What does AI  web traffic look like?

Cloudflare predicts an internet with 1,000x more traffic from AI agents than humans by 2031

The platform has already reported 1,700% growth in agentic traffic YOY. So, I wanted to see what agentic browsing looks like.

THE FUTURE OF INTERNET TRAFFIC

Human intent.
Machine scale.

Cloudflare's CEO predicts a huge shift in who uses the web. This guide makes that prediction visible: one person asks for help, and AI does the searching, comparing, and checking.

AN EXAMPLE GOAL

“Compare three laptops for travel. Check prices, battery life, and delivery.”

Same goal. Different routes. Here's how it works.

HumanAgent

You browse

YOU DO THE LEGWORK
You choose the next step. Open pages, compare what you find, and decide what to check next.

An agent helps

AI DOES THE LEGWORK
You give it a goal. It can check several sources, follow up on missing details, and bring you one answer.

Both routes can use the same sources. The animation illustrates the process, not a fixed traffic multiplier.

01 / FROM INTENT TO ACTIVITY

One task.
A whole web of work.

An AI agent can work through a task for you. When it needs current information, it can search the web, read pages, or get information directly from an online service, such as a store's latest price.[4]

01

Make a plan

An AI agent turns a broad request into smaller questions: which models, which features, which sellers?

02

Fan out

Search and fetch relevant sources. Some systems split the research across agents working in parallel.

03

Check the details

If prices differ between sites or delivery details are missing, the agent can look again.

04

Give you an answer

The agent puts what it found into one comparison. You get one answer, even if it checked many pages.

This is one possible workflow, not a fixed recipe. Some tasks need no web access; others use repeated or parallel searches.[4][8]

02 / HOW THE ACTIVITY ADDS UP

Small tasks.
Big totals.

More people ask agents for help. Each task may involve checking several pages. New questions and tasks people ask an agent to repeat add more activity. Try the controls to see how it adds up.[5]

TRY AN EXAMPLE

Start with a small group, then make it bigger.

How many pages does the agent look at for each task?

New questions or a task someone has asked it to repeat.

THE ACTIVITY BEHIND THE ANSWERS

240
page checks per day

Each bright dot represents 50 page checks.

These are example numbers, not measured internet traffic. Real tasks need different amounts of browsing. A repeated check of the same page counts again.

03 / WHAT’S HAPPENING.
WHAT COULD COME NEXT.

The shift is visible.
The scale is a forecast.

Cloudflare sees a large slice of web activity. Its observations show growing automation. Its CEO's longer-range prediction goes much further.

REPORTED OBSERVATIONMay 2026
Automated traffic has already passed human traffic.

Cloudflare says this happened in May 2026 in the traffic it observes. “Automated” includes AI agents, search crawlers, and other bots. Those groups together have passed humans; this does not mean AI agents alone have.[2]

REPORTED OBSERVATIONSept. 2026
>1,700%

Year-over-year growth in daily requests from AI agents on Cloudflare's network, reported September 30, 2026.[1]

MATTHEW PRINCE'S FORECAST
1,000×

Agentic traffic relative to human traffic
in about five years (around 2031).

Prince forecasted a future in which people keep using the web while agents generate vastly more activity around them. The ratio below makes the scale of that prediction visible.[2]

How big is that difference?

Human activity1 relative unit
Agent activity1,000 relative units
One human activity unit compared with 1,000 agent activity unitsEach filled dot is one equal relative unit. This is an illustration of a forecast, not measured traffic.

At 1,000×, each turquoise unit is matched by 1,000 pink units. Every filled dot uses the same scale. The ring simply highlights the human dot.

The 1,000× figure is a forward-looking extrapolation, not an established outcome. The keynote does not publish a numerical forecasting model or a precise definition of “traffic.” The 10× and 100× views are scale comparisons, not dated predictions.

04 / WHAT CHANGES FOR THE WEB

A page can inform a
decision without a
human visit.

An agent can retrieve a site's information and summarize it in another interface. The person may get an answer without opening the original page. Sometimes the agent sends a visitor or helps a customer act; sometimes it does neither, depending on the context.[1]

People and AI both read your site.

Make your information clear for people and the agents helping them.

More checks may not mean more visitors.

An agent can use your information without sending someone to your site.

You choose what agents can access.

You can allow some automated visitors and limit or block others.[3]

The big change: one person's attention can set far more traffic in motion.

Sources & how to read this guide

Research checked November 17, 2026. Observations are attributed to their source and network. The controls are hypothetical. The forecast is attributed to its speaker. General business implications are editorial interpretation.

  1. Cloudflare, The Internet has a second audience ↗Sept. 30, 2026. Reported agent-request growth and the distinction between retrieval and human referral visits.
  2. Cloudflare Connect 2026, Matthew Prince's keynote ↗See 4:49–6:20 in the transcript. Automated crossover and the five-year, 1,000× agentic-traffic extrapolation.
  3. Cloudflare, New options to manage AI traffic ↗Search, agent, and training behavior categories; site access controls.
  4. Anthropic, How we built our multi-agent research system ↗June 13, 2025. A real example of planning, parallel research, repeated searching, and synthesis.
  5. Cloudflare Agents, Schedule tasks ↗Recurring work is technically possible. A scheduled task generates external traffic only when it actually retrieves or sends information.
  6. Cloudflare Radar, HTTP requests ↗Scope of the HTTP dataset: requests reaching websites served by Cloudflare's CDN.
  7. MDN, Overview of HTTP ↗A page load can fetch many resources. Requests and transferred bytes measure different things.
  8. Anthropic, Web search tool ↗Search use depends on the task, and applications can set a maximum search count.
  9. MDN, HTTP caching ↗Stored responses can be reused; a request reaching a cache need not reach the origin server.


Background: 

I started pulling on this thread because I was curious, and building something visual felt like a more interesting way to learn.
‍
Furthermore, knowing how agentic traffic works strongly overlaps with the trends in AI Search. And getting into the technical details of how the technology is built and behaves is helpful to me.

When I was getting started in SEO, one of the best things that I did for my career was read a research paper on how search engines function and organize information. It was actually part of an investment thesis that I stumbled across in my eTrade account that was available for Google shareholders and prospective investors, that was speaking to Google's economic moat as a business. An unlikely source nonetheless, but invaluable.

Point being, having a firm and evolving understanding of how the tech functions is how I like to learn, and this CloudFlare claim struck my interest, so I wanted to dig deeper into how this would/could work.
‍