W3C and the State of the Web in the AI Era
Prepared by Philippe Le Hegaret, W3C VP Technical Strategy, finalized on 20 October 2026
About this paper
This W3C paper from the Strategy Team is intended to foster discussion in the technical community at TPAC — the most important W3C community event of the year — about how the W3C technical agenda for the web needs to evolve based on the rise of AI agents. It is part of building a consistent, coherent shared vision for evolving the web, starting with AI.
The W3C Strategy Team is looking for feedback on this analysis from the W3C Membership and the community. This report will be presented during the web and AI IG session on 29 October during TPAC 2026 and feedback is also welcome to [email protected] and the w3c/webai-roadmap repository.
Introduction
The rise of AI agents is dramatically changing the landscape of software that users — “users” in this paper means “humans” — use to interact with web content, which is having important impacts on business models, trust, and online safety.
In the two years since a 2024 Team report flagged the risks and opportunities of machine learning on the web, a major transformation has taken place: AI agents have advanced from merely reading the web to interacting with it. Every major AI platform now embeds an AI-agent driven web browser.
W3C needs to position itself as the venue where AI and web stakeholders discuss the risks and opportunities of this evolution of web user agents to develop responsible technology standards that align with W3C’s vision for the web. W3C’s core vision is to develop the web for all humanity: we have to keep humanity at the center of the web while allowing AI agents to act on their behalf.
The role of this paper is twofold: it brings together the W3C Strategy Team’s understanding of how current and emerging AI technologies intersect with the W3C's roadmap; it helps TPAC 2026 participants prepare for the many discussions on the agentic web that will take place in group meetings, a mini-Workshop, and breakouts. This paper does include some early recommendations that we expect to socialize during TPAC and seek community feedback.
AI on the web by the numbers
Caveat: the figures below come from vendors' own networks (Cloudflare, DataDome, HUMAN), each measuring a different slice of traffic, so they are not directly comparable. A limitation of these figures is that Asia-Pacific adoption is underrepresented.
The web remains the world's dominant information system and AI relies heavily on it.
In 2026, for the first time, more than half of the traffic reaching sites on Cloudflare's network was not initiated by human interactions with the website. About 2.5 billion people now use generative AI regularly (Cloudflare, July 2026) and a large portion of AI users don't check the sources, sending less traffic to the web. Daily requests from AI agents grew by more than 1,700% in the year to September 2026 (Cloudflare, Sept 2026). Ruishu Information, a Chinese security provider, reports figures drawn from monitoring and protection logs in its own protection network, covering early 2025 through Q2 2026. It says bots reached about 68% and AI agent-driven traffic rose from below 1% to 8–12%. On DataDome's network, AI accounted for 1.7% of all requests in June 2026, up 82% in a year (DataDome). AI training made up 52% of crawler requests in June 2026. Some heavily crawled categories lost up to 40% of their human traffic in under a year, while Google still sends about 88% of referral traffic.
In June 2026, Comet generated 47.6% of agentic traffic, Claude 20.8%, Atlas 16.5% and ChatGPT Agent 7.0%. Ecommerce drew 43.8% of that traffic, media 41.3% and travel 13.5%. Some 79% of agent visits landed on product and search pages, against only 2.3% on checkout and payment (HUMAN Security).
Long-held Web assumptions are changing
AI traffic and use is challenging a number of long-held assumptions for the Web in important ways.
The implicit contract of content providers with search engines — crawl my content, send me visitors — has yet no systematic equivalent for AI systems. websites want to control, and sometimes prevent, consumption for a variety of reasons, including privacy, intellectual property, and business models. Answer engines and crawlers reduce click-through and server-side rules fall short of controlling requests from AI crawlers or agents. The cost of providing free information is rising: “the toll of scraping and the behavior of some of the big technology and frontier AI companies takes a toll on our infrastructure” (Bernadette Meehan, Wikimedia Foundation’s CEO, United Nations General Assembly, September 2026). Yet, agents would benefit from clearer semantics and explicit descriptions alongside existing HTML and accessibility information: common HTML falls short for AI Agents (token bloat and noise extraction errors) and agents currently prefer text (lightly formatted with markdown) and structured data (json, yaml, csv).
