AI agent traffic statistics for 2026: the H1 numbers and what to do about them

Published July 24, 2026 · Last reviewed July 24, 2026

The question a lot of site owners started asking in the first half of 2026 is simple: how much of my traffic is AI agents now? The reports that answer it have finally landed, and the numbers are large enough that “some” is no longer an acceptable answer. Agent requests are now counted in the billions per quarter, and on at least one major network they already outnumber human requests.

A raw count, though, is the least useful thing to take from these reports. The number that decides what you do next is the mix. Most agent traffic is a crawler feeding a model, and you cannot sell to a crawler. A smaller and faster-growing stream is agents acting for a real person, and those are the ones worth answering. Every honest read of the 2026 data has to separate the two before it means anything.

This is a synthesis of the public H1 2026 datasets, sorted, sourced, and dated. Each figure comes from a different vendor measuring a different part of the web, so the goal is not one grand total. It is a clear picture of the scale, the trend, and the single distinction the headlines skip.

The H1 2026 numbers, in one place

Four datasets carry most of the signal for the first half of the year. They do not add up, because each measures a different network with a different definition. Read them side by side, not stacked.

DataDome, Q2 2026. DataDome’s AI Traffic Report for the second quarter, published July 16, 2026, recorded 17.7 billion AI agent requests across April, May, and June, up 45% from 12.2 billion in the first quarter. The report draws on the vendor’s network of more than 400 enterprises. DataDome’s own report page blocks automated fetches, so these figures are quoted from its published writeup and the Businesswire syndication carried by Morningstar and others. It is vendor data from one security network; label it as such, but it is the freshest large public snapshot of the year.

Cloudflare Radar, early June 2026. On Cloudflare’s network, automated requests crossed the halfway line: about 57.5% of HTML page requests were automated against 42.5% human, per Matthew Prince in early June 2026. This is the crossover point people quote when they say bots now outnumber humans, and it is worth pinning to Cloudflare’s specific network and metric rather than treating it as a law of the whole web.

TollBit, publisher network, through end of 2025. TollBit’s State of the Bots, measuring around 400 publisher sites, reported roughly one AI bot visit for every 31 human visits by the fourth quarter of 2025, up from one per 200 at the start of that year. In the second quarter of 2025 the ratio was about one per 50, and human page requests on those sites fell 9.4% quarter over quarter. This is publisher traffic, not B2B software, and the freshest TollBit figure is a 2025 reading, so it sets the trajectory into 2026 rather than measuring it.

Salespeak, B2B SaaS slice, February 2026. For a narrower and more commercial view, the vendor Salespeak reported 640,000 AI agent visits to B2B SaaS sites over 30 days, with ChatGPT-linked crawling making up about 91% of that volume. The data was captured in January and February 2026, so it is now the stale baseline in this set, useful mainly as an early B2B-specific data point against the newer macro numbers.

Put together, the shape is not subtle. Agent traffic is measured in billions of requests per quarter, it has crossed the human line on at least one major network, and it is climbing fast across every source that publishes a trend.

Why the growth rate matters more than the count

The 17.7 billion figure will be out of date by the next report. The 45% quarter-over-quarter jump is the part that tells you where this is going, and the composition underneath it matters even more.

In DataDome’s Q2 data, the growth was not evenly spread. Meta’s training crawler grew about 74% quarter over quarter, and its retrieval crawler grew about 163% in the same window. The agents generating the most traffic in Q2 were not the same ones that led in Q1. That churn is the real lesson: this is not a stable population you can profile once and forget. A rule you write in April to handle the top crawlers can be pointing at the wrong ones by July.

So the useful metrics are not “how many agents hit my site.” They are the rate of change and the mix. A site seeing agent traffic double every two quarters is on a different planning horizon from one seeing it grow ten percent, and the response depends far more on which agents are arriving than on how many.

Two kinds of agent traffic, and only one is a buyer

Here is the distinction every one of these reports contains and most summaries flatten. Agent traffic splits into two populations that happen to share an internet connection.

The first is crawlers: training bots and retrieval bots that read your pages to feed a model. They do not represent a person making a decision. In DataDome’s Q2 numbers, Meta’s two crawlers alone generated the majority of agent volume, on the order of 9.1 billion requests, while sending almost no visitors back. That is retrieval at scale, not demand. It costs you bandwidth and gives you nothing directly, which is exactly why edge blocking exists.

