Two partners at Andreessen Horowitz argued on November 11, 2026 that companies which spent years building defences against automated traffic will now have to compete for its business. In an essay titled "Welcome to the B2A Economy," published on the firm's a16z News newsletter, Olivia Moore and Justine Moore described AI agents as the web's primary users and set out who they expect to gain as those agents begin choosing products, subscriptions and places to spend money on behalf of people.

In Short

Two a16z partners published an essay on November 11, 2026 arguing that AI agents, rather than people, are becoming the main users of the web, and that businesses will have to sell to them the way they once sold to searchers and app users. It matters to retailers, software companies and advertisers because agents, led by Meta's Muse and OpenAI's products in HUMAN Security's October data, are starting to decide which sites, tools and paid plans get chosen. Several of the essay's headline claims run ahead of its own charts: the traffic chart counts all bots rather than AI agents, and the content chart shows AI and human writing at parity, not an AI majority. The practical change the authors describe is that being picked by an agent becomes a distribution channel in its own right, a step beyond appearing in an AI answer.

The argument: from walls to storefronts

The essay opens with a flat statement. "The primary users of the Internet are no longer human. They're AI agents," the authors wrote. They then claim that agents "make up the majority of web traffic, write more than half of new articles published online, and generate most of the code pushed to production."

From there the piece builds a historical frame. Each earlier platform shift, according to the authors, paired a new capability with a new way for businesses to reach customers. An illustration in the essay lines up four eras: phones, with Yellow Pages directories described as "the original search engine"; the web, with digital products and link aggregators such as Google and Yahoo; mobile, with location-based, on-demand products distributed through the App Store; and AI, with "products that deliver outcomes, not information, through agents."

The authors argue that AI has delivered on the first half of that pairing and not yet on the second. "AI has overperformed on the first and underperformed on the second, which is a big reason consumer adoption has stayed so limited outside ChatGPT," they wrote. The early stage of a platform shift, in their description, is one where "adoption is measurable, but the business models it enables mostly aren't visible yet."

The commercial claim follows. As agents move from answering questions to completing tasks, according to the essay, they will decide "which products to use, which services to subscribe to, and where to spend money on our behalf." Businesses that treat agents as first-class customers "have a shot at a wave of growth," while those that do not "may find themselves under threat." The line that gives the piece its subtitle is blunt: "For years, retailers and platforms built walls to keep bots out - now, they will have to fight for bots' business."

The authors also reject a narrower label. "Calling it 'AI commerce' undersells the opportunity," they wrote, arguing that agents will drive transaction volume to physical goods and services but also "reshape how traffic, tokens, and revenue are routed to digital products." The essay is described as the first in a series. Over the coming weeks, according to a16z, the firm plans to publish data on what consumer agent usage looks like, how agents decide which sites and products to send traffic to, how infrastructure is developing around them and what hurdles remain before agents can complete transactions.

The essay closes with what the authors call their "best current mapping of the market," covering both agents and the infrastructure beneath them. The piece opens with a note congratulating Counsel Health on its acquisition by OURA, and ends with the newsletter's standard disclaimer, which states that the content may draw on third-party sources and that "a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information." The same disclaimer notes that investments or portfolio companies may be mentioned.

What the seven charts show

Most of the essay's weight sits in its charts, which pull together figures from security vendors, analytics firms, a bank, a model marketplace and individual companies. Each has a different method and base.

Bots and the web

The first chart, titled "Bot vs. Human Web Traffic Share," shows bots rising from 42% of web traffic in 2021 to 47% in 2022, 49% in 2023 and 51% in both 2024 and 2025. The 2026 bar, at 54%, is labelled a forecast. The sources listed are NP Digital (January 2026), Security Magazine, Imperva and Soax. PPC Land has tracked similar measurements from other vendors: Cloudflare Radar data for the seven days to June 5, 2026 showed bots at 57.4% of web traffic to HTML content, and a July 2026 Experian release cited Imperva's figure of 53% of web traffic as automated.