In parallel, generated text, images, audio and video are cheap and appear credible to a large portion of users. AI systems now need to ensure that their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (as required for instance by the article 50 of the EU Artificial Act as of August 2026). However, there is no way to guarantee that content was not machine-generated, which raises the stakes for provenance, identity and credentials.
Since the birth of the web, users have primarily used web browsers to interact with it. A year ago the common bet was that agents would bypass the browser for APIs and MCP servers. It is however unlikely that MCP will ever provide endpoints at the same scale as websites. Service providers keep relying on the websites because the web browser already solves sandboxing, authentication, discovery, trust (HTTPS and origins), identity segmentation, and, most importantly, a place for the human to step in. All major AI platforms provide a web browser or a way to interact with one.
We’re moving into a model where the web browser and the AI agents, both acting as web user agents, need to cohabit and interact with each other, while properly handling user consent/delegation, privacy, and liability.
Risks to users and the ecosystem
- Security. agents having access to private data, interpreting untrusted content offer an exfiltration path similar to cross-site scripting. The important difference is that an attacker no longer necessarily needs to inject executable JavaScript. Malicious instructions can be carried by ordinary text in a web page, document, or tool response. The mechanism is different, and success depends on the agent's authority and how it handles untrusted content. Prompt injection and misaligned agents undermine the same-origin protections the web relies on.
- Privacy and consent. Data handled by many agents crosses contexts and is used to train models. Users may not meaningfully consent to what an agent does, including oversharing information and confirmation dialogs risk prompt fatigue. Speaker identification and inference from audiovisual metadata amplify tracking. Media recordings of users can be reused by agents to create deepfakes. Without modelling the information flows, it will be challenging to enforce separation between the user, agent memory, websites and sessions.
- Accessibility. Optimizing ARIA for agents may break the human experience. Generated alt text may be inaccurate and should not entirely replace authored descriptions or author intent. Misrepresentation of user needs due to bias in training data. AI may amplify the existing accessibility debt in existing large codebases.
- Trust and provenance. Credible synthetic content and impersonation raise the cost of trust online. Disclosure, provenance evidence and usage preferences are different things.
- Centralization and lock-in. Agents locked to or favoring partner-approved sites, dependency on specific models, and vendor MCP tool registries all threaten one interoperable web.
- Publisher sustainability. Less human traffic means less incentive to publish openly, with a disproportionate effect on less-resourced creators.
- Equity and bias. Models underperform for minority languages, scripts and cultures; averages can hide failures. Bias falls hardest on users already poorly represented in training data.
- Sustainability. AI's footprint covers electricity, water, minerals and hardware across the full lifecycle; W3C still lacks a framework to evaluate the environmental impact of its standards.
Opportunities
AI can make the web more capable, more accessible and more personal if it is built on the web's existing guarantees.
- Agents that use the website a person already uses, with access to their active sessions, can guide the users and streamline their online interactions.
- On-device inference keeps personal data on the device.
- User-controlled data via personal storage is possible with delegated, revocable authorization.
- Accessibility gains: computer vision/image recognition in AI models are useful. Generated audio descriptions, generated sign-language, use poorly coded content, improve poorly coded content. Other useful aspects include real-time captioning with richer metadata (speaker, flash warnings), translation, read-aloud, adaptive layouts, access content from inaccessible documents instead of using Optical Character Recognition (OCR) to extract the text content. Summarizing complex documents can be helpful to neurodiverse people.
- Structured data: Knowledge graphs bring advantages to Retrieval-Augmented Generation (RAG) applications (contextual understanding, knowledge Integration, explainability and transparency). JSON-LD, RDF, DCAT and ODRL become more valuable as inputs and guardrails for agents.
- Trust infrastructure: Verifiable Credentials and DIDs can identify agents and express whom they act for.
- Better standards work and adoption: AI-assisted threat modeling, horizontal review, code generation and testing can be beneficial under a clear W3C AI contribution policy.
The value of the web in the AI era
This is a proposal for a declaration of the value of the web in the AI area:
- The web is for all people — including those who choose to use agents. Agents serve people.
