return.

Imperva's 2025 Bad Bot Report found that automated systems generated 51% of web traffic in 2024, up from 49.6% in 2023. However, that total includes malicious bots, traditional search crawlers, monitoring software and other automated systems. It certainly doesn't mean AI agents account for half of internet activity. Half the web being bots tells you little about agents, which is why we separate real progress from hype in our agentic AI reality check.

DataDome's 2026 State of Bot & Agent Security Report gives us a clearer picture. Across more than 75,000 customer websites, the company detected 52.7 billion requests from AI agents and large-language-model crawlers between July 2025 and June 2026. Their share of total requests increased from 1.2% to 1.7%, while absolute AI traffic grew 82.3%.

Cloudflare measured a somewhat larger share across its own network. During 2025, identified AI bots excluding Googlebot generated an average of 4.2% of HTML requests. Googlebot alone accounted for another 4.5%, although Google's crawler also serves conventional search.

These figures measure different websites and types of requests, so we cannot combine them into one global percentage.

What would it mean for AI agents to kill the internet?

AI agents could seriously damage the open web by taking away the visitors and revenue that keep websites running. The internet's technical infrastructure can survive that change, but thousands of publishers and online businesses may struggle to adapt.

For roughly two decades, search engines and publishers operated through a fairly simple arrangement. Publishers allowed Google and other search engines to index their content, and search engines directed people back to the original websites. Those visits generated advertising revenue, subscriptions, purchases and customer relationships.

AI changes the arrangement. ChatGPT, Gemini and other assistants can retrieve information, summarize multiple sources and answer questions without requiring users to open the original pages. Autonomous agents can go further by comparing products, interacting with services or completing tasks on someone's behalf.

There are three different concerns here. Automated requests can put pressure on website infrastructure. AI-generated answers can reduce the number of human visitors. And autonomous agents can gradually take control of the interactions that businesses previously handled themselves.

The evidence is currently strongest for declining referral traffic and growing pressure on publishers. Fully autonomous transactions are still developing, while an internet-wide infrastructure crisis remains unsupported by the available data.

How fast is AI agent traffic growing right now?

AI agent traffic is growing extremely fast, with DataDome recording an 82.3% increase over twelve months. But the growth isn't evenly distributed, and we should be careful about interpreting every automated request as evidence of a new autonomous agent.

DataDome recorded approximately 3.5 billion AI requests in July 2025 and 6.6 billion in June 2026. Its monthly measurements show some interruptions, including noticeable declines in September 2025 and February 2026.

The second half of 2025 generated approximately 23.06 billion requests, compared with around 29.65 billion in the first half of 2026. That works out to roughly 28.6% growth between consecutive six-month periods.

The latest three months in that series were particularly strong. Monthly requests reached approximately 5.5 billion in April, 6.4 billion in May and 6.6 billion in June. Growth had continued well beyond the initial surge.

Cloudflare's 2025 Year in Review provides another useful measurement: crawling triggered by user actions increased more than fifteenfold during the year. These requests are more closely connected to people asking AI assistants to retrieve information than traditional crawlers collecting large datasets.

Whether request volumes continue growing at the same pace will depend partly on caching, direct data integrations and how often AI assistants need to retrieve fresh information.

Are claims that bots generate half of internet traffic misleading?

Claims that bots generate more than half of internet traffic are supported by major security reports, but they become misleading when people assume those bots are all AI agents.

Imperva measured automated traffic at 49.6% in 2023 and 51% in 2024. Within that total, malicious bot traffic increased from 32% to 37%. Those categories cover activities ranging from credential attacks and scraping to ordinary indexing.

Cloudflare's 2025 HTML-request analysis shows why the distinction matters. Identified AI bots excluding Googlebot accounted for 4.2% of requests, while Googlebot represented 4.5%. Conventional non-AI bots accounted for a much larger share.

Request counts are also different from visitor counts. One person opening a webpage can trigger numerous requests for images, scripts and other resources, while a crawler downloading a document may generate only one or several requests.

Measurement	Reported result	What it measures
Imperva, 2023	49.6%	All automated web traffic
Imperva, 2024	51%	All automated web traffic
Imperva, 2024	37%	Malicious bot traffic
Cloudflare, 2025	4.2%	AI bots excluding Googlebot, among HTML requests
Cloudflare, 2025	4.5%	Googlebot's share of HTML requests
DataDome, July 2025	1.2%	AI requests within its customer traffic
DataDome, June 2026	1.7%	AI requests within its customer traffic

Which AI companies are crawling the internet the most?

