On this page
If you run a company of a few hundred people, you have been told the Agentic Web, the part of the web where AI software browses, researches and sometimes buys for a person, will change how you sell: your customers will soon send AI agents to do their buying, so get the company ready.
In every 2025 and 2026 source that separates an agent reading from an agent acting, reading comes out far ahead. Machines make most web requests, and the AI ones mostly read. Business buyers let assistants shape their shortlists and then check with a person. Consumers let AI finish some small purchases and keep most decisions for themselves.
For a company your size, that means your buyers' assistants read about you before your sales team hears the buyer's name, so what those assistants read is now part of how you are found.
It also gives you the shape of your own first agents: an agent does the reading, and a person makes the decision.
The purchase data I found covers US consumer retail and travel. No source measures purchases that business buyers hand to an agent, and most business-buyer figures come from surveys run by companies with a stake in the answer. No source here measures what an agent costs to run or what it returns.
Business Buyers Build Their Shortlists With Assistants
The buyer researching your company is often another business, and that is where assistants reach a company your size first.
Forrester's Buyers' Journey Survey for 2025 found 94% of business buyers using AI in their buying process, up from 89% a year earlier, and more of them named generative AI or conversational search their most meaningful source of information than named vendor websites, product experts or sales.1 61% use private AI tools their own company provides.1
Gartner's survey of 645 B2B buyers, run in August and September 2025, shows the order they use it in.
Gartner: Survey of 645 B2B buyers, August to September 2025 (2026)
In Gartner's survey the assistant's job is mostly the early research, and the person's is the end: reps stayed the most important source when buyers picked a preferred supplier and finalized the purchase.2 Asked who misleads them more, the same buyers split 51% for the chat assistants and 49% for sales reps.2

Forrester's 94% and Gartner's 45% measure different things: any AI use across the buying process, and use of the chat assistants in one recent purchase.
Software buyers have moved furthest. G2, a software review site with an interest in the finding, surveyed 1,076 B2B software decision makers in North America, Europe, the Middle East, Africa and Asia-Pacific in March 2026: 51% now start their research in an AI chatbot more often than in Google, up from 29% in April 2025, and weighing vendors' strengths and weaknesses is the most common use (41%).9
Most software buyers changed their pick because of a chatbot, and most meet wrong answers
share of software buyers
View the data
The chatbot moves buyers while it gets things wrong, which is reason enough to check.
Vlad ShvetsPartner @ QveryG2's 69% is the number I would take into a sales meeting. It counts buyers who ended up choosing a different vendor than they planned to, and nothing in your pipeline records where that change came from. A buyer who changed their pick because of a chatbot did it in a conversation your team never saw, and the only way to answer for that change is to read what the chatbot said.
Forrester's 2026 report on business buying describes the check. Answer engines often deliver incomplete or unreliable information, so buyers seek validation from trusted sources, inside a typical decision that involves 13 people from the buying company and nine outside influencers.10
None of this reaches the seller's own reports. A buyer who reads an answer and never clicks leaves no visit in your analytics, no form and no call log, and Pew's 8% is the shape of that on Google alone.5
The shortlist exists only in the assistant's answer, so the way to see it is to ask your buyers' questions the way they ask them, on the assistants they use, and keep every answer with the pages it read: one reading is a sample, and a run of them shows who gets named and why.
It is an estimate of what buyers see, never a copy of their conversations, because a real buyer's context and follow-up questions change the answer.
Asking by hand for a week gives you a snapshot; telling a pattern from one day's noise takes the same questions asked again and again, the same way.
Agents Pass Tests Faster Than Companies Adopt Them
OSWorld is a standard test of 369 real computer tasks across web and desktop applications, with a human score for comparison. When it was published in April 2024, the best model completed 12.24% of the tasks and humans 72.36%.11 OpenAI's computer-using agent reached 38.1% in January 2025.12 In December 2025, Simular reported 72.6%, just above the human baseline, with a method that runs several agents and keeps the best attempt.13
The best reported OSWorld score rose from 12.24% to 72.6%, past the paper's 72.36% human baseline
share of tasks completed
The three scores come from different setups and versions of the test, so read the columns as direction. The score also depends on the test.
Agent success rates run from 87% on simple live-site tasks to 30% in a simulated company
share of tasks completed
View the data
Short tasks on public websites score highest, and work that strings many steps across tools scores lowest: TheAgentCompany's authors found simpler tasks within reach and long, multi-step ones beyond it.14
Piotr BabelPartner @ QveryWhen we plan a build, the 30% from TheAgentCompany matters more than the 72.6% on OSWorld. A long job that crosses several tools fails somewhere in the middle, and nobody sees where. So we cut the job into short reading steps, each checked against something written down, such as your criteria for a supplier quote, and the agent stops and hands the step to a person when it does not match.