Articles

A second chart, "AI Now Writes as Many Online Articles as Humans," plots the quarterly share of articles classified as primarily AI-generated against those classified as primarily human-written, from the first quarter of 2020 to the first quarter of 2026. The AI line sits near zero until ChatGPT's launch in November 2022, which the chart marks, then climbs steeply through 2023 and 2024. By early 2026 the two lines meet at roughly 50%. The source is an updated Graphite study dated May 15, 2026, averaging classifications from Pangram, Copyleaks and GPTZero. An earlier Graphite study from October 2025 had already put the share of new written content likely produced by AI at 50%.

Tokens

The third chart, "Agents Are Using Far More Tokens Than People," draws on OpenRouter rankings accessed November 10, 2026. It shows a seven-day average of tokens processed on the platform, split into agentic, human and mixed usage. According to the chart's annotation, agentic tokens overtook human tokens on February 6, 2026; the agentic line then peaks at 7.3 trillion tokens, which the chart describes as a 14x increase. Human usage, by contrast, peaks at roughly 1.5 trillion. PPC Land reported in August that a16z's "Charts of the Week" newsletter had circulated the same OpenRouter series, showing agent usage up fourteenfold since February and agents consuming close to five times as many tokens as human users.

Agent traffic

The data most directly tied to the essay's thesis comes from HUMAN Security. According to the essay, HUMAN "saw agentic browser traffic grow 8x between August and October, with Muse driving the largest share of activity (60%) and OpenAI products reaching a combined 30%."

The accompanying chart, "AI Agent Traffic Grew 8x in Two Months," indexes monthly agentic web traffic to February 2026 at 100. The line barely moves through the spring and summer, reaches 204 in August, jumps to 631 in September and ends at 1,626 in October. By PPC Land's calculation, October's level is about 16 times February's, and the August-to-October ratio of 1,626 to 204 is 7.97, consistent with the 8x headline. The chart's footnote matters: the index was "calculated from published growth rates" in HUMAN's April, September and October 2026 reports, the March to July values "are visually estimated from the supplied chart," and the underlying data covers "observed requests on HUMAN customer sites, including allowed and blocked activity."

The second HUMAN chart, "Meta and OpenAI Drive ~90% of AI Agent Traffic," breaks down October 2026 agent requests on HUMAN customer sites, from a report published November 9:

  • Meta Muse: 59%
  • ChatGPT Agent: 16%
  • Codex Browser: 8.8%
  • Codex Chrome extension: 5.2%
  • Perplexity Comet: 3.6%
  • Claude Chrome extension: 3.5%
  • Other: 2.4%
  • Google Agent / Gemini: 0.8%
  • xAI Grok Agent: 0.7%

The three OpenAI products add up to exactly 30%, and Meta and OpenAI together come to 89%, which the chart rounds to roughly 90%. The distribution has shifted sharply in a short time. In May 2026, HUMAN's data showed Comet at 47% of agentic traffic, OpenAI's Atlas at 20.3% and the Claude Chrome extension at 18.6%. OpenAI has since folded Atlas into a new ChatGPT Work agent, and Muse arrived. A HUMAN blog post dated September 28 had put Muse at an average of 72% of daily AI agent requests, with daily volume 5.2 times that of ChatGPT Agent. October's 59% is lower than that late-September average, though the two figures come from different HUMAN publications and periods.

Referral value

On the commercial side, the essay cites Shopify, which reported that AI-driven traffic to its stores "grew eightfold year over year in Q1 '26, while orders from AI-powered searches grew nearly thirteenfold." Shopify's later numbers point to slower growth from a larger base: in its third-quarter 2026 results, the company said AI-referred orders rose 3.3 times year over year and AI traffic to merchant storefronts 2.6 times, without quantifying the base.

The authors argue that agents should be a better referral source than websites ever were, because they hold "deep and specific data on the user and their needs." According to the essay, AI referral traffic to consumer sites "is higher engagement and converts better to revenue than non-AI traffic, with a lower bounce rate."