The second population is transactional agents: an assistant acting for one human or company, arriving to answer a question, compare options, qualify a vendor, or complete a task. This stream is smaller but it is where the intent is. On the referral side, ChatGPT-linked traffic dominated, commanding roughly 80 to 88% of AI-driven referrals each month in DataDome’s Q2 window and growing about 17% quarter over quarter. Referrals are a rough proxy, but they point at the same thing: a rising slice of agent activity ends with a real person or their assistant actually engaging.

This maps directly onto a distinction we have written about before. Making your site readable to crawlers is one job, and making it addressable to the agents that arrive to transact is a different one. We laid out the difference in agent-readable versus agent-addressable. A traffic report tells you the crawlers found you. It does not tell you the buyers could do anything once they did.

The number nobody can fully trust yet

There is a measurement problem underneath all of these statistics, and honesty requires stating it plainly. Most agents do not identify themselves accurately.

DataDome’s earlier 2026 analysis found that about 80% of AI agents do not declare themselves properly, and that 79.7% of roughly 698,000 tested sites let a spoofed ChatGPT-style user agent through. Meta-ExternalAgent was the most-impersonated agent in that sample, at 16.4 million spoofed requests, followed by ChatGPT-User at 7.9 million. User-agent strings are trivial to fake, so a large share of what any tool labels “ChatGPT” or “Meta” is an educated guess.

The practical consequence: every clean-looking pie chart of agent versus human traffic is part measurement and part estimate. It does not make the reports useless, the orders of magnitude are real, but it does mean you should distrust any breakdown presented to three decimal places. The fix is not better guessing. It is cryptographic identity, so a site can verify that an agent calling itself ChatGPT actually is one. That is the whole point of emerging signing standards, which we covered in how your site tells real AI agents from spoofs. Until they are widely adopted, read every agent-traffic number with a margin of error attached.

What a B2B site actually does about it

Detection is where most teams stop, and it is only the first of three steps. The reports hand you a defensive instinct: agents are flooding in, so filter them. For crawlers you do not want, that instinct is correct, and edge tools already do it well. For the buyers, it is a mistake that looks like prudence.

Think of it as three moves, in order. Detect: know an agent when it arrives and separate the crawlers from the transactional visitors. Greet: give the transactional agent something it can actually use, a clear, structured answer to the question it came with, instead of a JavaScript-heavy page and a contact form it cannot fill. Convert: let that agent qualify fit and route to a human when there is real intent, so a promising conversation does not dead-end at a lead capture no agent will complete.

Most sites are stuck on detect. They can tell you their agent traffic went up and to the right, and they have blocked some of it, and that is the end of the plan. The datasets in this piece are a signal to move past that. A rising share of your visitors will be agents sent by real buyers, and the sites that win are the ones that answer them, not just measure them. We walked through what changes when the buyer on the other end is an agent in what agentic browsers change for your website.

What these numbers do not tell you

A stats anchor is only as honest as its caveats, so here are the limits of everything above.

Each figure is a single vendor measuring its own network. DataDome sees the traffic hitting the sites it protects; Cloudflare sees its own edge; TollBit sees publishers; Salespeak saw a B2B SaaS slice months ago. None is a census of the web, and their definitions of “agent” do not match, which is why the numbers cannot be added or directly compared. Publisher data in particular says little about a B2B software or services site, where the buyer mix is completely different. And because of the identification problem above, even within one network the split between agent and human is estimated, not counted.

Read these as direction and magnitude, not precision. The direction is not in doubt: agent traffic is large, growing quickly, and increasingly transactional. The precise share for your specific site is something only your own logs, with real agent verification in place, can eventually tell you.

How this connects to Tobira

Tobira sits on the transactional side of that split, and it is complementary to the crawler-facing work, not a replacement for it. Readiness tools and edge controls handle how agents read and how crawlers are filtered. Tobira adds the other half: a way for the agents that arrive to transact to reach a representative tied to a human name, qualify fit through a real conversation, and route to a person only when both sides consent. That is the agent-addressable and networked layer, which is a different job from making a site machine-readable. Per Tobira’s founder update in late May 2026, the network listed 641 public agents, including 102 business agents, with more than 1,000 signups; it is free during beta, with a paid tier planned. For the underlying distinction between reading a site and reaching its representative, see agent-readable versus agent-addressable.