Companies moved more slowly. In McKinsey's 2025 survey, 23% of respondents said their organizations were scaling an agent somewhere and another 39% were experimenting.15
at most of McKinsey's respondents were scaling AI agents in any single business function15
Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, for escalating costs, unclear business value or inadequate risk controls.16
Gartner's three reasons point at decisions a company can make before anything is built: what the job is and what doing it is worth, who answers for the agent's output and what it may do on its own, and what gets counted against what the agent costs to run.
Machines Make Most Web Requests
Automated traffic made up more than 53% of all web traffic in 2025, up from 51% the year before, and human activity fell to 47%, according to Imperva's 2026 Bad Bot Report.3 That is a share of requests, not of people, and it counts every kind of bot: search crawlers, scripts and attack bots as well as anything built on a language model.
Cloudflare measures the AI share separately, and on its network it is small. Most of it is crawling to fill a training set.
A bot that loads a page because a person asked an assistant a question made up, together with crawling that does not say who sends it, less than 5% of AI bot traffic on Cloudflare's network in that July window.17 It is also the slice growing fastest: that kind of crawling grew more than 15 times over during 2025, from a small base, and at its peak training crawling still ran 32 times larger.4
Fastly saw the same shape on its own network from mid-April to mid-July 2025: AI crawlers were almost 80% of AI bot traffic, and OpenAI's bots made 98% of the requests from bots that load a page because a person asked an assistant a question.18
The reading sends few people on. Cloudflare's crawl-to-refer ratio compares how often a platform crawls a site with how often it sends that site a visitor.
In 2025 Anthropic's ratio peaked at 500,000 crawls per visitor referred and OpenAI's at 3,700, while Perplexity's stayed below 200 from September and Google's moved between about 3 and 30.4
People behave the same way when the answer sits on the search page. Pew Research tracked the browsing of 900 US adults in March 2025: when Google showed an AI summary, they clicked a traditional result in 8% of visits, against 15% without one, clicked a link inside the summary in 1%, and ended their browsing on 26% of those pages, against 16% without a summary.5
Sorted by what each visit can tell a seller, the machines fall into three kinds.
| Visitor | What it does | What it tells a seller |
|---|---|---|
| Crawler | Reads pages in bulk on its own schedule | Your pages are being read for an index or a training set |
| Bot loading a page for a person | Loads a page because someone asked an assistant | A real question was asked; the person may never visit |
| Acting agent | Clicks, fills in forms or checks out for a person | A real attempt to do business; not counted apart here |
For an owner, a jump in traffic is not a jump in prospects until you know how much of it came from machines.
In finance the machines are often hostile as well: Imperva attributes 24% of all bot attacks and 46% of break-ins to a customer's account by a stolen login in 2025 to financial services.3 For a finance company, deciding which agents to let in is a security decision before it is a sales one.
Most Consumers Keep the Purchase Decision
Transactions are counted in consumer commerce, so that is where to test whether agents do the buying.
Adobe's retail data covers more than a trillion visits to US retail sites. AI-referred traffic grew 393% year over year in the first quarter of 2026, after rising 693% over the 2025 holiday season,19 and grew 125% between April and June 2026.20
Those visitors buy on the retailer's own site, and now more often than visitors from other channels: in March 2026, AI-referred visits converted 42% better than other traffic, a year after converting 38% worse.19
Travel runs a step behind. AI traffic to US travel sites rose 119% in July 2026 and converted 1% lower than other channels, and users click through to the brand's own site to book.21
Gartner's consumer surveys show where people draw the line. At most 11% of US consumers would let AI make a purchase decision, and only in low-stakes categories such as personal care and household supplies, while 31% would let it narrow the choices for household supplies and 28% for personal electronics.7
The help creates work too: 54% of consumers who used AI on a recent purchase had to double-check everything it told them.7
The strongest evidence that agents do some buying comes from Menlo Ventures' July 2026 survey of 5,067 US adults, which asked shoppers who use AI what they had let it do in the past six months.6
The closer a step sat to the payment, the fewer shoppers handed it to AI
share of AI-using shoppers
View the data
Only the last bar is an agent finishing the job, and it is the shortest. Separately, 32% of AI users have let AI act for them at least once without a final sign-off, across purchases, emails and connected accounts.6
Menlo asked AI-using shoppers what they had done, and Gartner asked US consumers in general what they would allow, so the 26%6 and the 11%7 answer different questions.