- An agent acting for a person is a user agent. It owes that person loyalty, transparency and safety, as a browser does.
- People stay in control. Consequential actions need real human consent; people can see what was done, undo where possible, and choose which agent acts.
- One interoperable, world-wide web. No agent lock-in nor AI-only side-channels that replace the human web.
- The web's security model is not optional, but agents are a new composant of the model. The interaction varies for browser-integrated agents, extensions and external agents controlling the browser. Nevertheless, origins, sandboxing and credential isolation apply to agents as they apply to scripts.
- Creators have a voice in how their work is used. Open publishing needs a fair, machine-readable bargain model for training and reuse.
- Accessibility for people comes first. Semantics, alt text and author intent are not to be repurposed for machines to the detriment of humans.
- Trust must be verifiable. Provenance and credentials let people know who and what they are dealing with.
- The web must be sustainable. AI features added to the web should be necessary, efficient and transparent about their resource costs across their whole lifecycle.
This declaration would extend the Ethical web Principles and the Ethical Principles for Web Machine Learning rather than replace them.
We believe these values have played a critical role in making the Web a flourishing ecosystem in which users, content and service providers and technology providers have been able to build constructive relationships and want to advocate their adoption in the agentic web context.
What the W3C community is telling us
TPAC has plenty of agentic-web proposals, while nearly half of W3C groups have not answered the AI roadmap survey.
TPAC 2026 breakouts: The agentic web is the dominant thread, centered on the interface model (HTML-as-is vs new markup vs WebMCP), consent for consequential actions, agent alignment and neutrality, and agent identity for commerce. Agent identity is attempting to converge with existing identity work.
Roadmap survey of 48 groups: Half of the groups responded. Most report an unchanged mission with rising relevance. The clearest priority shift is LWS raising the need for authorization and delegation. Timed Text and MEIG created the AIRFOT task force for AI-sourced captioning requirements. ARIA participants cautioned against bending human-centred semantics to suit agents. Sustainable web flagged lifecycle impacts without reaching consensus against AI.
In addition, feedback on the Browser Testing and Tools Working Group charter requested to expand the scope of WebDriver BiDi to cover agentic use cases.
Current work and gaps
W3C already has relevant work in most areas. W3C should continue to leverage AI on the web (platform capabilities, data, agents, identity) and shape its impact (ethics, security, privacy, accessibility, sustainability).
| Theme | Existing or proposed work | Main gap |
|---|---|---|
| Agents on the web | Proposed Agentic Web WG (WebMCP, agent identity, permissions, transparency), webSkills, Autonomous Agents on the web CG, web Content Browser for AI Agents CG | Declarative vs imperative model; human handoff; tasks across navigation; |
| AI in the platform | WebML WG (WebNN), built-in AI APIs, WebGPU, WebAssembly, Web Speech, Generative UI | Testing non-deterministic behavior; cross-engine support; user control of where models run |
| Identity and trust | DID WG, VC WG, Federated Identity WG, DID Methods WG, LWS WG, AI Agent Protocol CG, Agent Conformance and Benchmarking CG, Agent Identity Registry Protocol CG, Agent Trust Protocol (ATP) CG | Agent identity profiles; delegated authority; user intent and consent or approval; |
| Content, provenance and rights | Publishing, Timed Text/MEIG (AIRFOT), AI Content Disclosure CG, AI Computed Provenance CG, TDMRep, proposed Digital Policy WG (ODRL), Agent-to-Web Framework (A2WF) CG | Fragment-level metadata; separating disclosure, provenance and training preferences; commerce for content |
| Data and semantics | JSON-LD, RDF & SPARQL, Dataset Exchange, LWS, AI KR (Artificial Intelligence Knowledge Representation) CG | Best-practice guidance; authorization delegation |
| Payments | Web Payments WG, Web Payment Security IG | Concrete agentic-payment work items; split with EMVCo, FIDO, IETF |
| Horizontal review | TAG, Privacy, Security IG, APA/ARIA/AG, i18n, Sustainable web | AI-specific review guidance; environmental evaluation framework |
Policy engagement
Policymakers are already regulating AI content, training data and agent liability; W3C should offer them standards that can carry those rules, without taking positions on the regulation itself.