Meta and OpenAI dominate identified AI crawling traffic in DataDome's latest measurements, together accounting for more than 80% of the activity. That concentration is much greater than most discussions about the growing number of AI companies would suggest.

During the first half of 2026, Meta represented 46.3% of identified AI traffic across DataDome's network, followed by OpenAI at 34.6%. Huawei's PetalBot and Amazon each accounted for 5.3%, while Anthropic and Perplexity represented smaller shares. Meta and OpenAI dominate the crawl, but we cover the wider list of companies that lead today in our report covering agentic AI.

Meta's position is particularly revealing. Its Meta-ExternalAgent and Meta-WebIndexer generated approximately 8.54 billion and 5.30 billion requests respectively during the period.

These crawlers support AI-related information retrieval and indexing, but they shouldn't be confused with millions of independent agents completing tasks for users.

OpenAI also operates crawlers with different purposes. GPTBot is associated with training-related collection, while ChatGPT-User requests are linked to user-directed retrieval. Google's indexing infrastructure is even harder to classify because Googlebot serves both traditional search and AI-related functions.

Meta-affiliated AI bot traffic increased 164.9% between January and June 2026, according to DataDome. DuckDuckGo's AI bot traffic grew 423% over the same period, showing that smaller operators can expand quickly even while the largest platforms dominate absolute volumes.

Are AI agents already replacing human visits to websites?

AI-powered answers are already replacing some visits to websites, especially when people search for straightforward information. We have direct evidence of fewer clicks, although autonomous agents are only one part of the broader change in search behavior.

Pew Research Center analyzed browsing activity from 900 US adults in March 2025. When Google displayed an AI-generated summary, users clicked a traditional search result in 8% of visits. Without an AI summary, they clicked a search result in 15% of visits.

That represents an observed click rate roughly 47% lower. Only 1% of visits involving an AI summary produced a click on a source link within that summary.

Pew also found that 26% of visits involving AI summaries ended the browsing session, compared with 16% of searches without them. Those figures suggest that AI answers can satisfy users before they reach another website.

There is a caveat. Searches that trigger AI summaries aren't identical to searches that don't. Longer and question-based queries are more likely to produce AI summaries, so we cannot attribute the entire click-through difference to the feature itself.

Wikipedia provides another independent example. After improving its bot detection and revising its measurements, the Wikimedia Foundation reported approximately 8% fewer human pageviews during parts of 2025 compared with the same months in 2024.

The Foundation connected that decline partly to generative AI and social platforms changing how people obtain information.

How much traffic are publishers losing because of AI search?

Publishers are losing a large share of their search traffic, and the decline has been getting worse. AI-generated answers are contributing to that pressure, although Google's ranking changes and wider shifts in audience behavior also play a role.

Chartbeat's analysis found that Google Search pageviews across its publisher network fell 34% between December 2024 and December 2025. Its subsequent 2026 Publisher Playbook reported a 40.2% year-over-year decline through July 2026, compared with a 21.9% decline in the preceding comparable period.

Google Discover also weakened, with its year-over-year decline moving from 6.6% to 34.3% across the periods examined in the later report.

Smaller publishers appear especially vulnerable. Chartbeat found that search referrals had fallen 60% over two years for websites receiving 1,000–10,000 daily pageviews, compared with 47% for medium-sized publishers and 22% for large publishers.

A smaller publication often has less direct traffic, weaker brand recognition and fewer alternative distribution channels. Losing search visitors can quickly affect advertising income and the ability to fund new reporting.

If a publication previously received 100,000 monthly search visits, a 40% decline would leave it with roughly 60,000, assuming its traffic followed the measured aggregate. Keeping advertising revenue unchanged would then require materially better monetization or additional visitors from elsewhere.

Can ChatGPT replace the traffic publishers are losing from Google?

ChatGPT is sending more people to external websites, but AI referrals are still far too small to replace the traffic publishers have lost from Google.

Chartbeat reported that ChatGPT referrals increased more than 200% year over year. Despite that growth, AI chatbot referrals accounted for less than 1% of pageviews across its publisher network.

DataDome tracked 89.4 million visits referred by AI chat platforms to its customer websites during the first half of 2026. ChatGPT generated 83.4% of those visits, followed by Gemini at 7.9%, Perplexity at 4.8%, Copilot at 2.2% and Claude at 1.5%.