Morgan Stanley's estimate that roughly 23% of Americans made a purchase using AI in the past month reads the same way: using AI, not handing it the decision, and most often for groceries and consumer packaged goods.22
If your clients are people rather than companies, as a law firm's or a brokerage's are, Menlo has a number for you too: for legal advice, 59% of US adults prefer a human expert and 21% prefer AI.6
ChatGPT's Agent Checkout Now Sends the Buyer to the Merchant's Site
Agent buying has to pass through the payment rails, and each one here either keeps the buyer's confirmation or permission in the last step or moves the purchase to the merchant's own site.
| Rail | What the agent can do | Where the person stays |
|---|---|---|
| OpenAI (ChatGPT) | Discover and compare products in ChatGPT | The shopper pays on the merchant's own site or app8 |
| Visa | Agent payments from comparison to checkout, still being deployed | Payment completes with the user's permission, under spending limits and approvals23 |
| Cloudflare | Signed agents let a website check which agent is knocking24 | A signature proves which agent sent a request, not that its person agreed to buy24 |
OpenAI's checkout launched in September 2025 for US users buying single items from US Etsy sellers, with over a million Shopify merchants announced as coming soon, and the launch post says users explicitly confirm each step before any action is taken.25 Its developer documentation still describes that checkout for approved partners, with the merchant accepting or declining each order and charging the payment on its own systems.26
The merchant page sets a different direction. OpenAI says it is moving away from a standalone Instant Checkout in favor of merchant-owned checkout, merchants send shoppers to their own site or app by default, and shopping is live for ChatGPT users in the US.8
For a seller, the page a buyer lands on after the answer is part of the purchase again. For a consumer brand, the work on that page starts with the product records ChatGPT reads.
Three forecasts put a size on agentic commerce, and each measures something different.
| Forecast | Figure | What it counts |
|---|---|---|
| Morgan Stanley | $190B to $385B of US e-commerce by 203022 | Spending by agentic shoppers |
| McKinsey | Up to $1T in US B2C retail by 203027 | Sales an agent set up |
| Gartner | 90% of B2B buying by 202828 | Buying carried out through an AI agent |
They cannot be added into one number. Gartner's is the one that speaks to a company selling to companies, and the same firm's survey of 645 buyers found 69% preferring to check AI's answers with a rep.2
Buying carried out through an AI agent is not buying decided by one. On Gartner's own numbers, the agent sits in the research and the comparison, and the buyer still wants a person before committing.
Every source here draws the same line between reading and deciding, and it is where your own first agents should stop until their work has earned more.
Frequently asked questions
What is the Agentic Web?
The part of the web where AI software browses, researches and sometimes buys for a person. In 2025 most of that activity was reading: crawling to train models made up nearly 80% of AI bot crawling on Cloudflare's network, and bots acting for a person were a small slice.17
Are AI agents buying from companies yet?
Consumers have started, mostly at low stakes: 26% of AI-using US shoppers in Menlo Ventures' survey have let AI complete a purchase,6 while at most 11% of US consumers would let it make the decision.7 No source here measures purchases business buyers hand to an agent, and 69% of B2B buyers prefer to check AI's answers with a sales rep.2
Can shoppers still buy directly inside ChatGPT?
OpenAI says it is moving away from a standalone Instant Checkout in ChatGPT: shoppers discover and compare products there, and purchases are completed on the merchant's own website or app.8
How much web traffic comes from bots and AI agents?
Imperva measured automated traffic at more than 53% of all web traffic in 2025.3 A separate measurement on Cloudflare's network puts AI bots other than Googlebot at an average of 4.2% of HTML requests in 2025.4
Do B2B buyers use AI assistants to find vendors?
Most do. 94% of business buyers in Forrester's survey use AI in their buying process,1 45% of the B2B buyers in Gartner's survey used the chat assistants to gather information on vendors and products,2 and 51% of software buyers in G2's survey start their research in a chatbot more often than in Google.9
Do buyers trust what AI chatbots tell them about vendors?
Not fully. Asked who misleads them more, the B2B buyers in Gartner's survey split 51% for the chat assistants and 49% for sales reps,2 and 64% of software buyers in G2's survey meet inaccurate recommendations often.9 Forrester's 2026 report says answer engines often deliver incomplete or unreliable information.10
Do AI answers send buyers to company websites?
They send fewer than plain search results do. Google users in Pew Research's study clicked a traditional result in 8% of visits when an AI summary appeared, against 15% without one,5 and AI platforms crawl sites far more often than they refer visitors to them.4 The AI-referred visitor who does arrive at a US retail site now converts 42% better than other traffic.19
How can a company see what AI assistants tell buyers about it?