What to remember

• Agent traffic is now measured in billions of requests per quarter. DataDome logged 17.7 billion in Q2 2026, up 45% from Q1, and Cloudflare reported automated requests at 57.5% of HTML traffic in early June.
• The growth rate and the mix matter more than any single count. The composition shifts fast: Meta’s retrieval crawler grew about 163% quarter over quarter in DataDome’s Q2 data.
• Most of the volume is crawlers feeding models, not buyers. Meta’s two crawlers alone drove roughly 9.1 billion Q2 requests while sending almost no visitors back.
• The numbers carry a real margin of error, because about 80% of agents do not identify themselves properly and user-agent strings are easy to spoof.
• Detection is step one of three. The sites that benefit greet and convert the transactional agents, rather than only measuring and blocking.

Frequently asked questions

How much of web traffic is AI agents in 2026?

It depends on whose network you measure, because there is no shared census. On Cloudflare’s network, automated requests reached about 57.5% of HTML traffic against 42.5% human in early June 2026. DataDome logged 17.7 billion AI agent requests across April to June 2026, up 45% from 12.2 billion in the first quarter. TollBit, measuring around 400 publisher sites, reported roughly one AI bot visit per 31 human visits by the end of 2025, up from one per 200 at the start of that year. The figures differ because each vendor measures a different slice of the web with a different definition of agent, so treat them as directional rather than a single audited number.

What is the difference between a crawler and a buying agent?

A crawler reads your site to feed a model, either training data or real-time retrieval, and it does not act for a specific person. A buying agent visits on behalf of one human or company to answer a question, compare options, qualify fit, or complete a task, and there is a real decision behind it. In DataDome’s Q2 2026 data, Meta’s two crawlers generated the majority of agent volume, roughly 9.1 billion requests, while sending almost no visitors back. That bulk is retrieval, not demand. The smaller stream of transactional agents is the one a business site should care about converting.

Why can I not just trust my analytics to count AI agent traffic?

Because most agents do not identify themselves honestly. DataDome’s earlier 2026 analysis found that about 80% of AI agents do not declare themselves properly, and that 79.7% of roughly 698,000 tested sites let a spoofed ChatGPT-style user agent through. Meta-ExternalAgent alone accounted for 16.4 million spoofed requests in that sample. Until agents authenticate cryptographically, any breakdown of agent versus human traffic is part measurement and part estimate.

Should I block AI agents from my website?

Blocking makes sense for unwanted training crawlers that consume bandwidth and return nothing, and most sites already do some of this at the edge. It is the wrong reflex for buying agents. An agent arriving to compare vendors or qualify a service is a lead, and blocking it is closer to unplugging your phone than stopping spam. The useful posture is to separate the two: control the crawlers you do not want, and give the transactional agents a clear way to get answers and reach a human.

Is AI agent traffic growing or leveling off in 2026?

Growing, and the growth rate is the more important figure. DataDome reported a 45% quarter-over-quarter jump in agent requests into Q2 2026, with the composition shifting: Meta’s training crawler grew about 74% quarter over quarter and its retrieval crawler about 163%. The agents driving the most traffic in Q2 were not the same ones leading in Q1, which means any snapshot dates quickly. Watch the trend and the mix, not one quarter’s headline count.

Sources

• DataDome, “The AI Traffic Report Q2 2026: Agentic Traffic Surged 45%, With Meta Taking the Lead,” July 16, 2026 (report page; figures corroborated via Businesswire syndication on Morningstar and Yahoo Finance).
• DataDome, earlier 2026 AI Traffic Report edition, on agent identification and user-agent spoofing (about 80% not identifying properly; 79.7% of ~698,000 sites; Meta-ExternalAgent 16.4M spoofed requests).
• Cloudflare Radar, automated versus human HTML traffic share (about 57.5% automated), Matthew Prince, early June 2026.
• TollBit, “State of the Bots,” publisher-network data through Q4 2025 (about one AI bot visit per 31 human visits; one per 200 at the start of 2025; human page requests down 9.4% quarter over quarter in Q2 2025), via Press Gazette.
• Salespeak, B2B SaaS AI agent traffic analysis, 640,000 visits over 30 days with ChatGPT-linked crawling around 91%, data captured January to February 2026.