That works out to approximately 74.6 million ChatGPT referrals, 7.1 million from Gemini and 4.3 million from Perplexity over six months.

Interestingly, AI referral activity wasn't rising continuously. DataDome measured a peak of 19.1 million visits in March 2026, followed by roughly 13–14 million monthly visits through midyear.

AI-referred visitors may be more valuable than average because they often arrive after an assistant has helped narrow down their options. Someone asking ChatGPT to compare accounting software may have stronger purchase intent than someone casually searching for a definition.

But the arithmetic remains difficult for publishers. A channel accounting for less than 1% of pageviews cannot compensate for a major distribution channel falling by 30–40% without an extraordinary difference in revenue per visitor.

Are AI companies getting more from websites than they give back?

AI companies are extracting enormous amounts of information from websites while sending comparatively few visitors back. For many publishers, the exchange is looking increasingly unfavorable.

Cloudflare introduced a crawl-to-refer ratio to compare how many HTML pages an AI platform retrieves against how many visits it sends back to websites. Crawlers that take far more than they send back are one of the open problems we dig into in our agentic AI market report.

During June 19–26, 2025, Anthropic's measured ratio reached approximately 70,900 crawler requests per referral. Other platforms showed very different results, demonstrating that AI companies don't all use the web in the same way.

Cloudflare's methodology has an important limitation: referral tracking based on browser headers can miss visits originating from native AI apps. The ratio therefore shouldn't be interpreted as a precise count of all people an AI platform sends to websites.

Still, the gap can be enormous. A crawler may retrieve thousands of pages to train a model or prepare responses, while only a small number of users subsequently visit those original sources.

And not all publishers get the same deal. An exclusive financial database may license its information directly to AI developers, while a small advertising-supported publication may receive neither licensing payments nor meaningful referral traffic.

Are AI crawlers making websites more expensive to run?

AI crawlers are increasing infrastructure costs for some websites, particularly when they repeatedly request large amounts of content or trigger expensive database operations.

The Wikimedia Foundation raised this problem publicly in April 2025. Automated crawlers were putting substantial pressure on its infrastructure, especially when requests bypassed caching systems designed around ordinary human reading patterns.

Imagine thousands of people reading the same popular Wikipedia article. A cached copy can serve many of those visits efficiently. A crawler requesting thousands of obscure articles may force the infrastructure to retrieve far more different pages.

DataDome's 2026 security report points to a broader rise in automated pressure. Across its customer network, bad-bot traffic increased 124% over twelve months, while scraping-related activity increased 185.2%. Scraping accounted for 70.9% of classified malicious automation.

Those increases cannot all be attributed to AI. Scraping existed long before modern language models, and many malicious bots have nothing to do with autonomous assistants.

The cost also depends heavily on what the website does. Serving a static article through a content delivery network is generally cheaper than calculating live flight prices, checking inventory or querying a complex database.

Are AI agents becoming a serious cybersecurity problem?

AI agents are becoming a serious security concern because they're increasingly accessing login pages, forms and payment systems, where unauthorized automation can cause much more damage than ordinary crawling.

DataDome detected 605.6 million AI-related requests to sensitive website endpoints during the first half of 2026. These included login pages, forms, shopping carts, payment flows and account-creation pages.

Login endpoints accounted for 51.7% of that activity, equivalent to approximately 313 million requests. Monthly AI requests targeting login pages rose from 11.9 million in January to 99.7 million in June, an increase of more than eightfold.

That doesn't mean there were 313 million successful attacks. DataDome counts requests, including activity that may have been legitimate or blocked before any harmful action occurred.

The security problem is becoming harder because authorized agents and malicious bots can look similar. A personal assistant checking a user's account may generate automated login requests. An attacker testing stolen passwords or probing account systems can produce comparable traffic.

DataDome also reported that AI-agent impersonation increased 45% between February and July 2026. Attackers can pretend to be recognizable AI crawlers to bypass simple defenses based on user-agent labels.

Why are Amazon and other websites fighting AI shopping agents?

Amazon and other major platforms are resisting unauthorized AI agents because those agents can weaken their control over customers, shopping data and transactions.

Amazon's dispute with Perplexity illustrates the conflict. Perplexity's Comet browser allows an AI assistant to navigate shopping websites on behalf of users. Amazon objected to automated access that it had not authorized, leading to litigation and an injunction dispute.

