The next contract: structuring the page for an unknown reader

Every argument in this module has rested on one clause of the platform contract: a document is written once, and its rendering is delegated to a user agent the author never meets. Unit 1 named that clause as the origin of the web's resilience. Unit 2 named its first non-visual reader, the assistive technology that consumes the accessibility tree. The clause is now doing new work: the software that fetches a page is no longer only a browser waiting for a person, but an agent pursuing a goal, a generative system assembling an answer, a crawler deciding what a page is worth. The contract is being renegotiated from rendering for an unknown device to structuring for an unknown reader. The strongest counter to this unit's thesis is that nothing in the architecture has changed. ActivityPub (W3C, 2018) was a Recommendation five years before the agent benchmarks, and Grigalis (2013) was extracting structured records from template pages a decade before them.

The design response is continuous with the nine units before it: declared structure is what assistive technology already needs. What is new is that this reader, unlike a browser, competes for the attention the page was written to earn. This one asks who the reader has become.

The machine reader reconstructs structure whether or not it is declared

Machines have read pages written for people for decades. Grigalis (2013) set out an unsupervised, domain-independent method for extracting structured records from template-generated pages. The method clusters visually similar elements by formatting and structure, and it derives reusable wrappers from their tag paths. A page that declares nothing has therefore not prevented inference. It has made inference brittle. That non-visual reading was already the condition of the reader Unit 2 described.

Work the inference on one page. A product listing writes each item as a div, with the price in a span and the title in a link. The markup says nothing about which text is the name and which is the price. An extractor clusters the elements that repeat with the same tag path and formatting, and it infers the field boundaries from the repetition. The reading is possible, and it is brittle. A redesign that renames the class, or a price sometimes struck through, breaks the inferred wrapper.

Agents are now asked to use the ordinary web, and they are not good readers yet

An autonomous web agent receives a goal in language and must complete it on a real site by interpreting the page and acting on it. Three benchmarks define the evidence.

• WebArena (Zhou et al., 2024). A reproducible environment of self-hosted websites from four domains, including e-commerce, social and content management, with long-horizon tasks scored on functional correctness. The best GPT-4-based agent completed 14.41 per cent of tasks, against 78.24 per cent for people.
• Mind2Web (Deng et al., 2023). Over 2,000 open-ended tasks from 137 real websites across 31 domains, built to end the dependence on simulated sites. Filtering raw HTML with a small model improves an agent's effectiveness and efficiency.
• WebVoyager (He et al., 2024). An agent built from large multimodal models that completes instructions end to end on live sites, evaluated on tasks from 15 popular websites. It reports 59.1 per cent task success, with a protocol that agreed with human judgement 85.3 per cent of the time.

The distance between 14.41 per cent and 78.24 per cent is the unit's first honest measurement of the unknown reader, and WebVoyager's 59.1 per cent is less a contradiction than a change of target. A 2026 study joins the agent question to the accessibility tree. Gubbi Mohanbabu and colleagues (2026) gave a computer-use agent 60 tasks: it succeeded on 78.3 per cent by default, on 41.7 per cent restricted to the keyboard, the condition a screen reader user works under, and on 28.3 per cent behind a magnified viewport.

Structure can be declared, and the hypermedia tradition is the template

If the response is to declare structure, the web has a tradition for it. Lanthaler (2013) argues that hypermedia-driven APIs can be built from Linked Data. A short list of design guidelines and a small vocabulary, on that argument, exposes the interface to a client with no prior knowledge of it.

The claim is not that every site should publish linked data. It is that a representation can carry its own controls, so a machine that has never seen the service can discover what it may do. ActivityPub (W3C, 2018) is the normative case, a Recommendation of January 2018 that defines a decentralised social protocol on the ActivityStreams 2.0 format, with a client-to-server API and a federated server-to-server API.

Work the contract on one document. An ActivityPub client that has never met a server fetches an actor's address. The representation it receives names that actor's inbox and outbox, and the client posts to the inbox it was told about rather than one it was configured with, because the controls travelled inside the representation (W3C, 2018).