- Where standards meet regulation. AI content labelling (Content Disclosure CG), training-use reservations (TDMRep, robots.txt evolution at IETF, ODRL policies), agent identity and accountability, and environmental reporting are natural hooks for regulation.
- Our role. Explain the web's architecture and principles, point to the standards that can implement policy goals, and flag where proposed rules would fragment the web or weaken user protections.
- Build on W3C members strength. Build positions through a Member-facing forum (TPAC breakout #43 is the starting point) and coordinate with IETF, ISO/IEC, C2PA, EMVCo and FIDO where work overlaps.
- Consistency. One set of messages derived from the declaration above, reused across regions, so W3C speaks with one voice.
Strategy Team recommendations and next steps
- Refine the declaration of the web's value in the AI era, in coordination with the Web and AI IG and the TAG
- Charter the Agentic Web WG with explicit scope for security and privacy, agent identity and delegation, and the WebMCP-HTML-ARIA relationship (TPAC #22).
- Protect accessibility semantics: ARIA/APA guidance that instructions targeted at agents must not degrade the accessibility tree, and guidance against prompt injection via ARIA.
- Advance a content-use bargain: Understand how TDMRep, ODRL (Digital Policy WG) and IETF work help in balancing the different interests.
- Clarify provenance: separate disclosure, provenance evidence and training preferences, at resource and fragment level.
- Close survey gaps: follow up with the groups, prioritizing horizontal-review groups.
- Refine threats: guidance on testing non-deterministic behavior and on user control over where models run.
Proposed next steps for W3C groups
The Strategy Team will submit these proposals to each group for discussion and next steps.
Groups with substantive input: turn direction into work items
| Group | Proposed next step | Coordinate with |
|---|---|---|
| Verifiable Credentials WG/CCG | Publish an agent use-case note: identity, delegation, responsibility; assess if the data model suffices | Web Payments |
| Web Payments WG | Name specific agentic-payment work items from the September e-commerce workshop | Payment Security IG, EMVCo, FIDO, KYAPay at IETF |
| Web Payment Security IG | Develop payment requirements for AI agents on the web | Web Payments WG |
| Publishing Maintenance WG | Capture AI disclosure, training-permission and alt-text needs as EPUB use cases | AI Content Disclosure CG, Digital Policy WG |
| RDF & SPARQL WG | Publish a short note on knowledge graphs as agent inputs | JSON-LD, Dataset Exchange |
| Sustainable Web IG | Draft guidance on AI usage and lifecycle impact | TAG |
| ARIA WG | Post a group summary; draft guidance against agent-oriented misuse of ARIA and prompts in aria-label | APA, AG, Agentic Web WG |
| Privacy WG | Agent consent, cross-context data, model-run location ( TPAC #72 ) | TAG, Agentic Web WG |
| Federated Identity WG, Web Authentication WG | Agent identity and human approval for consequential actions | VC, DID, LWS |
| Web Applications WG, Web Editing WG | Agent interaction with forms, editing and built-in AI APIs | Agentic Web WG |
| Media WG, Audio WG, WebRTC WG, Web Transport WG | speech and real-time agent sessions | MEIG |
Agent requirements, paired with user requirements
Groups whose technologies are already leveraged by agents should document what agents need from them, the way APA's Research Questions Task Force documents accessibility user requirements (XAUR, RAUR, MAUR). Each document would pair two lists: what an agent needs to act reliably, and what the person needs to stay safe and in control when an agent acts for them. Where they conflict, the person's needs win; the ARIA discussion makes this point.