The stakes go well beyond server traffic. If a shopping agent compares products, selects a seller and places an order, the retailer may lose some influence over how customers discover products, respond to recommendations and make purchasing decisions. If shopping agents start choosing the seller, not everyone comes out ahead, and we weigh the likely future winners in our agentic AI report.

That threatens valuable parts of online retailing, including advertising placements, sponsored products, cross-selling and customer loyalty.

Other companies are approaching AI shopping more collaboratively. Walmart and Shopify have explored ways to connect AI-powered discovery with controlled commerce systems, where product information, payments and customer permissions can be handled through authorized integrations.

There's an obvious commercial tension. Blocking every shopping agent could protect the retailer's existing business model, but it could also mean losing customers who increasingly prefer to shop through AI assistants.

Are AI shopping agents actually buying things for people yet?

AI shopping agents can already help people discover and compare products, but fully autonomous purchasing remains far less established. Recommendations are advancing faster than agents that reliably complete purchases from beginning to end.

The distinction becomes clear when we consider the steps involved in an online order.

Finding a suitable product is relatively straightforward. Completing the transaction requires accurate pricing, current inventory, delivery information, customer authentication, payment authorization and sometimes identity verification.

An agent making a mistaken product recommendation creates inconvenience. An agent charging the wrong amount or ordering an unsuitable product creates a financial and customer-service problem.

Retailers have therefore generally been more willing to experiment with AI-assisted discovery than unrestricted autonomous checkout.

Amazon's conflict with Perplexity and the controlled integrations being explored by other retailers show two approaches developing at the same time. One emphasizes protecting existing interfaces; the other tries to make AI-assisted purchases possible through approved systems.

For now, there's insufficient public transaction data to establish what share of online purchases are independently completed by agents without meaningful human involvement.

Is AI making traditional SEO less valuable?

AI is already making traditional SEO less valuable for websites that depend on informational clicks. Search visibility still matters, but appearing in an answer no longer guarantees a visit to the source.

The most exposed content is information that an AI assistant can summarize directly: definitions, basic explanations, straightforward comparisons and answers to common questions.

Pew's research found that Google AI summaries were particularly common for question-based searches. Approximately 60% of queries beginning with question words produced an AI summary, compared with just 8% of one- or two-word searches.

That's a tough situation for publishers who built their traffic around answering specific questions. Their content may help an AI system generate a response, while the user has little reason to open the original article.

Commercial searches offer a different picture. Someone comparing enterprise software, checking a product's availability or choosing a service provider may still need to visit a website before making a decision.

A business generating leads from a small number of high-intent visitors can also tolerate lower traffic better than a publisher earning revenue from advertising impressions.

Can publishers block AI crawlers or make them pay?

Publishers can now restrict AI crawlers and experiment with paid access, but neither approach has yet proved capable of replacing the revenue lost from declining search traffic.

Cloudflare introduced default blocking of unauthorized AI crawlers for newly onboarded domains in July 2025. It also launched Pay Per Crawl, allowing website owners to establish prices for automated access. Blocking bots by default and charging per crawl are new tools, and we check which technologies catch on in our report on agentic AI.

The idea is fairly simple. A publisher that produces valuable information should have more control over whether AI companies can retrieve it and under what conditions.

Direct licensing offers another route. The Wikimedia Foundation, for example, operates Wikimedia Enterprise, which provides paid access to Wikipedia content and services for organizations that need large-scale, reliable data delivery.

However, charging for content creates different problems depending on the business.

A financial database with exclusive information has stronger negotiating power than a general news website covering widely available events. An independent blog may also struggle to enforce restrictions against crawlers that ignore access policies or obtain information through intermediaries.

Blocking creates its own trade-off. Refusing all AI crawlers may protect content from automated extraction, but it can also reduce visibility inside AI assistants.

Website type	Potential value from AI access	Main commercial risk
News publisher	Referrals, licensing, brand exposure	AI answers replacing article visits
Specialist research publisher	Paid data access and professional visibility	Extraction of expensive proprietary research
Product review website	Qualified shoppers	Product comparisons reused without clicks
Online retailer	New customers and sales	Loss of checkout and customer control
Travel booking platform	AI-assisted bookings	Expensive automated searches and fraud
Public documentation website	Wider access to information	Infrastructure costs
Subscription database	Enterprise integrations	Unauthorized access to paid content
Community forum	Discovery of useful discussions	Uncompensated reuse of community content

Will AI agents replace websites with direct integrations?