The supporting chart, sourced to Adobe and J.P. Morgan and dated October 27, 2026, compares AI-referred traffic with non-AI traffic in three months:

  • Year-over-year traffic growth: 269% in March 2026, 138% in May and 62% in July
  • Engagement uplift: 12%, 15% and 14%
  • Time spent uplift: 48%, 53% and 59%
  • Pages viewed uplift: 13%, 23% and 22%
  • Conversion uplift: 42%, 54% and 60%
  • Relative bounce rate: -32%, -36% and -34%
  • Revenue per visit uplift: 37%, 53% and 53%

The chart supports the essay's claim on conversion and revenue per visit, both of which widened over the period. It also shows something the text does not mention: year-over-year growth in AI-referred traffic fell from 269% to 62% in four months. The March figure matches the 269% Adobe reported at Adobe Summit in April.

Not every measurement has pointed the same way. A study of 973 ecommerce sites covering August 2024 to July 2025 found that ChatGPT referrals made up under 0.2% of sessions and converted worse than organic search. And Walmart said in March that checkout inside ChatGPT converted three times lower than click-out links, after which OpenAI shut instant checkout down and moved product feeds into its Ads Manager.

The last data chart, from Similarweb (August 2026), tracks the share of US desktop shopper journeys that visit an AI platform before a transactional site. It rises from 4.5% in July 2024 to a peak of about 11.5% in May 2026 and ends at 11.4% in June 2026, with a dip from roughly 11% in October 2025 to roughly 10% in December 2025. The essay rounds this to "an estimated 11% of shopper journeys in the U.S." and describes it as a figure from before agents arrived. By PPC Land's calculation, the share grew about two and a half times over 23 months.

Where the essay runs ahead of its charts

Read against its own exhibits, the essay's opening sentences are stronger than the data beneath them.

The claim that agents "now make up the majority of web traffic" rests on a chart that measures bots, not AI agents. Bots in that series include crawlers, scrapers, monitoring tools and malicious automation, most of which predate consumer AI agents. HUMAN's own annual report, published in April, described agentic traffic as growing 7,851% year over year from a low 2024 base, with checkout pages at 2.31% of agentic activity. Rapid growth from a low base is a different claim from a majority of all traffic. The majority in the chart also depends on 2024 and 2025 readings of 51%, a margin of one point, and on a 2026 figure that is a forecast.

The claim that agents "write more than half of new articles published online" is not what the content chart shows. Its own title says AI writes "as many" articles as humans, and the two lines end the period at roughly the same level after crossing back and forth through 2025. The classification also depends on three AI-detection tools, whose results are averaged.

The claim that agents "generate most of the code pushed to production" has no source in the essay. The only code statistic given is that "upwards of 40%" of GitHub pull requests started with an agent's draft as of September 2026. Forty percent of drafts is not most of the code, and a draft is not production code.

Smaller differences also appear. The text gives Muse 60% of agent activity; the chart says 59%, with a note that figures are rounded. The 8x HUMAN figure includes blocked requests, so it measures agent attempts to reach sites, not completed visits.

The essay's description of the market also invites scrutiny. "Most enterprises are still blocking agent traffic due to concerns about fraud and chargebacks," the authors wrote, without a figure. Measurements covered by PPC Land suggest the opposite pattern on the open web. DataDome's September 2026 test of 21,491 sites found that 65.3% let all ten of its test bots through and only 2.4% stopped all ten. Blend's audit of 200 UK and US retailers found that 29 refused its AI shopping agent, or 14.5%. HUMAN's own blocking rate for agentic traffic was close to 9% in May. Blocking may be more common at large enterprises, which Blend's tier breakdown hints at, with 16 of 80 large retailers refusing the agent; the essay does not say how it defines "most enterprises."

None of this undoes the essay's central observation, which is that agent traffic is growing quickly and concentrating among a handful of products. What the charts do not establish is that agents already dominate the web, or that most of that traffic is commercially useful today.

From clicking around to plugging in

The essay distinguishes two ways agents reach businesses. Most activity today, according to the authors, happens "via browser use - where an agent either takes over the user's browser, or spins up its own virtual machine and browser instance, and completes the task by clicking around like a human would." That is the pattern of a computer use agent, and it is why such traffic is hard for sites to tell apart from people.

