Bots already account for nearly half of global web traffic. With the rise of large language model (LLM) agents, that number could reach 90% or more in the near future. But as the Economist puts it, the rise of the bots is risking the economic engine of the web – eye balls and clicks. Publishers are already seeing a huge drop in traffic since the introduction of ‘AI Summaries’ and now, as ChatGPT and Perplexity enable commerce and checkout straight from the chatbot, shopping is next. What does this mean for the future of the web and how we design websites and apps?
We’re talking about a very significant drop, not a blip. A study published by SEO platform Authoritas shows that when AI overviews (AIO) is present, publishers lose 47.55% of traffic. This is expected to affect 90% of queries in the future.
This isn’t the end of websites. It’s the beginning of a dual-use internet — one designed for both humans and machines. At Remagine Ventures, we believe this transformation will unlock billion-dollar opportunities for startups bold enough to build the infrastructure, tools, and interfaces of this new web.
From UI to API: Building for a Dual-Use Web
For the past three decades, websites have been designed primarily for humans — filled with buttons, forms, and friction. But AI agents don’t need drop-downs and sliders. They want direct access — via APIs and structured data.
Still, people aren’t going away. When you compare restaurants, you want to see photos, not hear a chatbot read them out. That’s why the web won’t go away — it will split. Some interfaces will be for us. Others will be for agents acting on our behalf.
And that duality is where the opportunity lies.
3 Startup Opportunities Emerging Now:
1. API Infrastructure for Agents
Even great products often expose mediocre APIs. LLMs still struggle with multi-step workflows and unclear docs.
- Opportunity: Build tools that help developers create agent-friendly APIs, with validation, multi-step orchestration, and usage simulation.
- Example: Speakeasy is helping teams build better developer portals and observability for API traffic.
2. Brand Control in Agent Interfaces
As agents summarize or act on your product, they risk flattening your brand into a sentence. Companies need new ways to show up with identity in agent interactions.
- Example: Shopify’s Sidekick is an embedded AI assistant that learns merchant context and answers questions, keeping interactions inside the Shopify universe.
3. LLM-Aware Tooling
Agents need tools that help them plan, reason, and act. Current models aren’t great at following logic across many steps or interacting with unfamiliar APIs.
- Example: Guardrails AI helps enforce structured outputs and logic across LLM workflows.
Goodbye UX, Hello AX: Designing for Memory and Intent
I was first introduced to the definition of AX, or Agentic Experience, by Greg Isenberg on X:
Traditional UX is screen-centric. you tap a button, product reacts, job done. every session starts from zero. designers pre-plan every path with hard-coded flows. users fill out forms and dropdowns because the product remembers nothing about you. success = fewer clicks and faster flows. trust = “interface looks clean so it must work.” agentic experience is relationship-centric. the agent keeps track of ongoing goals, nudges next steps, improves over time. you’re never starting over. the system plans its own path – it senses, infers, chooses actions the designer didn’t script. context is learned, not asked. preferences, patterns, even team norms are remembered. success = earned trust and compounding value. metrics shift to retention, satisfaction with decisions, how much autonomy you hand over.
The idea is simple: until now, we designed websites and apps for humans, optimising for how humans think and navigate. But now, with the rise of AI Agents, autonomously browsing and taking actions on behalf of users, we need to re-think traditional user experience (UX). Another difference is that AI agents don’t forget. They build context, remember preferences, and take initiative.
Imagine: Instead of navigating Linear to check bugs, your agent already knows you prioritise mobile bugs with estimates and surfaces them proactively.
Founders should ask: What patterns could your product remember? What tasks could it anticipate? Even smart defaults can create meaningful differentiation.
- Example: Rewind AI is building a memory layer for humans, but the same idea applies to agents — a persistent memory layer for digital behavior.
New Infrastructure Needs for the Agentic Web
The rise of AI agents is fundamentally transforming the internet, making them the new primary interface. This shift means that much of the web will integrate into an “agentic chat layer,” leading to an internet where “everything becomes headless”. The tooling and infrastructure that supported the traditional human-centric web are no longer sufficient. As AI agents evolve to handle complex tasks and act on behalf of users, a new infrastructure stack is essential.
Here’s where infrastructure builders should focus for the agentic web:
- Browser Automation 2.0 Traditional browser automation tools, designed for web scraping and testing, relied on brittle, heuristic-based scripting that struggles with the dynamic and complex tasks AI agents will pursue. Unlike humans who visually navigate, AI agents will operate using headless browsers and browser automation frameworks to interact with web pages programmatically, not with a cursor or keyboard. These next-generation systems need to navigate complex application sequences ad-hoc and understand sites based on semantic content and rendered visual interfaces, rather than just HTML code. The ability for LLMs to generate automation scripts will further increase the demand for more advanced browser and automation infrastructure.
- Machine Authentication Standards Currently, many actions an AI assistant might take on a user’s behalf—like booking a flight or changing a reservation—require the agent to log in. In the short term, agents might “spoof human activity” to achieve this, a process that is complex, requires managing sensitive credentials, and carries a high risk of account bans. For a more secure and reliable future, there will be a need for separate pathways for machine authentication and authorisation, possibly resembling consumer-directed service accounts. These secure and standardised access protocols, akin to an “OAuth for AI,” are necessary to enable agents to take actions safely and efficiently.