The unknown reader is also a gatekeeper

None of this would be a design problem if the machine reader's interests matched the person's. The recent evidence, all of it preprint, says they do not. Khosravi and Yoganarasimhan (2026) compare external-search referrals to English Wikipedia articles with the same articles in German and French, across the staggered rollout of Google's AI Overviews. They estimate that default availability reduced English search traffic by 5.45 per cent and 4.82 per cent. Lin (2026) uses daily robots.txt snapshots across 492 and 825 news websites, and finds no detectable average ranking effect from opting out of AI training, but a 4.3 per cent average worsening from opting out of AI search.

Both are preprints, so both are claims under review. Together they show the new reader is not a neutral visitor: it reads at scale, and it changes who reaches the page afterwards. Structuring for it is distributive, as Unit 8 described for performance, because benefit and cost fall on different parties.

The machine reader is already inside the tools that write the page

Generative support now sits inside the authoring tools themselves. Wadinambiarachchi and colleagues (2024) tested its effect in a between-participants experiment with 60 participants. AI image support produced higher fixation on an initial example, and users generated fewer ideas, with less variety and lower originality than the baseline.

Kim and Ji (2025) varied response disclosure and prompt initiative in a two-by-two within-subject experiment with 36 participants. Progressive disclosure and high prompt initiative increased perceived knowledge change, engagement and germane cognitive load.

Song and colleagues (2025) show where the argument meets accessibility: existing alt text tools give one language, written for one reading level, or a literal machine translation that misses cultural nuance, and their tool generates descriptions with culture-specific, region-appropriate conventions.

Contested question: is the new reader a rewriting of the contract, or a claim on it?

One position holds that the contract is being rewritten: the page becomes a representation for machines first, and the response is declared structure, self-describing hypermedia (Lanthaler, 2013) and decentralised federation (W3C, 2018). The other holds that nothing in the architecture changed, because the standards always delegated interpretation to software the author does not control. On that reading the agent wave tests whether sites ever honoured the existing contract.

The decentralisation strand tests a contract change against a slogan. Zhang and colleagues (2025) assembled over 3.1 million posts from 13,000 users across DeSo and Mastodon, with 11 million DeSo on-chain transactions. They found its social graph decentralised while its transaction graph remains centralised. What would settle the dispute is a matched comparison of agent task success between sites that declare structure and visually equivalent sites that declare nothing.

A second dispute is distributive rather than architectural. The shared-good position rests on convergence: declared structure is what assistive technology already needs, and Song and colleagues (2025) treat machine generation and human accessibility as the same task. The traded position rests on the economics and design studies. Khosravi and Yoganarasimhan (2026) estimate reductions of 5.45 per cent and 4.82 per cent in search referrals. Lin (2026) estimates a 4.3 per cent ranking penalty for withholding content from AI search. Wadinambiarachchi and colleagues (2024) found fewer and less original ideas once a generative tool entered ideation. What would settle it is a within-site comparison of human task success, agent task success and referral traffic before and after a change that declares structure. No cited study does that, so the two positions share the first move and differ on how much a machine may take.

While the question is open, a designer should declare the structure the page half-carries and decide what a machine may take. Mark up the content with the elements and properties that state what it is, so assistive technology and an agent read the same structure. Leave the interactive path usable without a machine reader, including on a slow or metered connection, and say in the interface where an automated answer has been used.

Where the argument is heading

Three directions are already research programmes. Measurement is consolidating on reproducible environments and on live-site benchmarks, the approaches Zhou and colleagues (2024) and He and colleagues (2024) represent. The representation is being pushed back towards self-description, the position Lanthaler (2013) argues for hypermedia-driven APIs. The economics is being formalised, with Lin's (2026) separation of training use from search use suggesting governance will split the same way.