The second route is newer. The authors describe the emergence of "the equivalent of an App Store for agents, but on steroids," made up of connectors, plug-ins and MCPs, through which agents reach third-party tools and sites more directly. MCP servers, which expose a service's functions to an AI model in a structured form, have spread quickly through advertising technology; Meta, for example, opened its ad system to Claude and ChatGPT through AI connectors built on an MCP server in April 2026.

To show what that looks like in practice, the authors asked ChatGPT to make a television advertisement using three plug-ins. The prompt, shown in a screen recording in the essay, read: "Can we put together a 10 second TV ad for this toy for dogs? Make it retro / 90s style," with Descript, Krea and ElevenLabs tagged. ChatGPT replied that it would make "a 10-second '90s TV spot with bright graphics, an enthusiastic announcer, and a catchy jingle, keeping the toy recognizable throughout." According to the authors, ChatGPT "coordinated across the products, and upsold us on a Descript paid plan to get the video done." The essay does not say which plan was offered or what it cost.

That upsell is the commercial point. An agent that can recommend a paid tier mid-task is, in effect, a sales channel for the software it calls. The authors expect the mechanics to get easier as platforms let users "'lend' their identity or tokens to other products," naming Sign in with ChatGPT as an example. They add that Claude "already allows enterprise admins to connect plugins for their entire organization," and that agents will eventually "select and install the tools they need without human intervention."

The essay also points to OpenAI's launch of Businesses in ChatGPT the previous day. On November 10, OpenAI opened a beta that lets companies claim a business, add product feeds and see mentions, profile views and website clicks inside a Business Hub. PPC Land had reported on November 6 that ChatGPT was building mention and click charts for verified businesses before the launch was official.

Four groups the authors expect to gain

The essay sorts potential winners of what it calls the B2A, or business-to-agent, economy into four categories.

Agents and agentic models come first. Most people, according to the authors, picture consumer agents as standalone harnesses such as Instinct, Muse and Wabi. "We believe every product will soon operate agentically, as the models behind them get more action-oriented," they wrote. They single out ChatGPT, "with 1.2 billion weekly active users," as having "a particularly intriguing opportunity here." OpenAI used the same 1.2 billion weekly figure in its Business Hub announcement.

Products newly enabled by agents form the second group. "Amazon needed the Internet to come to life, and Uber could not have existed without the mobile phone," the authors wrote, adding that agents should produce their own examples. They concede that "exactly how is hard to predict," and list three themes: self-running SMBs, marketplaces or network products that are AI-intermediated, and auto-optimizers.

Incumbents who optimize for agents are the third. Here the essay connects directly to search marketing. LLMs have already been a significant traffic source for many websites, according to the authors, and that traffic formed the basis of the generative engine optimization (GEO) industry. GEO, in their view, "largely stops at discovery. Agents will take it further, into evaluation and execution of purchase / signup choices, an even larger market."

Agent infrastructure is the fourth. Agents operating "in the real world on behalf of their humans," the essay says, will need their own tools, and the authors report "system strain at fairly low agent volume on payments, fraud and identity, and communication (iMessage, calls, email, etc.)." PPC Land's coverage has documented that strain from the site side. Arkose Labs found that 97.7% of legitimate AI agent sessions carried no verified identity, and a Cloudflare post on November 5 argued that websites bear the primary cost in nearly all of 14 problems agents create. On payments, Cloudflare and Stripe announced in April that agents could open accounts, buy domains and deploy code, subject to a default spending limit of $100 per month per provider.

Coding agents as the template

The authors point to software development as "one case study of how this might play out." Coding agents, according to the essay, "completely transformed developer behavior." As of September 2026, upwards of 40% of GitHub pull requests started with an agent's draft, against 17% six months earlier.

Like consumer agents, coding agents depend on outside services. A user asking an agent to build a website, the authors note, may need "a database, hosting, authentication, email delivery, and monitoring." Beyond the agents and harnesses themselves, which the essay lists as Cursor, Codex, Claude Code and Cognition, plus vibe coding platforms such as Replit, Lovable and Base44, the authors cite four infrastructure companies:

  • Supabase reported in June 2026 that more than 60% of new databases are launched by AI tools.
  • Neon reported in May 2025 that automated agents deploy 80% of its new databases.
  • Stripe reported in June 2026 that agents accounted for nearly 40% of documentation traffic and 70% of CLI requests for API resources.
  • Vercel reported in October 2026 that agent-triggered deployments had increased 28 times over the previous twelve months.