- Agent-First Frontends Today’s web applications are designed for visual consumption by humans, and their underlying code can be difficult for systems to interpret. To support AI agents, websites of the future may undergo redesigns that are invisible to the human eye but specifically help agents navigate. This could involve decorating HTML elements with additional comments to aid agent understanding or even creating entirely invisible “agent sections”. New standards like MCP-UI (Machine-Readable Procedural Information for User Interfaces) are emerging to allow agents to present standardized buttons and UI components, blending text with structured UI elements to give merchants more control over their brand’s appearance on third-party platforms. Utilising semantic HTML with meaningful tags like
<article>,<header>, and<section>also makes content easier for both humans and crawlers/agents to interpret. - Structured Data Layers The utility of a website to AI agents is directly proportional to how well its data is structured. Structured data, implemented using schema markup (like schema.org vocabulary), provides explicit clues about the meaning of your content, helping search engines and LLMs categorise and understand it more effectively. This enhances how content appears in search results and its likelihood of being included in knowledge graphs and answer boxes, which LLMs use to generate responses. Optimising for LLMs means focusing on semantic SEO, where content covers subjects comprehensively and uses synonyms and related phrases, rather than just keywords. This foundational emphasis on accessible, semantically rich, and authoritative content ensures that a site’s insights are readily available for AI-driven knowledge tools.
Agents and Ecommerce: The Headless Opportunity
The evolution of AI agents is rapidly making them the new interface for the internet, enabling chat or voice interfaces to scroll websites, scrape content, fetch data, call APIs, and use various tools. This transformation means a significant portion of the internet will be integrated into the “agentic chat layer,” making “everything headless”.
For commerce, this has major implications: OpenAI is working on integrating checkout directly into ChatGPT, and Perplexity already allows product purchases straight from chat. This development is both exciting and unsettling for merchants. While it offers a huge new distribution channel with massive user bases like ChatGPT’s over a billion users, merchants risk their entire brand being “compressed into the chatbot interface,” which they don’t control. This could lead to a loss of e-commerce “superpowers” like brand equity, cross-selling, up-selling, promotions, and user loyalty.
Historically, merchants have resisted dependence on single platforms (like Amazon), which is why Shopify emerged to “arm the rebels” and help them own their stores and futures. The rise of agentic checkout presents a similar dilemma, forcing merchants to decide how to stay in control.
A viable path forward is headless commerce, which separates the frontend user interface from the backend commerce logic, allowing developers to build custom UIs using APIs. This architecture involves four core API groups for the commerce experience:
- /catalog: For searching products, including images and metadata.
- /cart: To manage stateful sessions, adding or removing items, and returning to past sessions.
- /checkout: To process payment credentials and kick off the order and payment flow once the cart is filled.
- /post-order: For tracking shipping, order status, and managing disputes.
For many large merchants, most of these APIs already exist, but a key missing or insufficient component for agentic workflows is the /checkout API. Current checkout APIs often redirect to GUI-based checkout pages, breaking the unified experience of an agent (like one-click buying across merchants) and pulling users off the platform with extra clicks. To solve this, merchants need to open a checkout endpoint that can receive token-based payment credentials, enabling autonomous, one-click purchases directly within agent interfaces. This allows merchants to retain brand control and customer relationships while expanding distribution.
Additionally, new standards like MCP-UI (Machine-Readable Procedural Information for User Interfaces) are emerging to help agents present standardized buttons and UI components, allowing brands to maintain control over how their appearance is displayed on third-party platforms. This blends text with structured UI elements, giving merchants more control over their brand’s presentation in third-party contexts while leveraging the benefits of headless architecture.
The Next SEO: Generative Engine Optimization (GEO)
Search is shifting from keywords to questions. Instead of Googling “best hiking shoes,” people ask ChatGPT, Claude, or Perplexity. And these agents don’t rely on backlinks — they rely on structured, helpful, machine-readable content. I wrote about this in greater depth in my post ‘Could GEO overtake SEO as the primary driver of traffic?‘
That’s GEO: Generative Engine Optimization. Winning answers, not keyword stuffing.
- Example: Notion ranks for “best productivity tool for teams” not just because of backlinks — but because it answers the question clearly, semantically, and comprehensively.
This area is becoming a bit of a red ocean, but we’re still early in this market. For example, one of these tools was able to show a large consumer brand that the negative sentiment users were getting in LLMs about their brand wasn’t the result of bad reviews, but rather one influential post on Reddit. As models change, I suspect these companies are going to play cat and mouse with LLMs for the years to come to show their clients some results.
Closing the Loop: Your Strategic Playbook
We started with a question: What happens when most web traffic is non-human?
The answer: a web that splits in two — one for people, one for agents. The founders who understand this early will define the platforms, interfaces, and infrastructure of the next decade.
In summary, the evolution of the web into a dual-use environment necessitates innovative solutions across design philosophy, infrastructure, and business models, creating fertile ground for startups poised to shape the next era of digital interaction and commerce.
At Remagine Ventures, we’re actively investing in founders building:
- Agent-native infrastructure
- Memory-aware product experiences (AX)
- Commerce and content tools optimized for AI intermediaries
- New standards for trust, control, and machine authentication
This is the next layer of the internet. You don’t need to be OpenAI to win. You need to build the tools they and others will rely on.
I'm a former general partner at google ventures, head of Google for Entrepreneurs in Europe and founding head of Campus London, Google's first physical hub for startups.
I'm also the founder of Techbikers, a non-profit bringing together the startup ecosystem on cycling challenges in support of Room to Read. Since inception in 2012 we've built 11 schools and 50 libraries in the developing world.
- Remagine Ventures Ranked Among Israel’s Most Active VCs in 2025 - March 18, 2026
- The Global Race for AI Sovereignty: Where Does Israel Stand? - March 16, 2026
- The Web’s Next Customer Isn’t Human - March 13, 2026