| Group | Agent requirements to document | Paired user requirements |
|---|---|---|
| Proposed Agentic Web WG | discovery, tool description, state across navigation, handoff points | Consent for consequential actions, visibility of what the agent did, choice of agent |
| APA WG (RQTF) | How agents and assistive technology share the accessibility tree without degrading it | Screen-reader and cognitive-accessibility needs in agent-mediated tasks |
| Web Applications WG, Web Applications Security WG, Web Editing WG | Stable form semantics, field purpose, validation and submission state | Review before submit; no silent autofill of sensitive data |
| Web Payments WG | Payment intent, agent mandate and limits, merchant-side signals | Spending caps, revocation, clear receipts and dispute paths |
| Web Authentication WG, Federated Identity WG | How an agent presents delegated authority without holding credentials | Proof that a human approved; per-agent, revocable permissions |
| Media WG, Timed Text WG | Segment identification, rights and attribution for agents that cite or play media (Agent Ready Video CG) | Captions and metadata remain intact and accurate |
| Publishing Maintenance WG | Reuse and disclosure metadata | Readers can see what was AI-generated |
The TAG could also add agent considerations to their self-review questionnaires, next to security and privacy.
Security Interest Group: proposed actions
The Security IG should move agent-related security reasoning upstream, so groups arrive at horizontal review with their threat analysis already done.
- Threat models for the web and Agents, organizing them as a family: Threat Model for the web as a foundation, an Agentic Web model for agent interactions, WebMCP as a specific analysis, and a related AI in the Browser model for developer-facing AI capabilities. First, we would need to reconcile architectures, shared threats and trust boundaries. For the threat model for agents, consider prompt injection, the private-data / untrusted-content / exfiltration trifecta, confused deputies, spoofed tools (no "padlock" for tools), agents approving their own actions, etc. A web developer accessing browser-provided AI APIs (such as Prompt API) is different from an AI agent navigating on behalf of a user. The next step is to clarify how the approaches differ and where they share threats. This includes finishing the "AI in the Browser" model, covering built-in AI APIs, prompt observability traces (TPAC #63) and where models run (TPAC #72).
- Security requirements for human approval, with the Agentic Web Working Group: what makes a confirmation trustworthy and resistant to prompt fatigue.
- Agent identity threats, building on the decentralized-identities model, with DID, VC and FedCM groups and the KYAPay work at IETF.
- Facilitated threat-modeling sessions before horizontal review for the highest-risk work: WebMCP, built-in AI APIs, agentic payments.
- AI-assisted security review: evaluate portable AI skills for spec review (TPAC #33) with human sign-off, under the W3C AI contribution policy.
- Engage webAppSec on how sandboxing/isolation intersects with agents when they use web technologies.
Privacy Working Group: proposed actions
The Privacy WG should test whether W3C's privacy principles still hold when the user agent is an AI agent, and give groups concrete guidance before agent features ship.
- Review the Privacy Principles for agentic user agents. Identify gaps when an agent, not the person, reads pages, fills forms and moves data between sites, including duties of the agent to its user.
- Define meaningful consent for delegated actions, with the Agentic Web WG and Security IG: what a person must understand and approve, how long consent lasts, and how to avoid prompt fatigue.
- Guidance on cross-context data flows. Agents carry data between sites, sessions and memory; address secondary use, data ingested into models, and an "incognito" mode with no agent memory.
- Guidance on user control over where models run (TPAC #72): browser-provided and cloud inference that user agents can enforce.
- Consider fingerprinting from model-backed APIs and agents: built-in AI APIs, assistive-technology adaptations (TPAC #50), and agent-specific behavior that can identify users.
- Privacy of AI observability: what prompt and tool-call traces (TPAC #63) may record and who can see them.
- Inference from content and metadata: with MEIG and Timed Text, address AI inferring location, people and speakers from media, and the privacy risk of speaker identification.
- Anti-abuse without tracking: review anonymous-credential approaches such as PACT (TPAC #29) as an alternative to CAPTCHAs and invasive bot detection.
- AI-assisted privacy review: pilot AI tools to scale horizontal privacy review, with human sign-off and transparency about the tools used.
Other actions
- TAG and Web and AI IG: produce a declaration for W3C about the value of the Web in the AI era
- Proposed Agentic Web WG: host the WebMCP-HTML-ARIA discussion; own human-handoff and consent patterns.
- Proposed Digital Policy WG: provide the ODRL basis for training and reuse preferences used by Publishing and others.
NOTE: this report was developed with AI assistance, its human editor takes full responsibility for its content.