The Neon figure is the oldest in the essay by more than a year. The others fall within the last six months.

The authors then ask the question their thesis depends on: "Fewer than 1% of the global workforce is developers - what happens when the other 99% starts using agents?" The comparison has limits. Developer tools are built around APIs, documentation and command-line access, so an agent can call them without a visual interface. Retail checkout, travel booking and financial services are not, which is why most consumer agent traffic still arrives through a browser, by the essay's own account.

Why this matters for marketers

For advertisers and publishers, the essay's most consequential claim is about where distribution goes next. Search engines and app stores became the gatekeepers of the last two eras, and both turned that position into advertising businesses. If agents become the layer that chooses which product, plan or merchant a person ends up with, the economics of being found change again. a16z's framing treats GEO as an early, partial version of that shift, limited to discovery.

The HUMAN data cited in the essay shows how concentrated the agent layer already is. Two companies accounted for roughly nine in ten agent requests on HUMAN customer sites in October. Meta, the larger of the two by that measure, is also one of the largest advertising businesses in the world. PPC Land reported in October that, according to instruction text extracted by Profound, Muse is told to rank well-known merchants first among its top five picks. An Open Markets Institute report covered by PPC Land on November 9 warned that the same large platforms could control AI agents, though its recommendations are policy proposals, not law. Who decides what an agent picks, then: the person's instructions, the merchant's data, or the company that runs the agent?

Retailers are already setting terms. Walmart said in October it would let AI agents buy within customer-set spending limits through an invitation-only gateway, and that agents breaking its terms would be blocked. Behaviour on sites also differs from human patterns: Microsoft Clarity found checkout steps in 25% of AI agent sessions against 15% for humans, though PPC Land found the study's base was smaller than its headline suggested.

Consumer willingness remains an open question that the essay does not address. A YouGov poll for ACI Worldwide found that 60% of UK adults would stop using an AI shopping agent after one mistake, and only 19% trusted AI to make everyday purchases within their own rules.

The source of the essay is also relevant. Andreessen Horowitz is a venture capital firm that invests in AI companies, and its newsletter states that portfolio companies may be mentioned and that third-party data has not been independently verified. The essay presents a thesis about where value will accrue; it is not a neutral market measurement, and several of its charts are a16z's own constructions from other firms' published figures. The OpenRouter series it uses has appeared in the firm's newsletter before.

What comes next

According to the authors, the next pieces in the series will cover consumer agent usage, how agents choose which sites and products to direct traffic to, the infrastructure developing around agents, and the remaining hurdles for agents to complete transactions. No dates were given beyond "over the next few weeks." HUMAN has published agentic traffic reports several times this year, and its next release would show whether October's jump continued or, like May's dip, reversed for a month.

Timeline

Summary

Who: Olivia Moore and Justine Moore, partners at venture capital firm Andreessen Horowitz (a16z), using data from HUMAN Security, Shopify, Adobe and J.P. Morgan, Similarweb, OpenRouter, Graphite, NP Digital, Imperva, GitHub, Supabase, Neon, Stripe and Vercel.

What: An essay titled "Welcome to the B2A Economy" arguing that AI agents are becoming the web's primary users and that businesses which once blocked bots will have to compete for their business. It cites an eightfold rise in agent traffic between August and October 2026, with Meta's Muse at 59% and OpenAI's products at 30% of agent requests on HUMAN customer sites in October, and names four groups it expects to gain: agents, agent-native products, incumbents that optimize for agents, and agent infrastructure.

When: Published November 11, 2026, as the first in a series planned over the following weeks.

Where: On the a16z News newsletter, with data drawn mainly from US and global web measurements.

Why: The authors argue that AI has changed what people can do without yet changing how businesses reach them, and that agents choosing products and paid plans will create that new distribution channel. Several of the essay's headline claims go further than its charts support, and its blocking claim conflicts with site-level measurements PPC Land has reported, so the essay reads as an investor's thesis backed by selected third-party data.