Ask the questions your buyers ask, on the assistants they use, the same way every day, and keep each answer with the pages it cites. Compare every statement about you with a written fact sheet, then fix wrong facts where they live: your own pages first, then a request to the owner of any other page.
Which purchasing decisions should stay with a person?
The choice of supplier and the signature on the order. An agent can read and compare quotes against your criteria while the person who buys picks and signs, the way Visa's agent payments run under spending limits and approval workflows23 and ChatGPT's shoppers complete the purchase on the merchant's own site.8
Can AI agents do real computer work reliably?
They can on well-defined tasks, and the scores are rising. The best reported score on OSWorld, a standard test of computer tasks, reached 72.6% in December 2025,13 against 72.36% for humans,11 while the best agent in a simulated software company finished 30% of its tasks.14
Why do agentic AI projects get canceled?
Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, for escalating costs, unclear business value or inadequate risk controls.16 Each reason points at a decision made before building: what the job is worth, who answers for its output, and what gets counted.
How big will agentic commerce be?
The forecasts range from $190 billion to $385 billion of US e-commerce by 2030 (Morgan Stanley)22 to 90% of B2B buying carried out through AI agents by 2028 (Gartner),28 and each counts something different.
Does this apply to logistics, real estate, legal or finance companies?
Partly. None of the sources here counts agent purchases in these four industries. Financial services took 24% of all bot attacks in 2025,3 and 59% of US adults prefer a human expert to AI for legal advice.6 Checking what assistants say about you and reading supplier quotes are jobs in all four.
How often do Google users click a link inside an AI summary?
Rarely: in Pew Research's study of 900 US adults' browsing in March 2025, they clicked a link inside the summary on 1% of visits to pages that had one, and ended their browsing on 26% of those pages, against 16% without a summary.5
Do AI chatbots change which software vendor buyers pick?
Often, by buyers' own account: in G2's March 2026 survey of 1,076 software decision makers, 69% chose a different vendor than planned because of a chatbot, and 85% think more highly of a vendor a chatbot recommends.9
Sources
- Forrester: B2B Buyers Make Zero-Click Number One (2026)
- Gartner: Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights (2026)
- Imperva (Thales): Bad Bot Report 2026: The Internet Is No Longer Human and It's Changing How Business Works (2026)
- Cloudflare: The 2025 Cloudflare Radar Year in Review (2025)
- Pew Research Center: Google users are less likely to click on links when an AI summary appears in the results (2025)
- Menlo Ventures: 2026: The State of Consumer AI (2026)
- Gartner: Gartner Survey Finds Consumers Want AI Shopping Help, But Not AI Purchase Decisions (2026)
- OpenAI: Power product discovery in ChatGPT (merchant page) (2026)
- G2 via PR Newswire: New G2 Research: Half of B2B Software Buyers Now Start Their Research With AI Chatbots (2026)
- Forrester: Forrester's 2026 Buyer Insights: GenAI Is Upending B2B Buying As Leaders Face Mounting Pressure To Justify Every Dollar Spent (2026)
- arXiv (Xie et al.): OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (2024)
- OpenAI: Computer-Using Agent (2025)
- Simular: Simular's Agent S Outperforms Humans on OSWorld Benchmark (2025)
- NeurIPS 2025 (Xu et al.): TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (2025)
- McKinsey & Company: The state of AI in 2025: Agents, innovation, and transformation (2025)
- Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (2025)
- Cloudflare: A deeper look at AI crawlers: breaking down traffic by purpose and industry (2025)
- Fastly: New Fastly Threat Research Reveals AI Crawlers Make Up Almost 80% of AI Bot Traffic (2025)
- Adobe: Adobe report: U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines (2026)
- Adobe: Now available: Product discovery on LLM surfaces in Adobe Commerce (2026)
- Adobe: Adobe report: U.S. consumers are embracing LLMs to make travel plans, but many brands have AI visibility gaps (2026)
- Morgan Stanley: Here Come the Shopping Bots (2025)
- Visa: Enabling AI agents to buy securely and seamlessly (2026)
- Cloudflare: The age of agents: cryptographically recognizing agent traffic (2025)
- OpenAI: Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol (2025)
- OpenAI: Key concepts: Agentic Commerce (developer documentation) (2026)
- McKinsey & Company: The agentic commerce opportunity: How AI agents are ushering in a new era for consumers and merchants (2025)
- Gartner: Gartner IT Symposium/Xpo 2025 Orlando: Day 2 Highlights (2025)