The sources this unit rests on

• Zhou, S., et al. (2024). WebArena: A Realistic Web Environment for Building Autonomous Agents. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2307.13854 A reproducible environment of self-hosted websites across four domains, with long-horizon tasks scored on functional correctness, in which the best GPT-4-based agent succeeded on 14.41 per cent of tasks against 78.24 per cent for people.
• Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H. and Su, Y. (2023). Mind2Web: Towards a Generalist Agent for the Web. Advances in Neural Information Processing Systems 36. doi:10.52202/075280-1220 Over 2,000 open-ended tasks from 137 real websites across 31 domains, and the finding that filtering raw HTML with a small model improves an agent's effectiveness. It establishes that real pages, not simulated ones, are the target.
• He, H., et al. (2024). WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. doi:10.18653/v1/2024.acl-long.371 An agent built from large multimodal models that completes instructions end to end on live sites, evaluated on tasks from 15 popular websites, reporting 59.1 per cent task success with a model-based protocol that agreed with human judgement 85.3 per cent of the time.
• Gubbi Mohanbabu, A., Natalie, R., Kim, B., Guo, A. and Pavel, A. (2026). A11y-CUA Dataset: Characterizing the Accessibility Gap in Computer Use Agents. CHI '26. doi:10.1145/3772318.3791896 Sixty everyday tasks went to a computer-use agent: it succeeded on 78.3 per cent by default, 41.7 per cent keyboard-only and 28.3 per cent magnified, the conditions assistive-technology users work under.
• Grigalis, T. (2013). Towards web-scale structured web data extraction. Proceedings of the sixth ACM international conference on Web search and data mining. doi:10.1145/2433396.2433491 Unsupervised, domain-independent extraction of structured records from template-generated pages, clustering visually similar elements and deriving reusable wrappers from their tag paths. It shows that presentation is read as structure when structure is not declared.
• Lanthaler, M. (2013). Leveraging Linked Data to Build Hypermedia-Driven Web APIs. In REST: Advanced Research Topics and Practical Applications, 107-123. doi:10.1007/978-1-4614-9299-3_7 The chapter argues that Linked Data plus design guidelines and a small vocabulary produces loosely coupled APIs usable by a completely generic API browser, demonstrated with a prototype. The clearest statement that the contract can be extended to machines by structure.
• W3C (2018). ActivityPub. W3C Recommendation, 23 January 2018. https://www.w3.org/TR/activitypub/ The Recommendation defines a decentralised social networking protocol on the ActivityStreams 2.0 format, with a client-to-server API for content and a federated server-to-server API for delivery. It is normative, and it is the standard against which contract claims can be tested.
• Zhang, T., Mahanti, A. and Naha, R. (2025). Web3 vs Fediverse: A comparative analysis of DeSo and Mastodon as decentralised social media ecosystems. Online Social Networks and Media. doi:10.1016/j.osnem.2025.100337 Over 3.1 million posts from 13,000 users across DeSo and Mastodon, with 11 million DeSo on-chain transactions, showing DeSo's social graph decentralised while its transaction graph remains centralised.
• Khosravi, M. and Yoganarasimhan, H. (2026). Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia. SSRN working paper 6164926. doi:10.2139/ssrn.6164926 Preprint. A difference-in-differences estimate using AI Overviews' staggered rollout and Wikipedia's multilingual structure, in which default availability reduced English search traffic by 5.45 per cent and 4.82 per cent against German and French comparisons.
• Lin, Y. (2026). Opting Out of Generative AI? Training Data, Search Visibility, and the Traffic Rankings of News Publishers. SSRN working paper 7518878. doi:10.2139/ssrn.7518878 Preprint. Daily robots.txt snapshots across 492 and 825 news websites show no detectable average ranking effect from opting out of AI training but a 4.3 per cent average worsening from opting out of AI search.
• Wadinambiarachchi, S., Kelly, R. M., Pareek, S., Zhou, Q. and Velloso, E. (2024). The Effects of Generative AI on Design Fixation and Divergent Thinking. Proceedings of the CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3642919 A between-participants experiment with 60 participants found that AI image support during ideation produced higher fixation and fewer, less varied and less original ideas than the baseline.
• Kim, N. and Ji, Y. G. (2025). Exploratory Search with Generative AI: An Empirical Study on the Impact of Interaction Design Strategies on Information Exploration and Cognitive Load. SSRN working paper 5418661. doi:10.2139/ssrn.5418661 Preprint. A two-by-two within-subject experiment with 36 participants found that progressive disclosure and high prompt initiative increased perceived knowledge change, engagement and germane cognitive load and produced more exploratory behaviour.
• Song, H., Choi, J., Shin, M., Suh, B. and Jung, H. (2025). AltCAT: An Alt Text Authoring Tool with Automatic Generation and Culturally-Aware Translation. Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility. doi:10.1145/3663547.3759732 Noting that existing alt text tools give one language or literal machine translation that misses cultural nuance, the authors build a tool that generates descriptions and translates them with culture-specific knowledge and region-appropriate conventions.