AI agents are likely to push more internet activity toward direct integrations, especially for purchases and authenticated services. Traditional websites will remain useful, but companies are developing more efficient ways for machines to communicate with one another.

Today, many AI agents use browser interfaces originally designed for people. They navigate menus, read pages, fill out forms and click buttons.

That approach can be inefficient. Websites change their layouts, automated browsers make mistakes, and authenticating a machine as an authorized representative of a user can be difficult.

Structured APIs and authenticated integrations offer an alternative. A retailer can give an approved agent access to product availability, prices and ordering functions without requiring it to navigate every page visually.

The distinction is particularly important for payment systems, banking and travel reservations, where accuracy and authorization matter more than recreating a human browsing experience.

Cloudflare's automated-access controls, commerce integrations and disputes over unauthorized agent browsing all point toward a web where access depends increasingly on who is requesting information and what they're permitted to do.

What would prove that AI agents are really killing the open web?

We would have strong evidence that AI agents are seriously damaging the open web if human traffic and publisher revenue kept falling while AI-generated referrals, licensing payments and transactions failed to replace them.

Several parts of that pattern are already measurable. AI crawlers generate billions of requests, AI-powered search is associated with fewer outbound clicks, and major publisher networks have experienced sharp declines in search referrals.

However, the available evidence is much stronger for falling traffic than for a sustained decline in the number of independent websites, publishers or content creators.

We also need to distinguish businesses that monetize visits from those that monetize purchases or subscriptions. A retailer could receive fewer human visits but more completed transactions through AI agents. An advertising-supported publisher generally has fewer ways to benefit from that change.

The following measurements show how far the evidence currently goes.

Indicator	Available evidence	What we can conclude
Overall automated traffic	51% in Imperva's 2024 dataset	Bots generate a majority of measured requests
AI traffic growth	+82.3% in DataDome's twelve-month study	AI-related automation is expanding quickly
AI share of web requests	1.7% in DataDome's June 2026 dataset	Identifiable AI traffic remains a minority
Google AI Overview click-through	8% with summaries versus 15% without	AI summaries are associated with fewer clicks
Publisher search referrals	-40.2% year over year in Chartbeat's 2026 analysis	Traditional search distribution is deteriorating
AI referral contribution	Less than 1% of publisher pageviews in Chartbeat's analysis	AI traffic hasn't replaced lost search visits
Wikipedia human readership	Approximately -8% in adjusted 2025 comparisons	Direct visits to a major information source have declined
Bad-bot traffic	+124% in DataDome's report	Automated security pressure is growing
Independent publisher survival	No comprehensive internet-wide measurement	Long-term structural damage remains unproven

Is AI agent traffic killing the internet?

AI agent traffic is partly killing the traditional open-web business model, especially for publishers that depend on Google search traffic. But the internet itself is still growing, and claims that autonomous agents are overwhelming or destroying it go well beyond the evidence.

Cloudflare reported that global internet traffic grew 19% in 2025. That is difficult to reconcile with claims of a general collapse in internet usage. People are still using online services heavily, even as the way they discover and consume information changes.

The sharper problem is economic. Publishers are losing search referrals, AI systems are consuming large amounts of web content, and AI-generated visits have not become a meaningful replacement channel for most media websites.

As seen above, Chartbeat's publisher measurements show how severe the pressure has become, while Pew's research helps explain why AI-generated answers are changing click behavior.

There is also a growing struggle over who controls internet transactions. Amazon's dispute with Perplexity, Cloudflare's paid crawling tools and emerging AI commerce integrations suggest that companies are already preparing for a world where machines handle more browsing and purchasing activities.

Our judgment is that AI agents are accelerating a fundamental change in the internet's economics, but the damage is concentrated rather than universal.

Independent publishers, informational websites and advertising-supported services have the most immediate reason to worry. Retailers and transaction-driven platforms face a different challenge: retaining control of customers while benefiting from AI-assisted purchases.

The longer-term danger is a web where original information becomes less profitable to produce, while a small number of AI platforms capture more of the value created by that information.

If publishers cannot recover that value through referrals, subscriptions, licensing or paid access, we could end up with fewer independent sources and a more concentrated internet.

That outcome is increasingly plausible. It is not inevitable, and it hasn't happened at internet-wide scale yet.