{
  "schemaVersion": "1.0",
  "kind": "nilu.article",
  "id": "web-agent-ready-overview",
  "slug": "web-agent-ready-discover-cite-execute",
  "locale": "en",
  "sourceLocale": "pt-br",
  "localizationStatus": "translated",
  "title": "Web agent-ready — what changes when part of the audience isn't human",
  "summary": "Three layers — discover, cite, execute — organize what changes when bots and agents read your site, with clear priorities and limits.",
  "author": "NiLu",
  "publishedAt": "2026-08-07T00:00:00.000Z",
  "updatedAt": "2026-08-09T00:00:00.000Z",
  "canonicalUrl": "https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute/",
  "representations": {
    "html": "https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute/",
    "markdown": "https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute.md",
    "json": "https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute.json"
  },
  "bodyMarkdown": "# Web agent-ready — what changes when part of the audience isn't human\n\nThree layers — discover, cite, execute — organize what changes when bots and agents read your site, with clear priorities and limits.\n\nAuthor: NiLu\nLocale: en\nPublished: 2026-08-07\nUpdated: 2026-08-09\nCanonical: https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute/\nMarkdown: https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute.md\nJSON: https://blog.nilu.pro/en/blog/web-agent-ready-discover-cite-execute.json\n\n## In brief\n\n- Part of the traffic that reaches a website has always been automated; what changed is that a fraction of it now also generates AI responses — and, still experimentally, acts on a user's behalf.\n- Three distinct jobs — discover, cite, and execute — organize priority better than treating everything as \"SEO with more schema.\"\n- For Google, optimizing for AI features in Search is still SEO: no special markup and no llms.txt requirement.\n- Agent execution (WebMCP) is experimental and should not drive budget decisions today.\n\n## Contents\n\n- [Part of the audience isn't human](#audiencia-nao-humana)\n- [Three layers: discover, cite, execute](#tres-camadas)\n- [Why it's not just \"SEO with more schema\"](#por-que-nao-e-so-seo-com-schema)\n- [What to prioritize first](#o-que-priorizar-primeiro)\n- [Where the hype lies](#onde-fica-o-hype)\n- [Uncertainties](#incertezas)\n\n<a id=\"audiencia-nao-humana\"></a>\n\n## Part of the audience isn't human\n\nNot every visitor to a website is a person reading in a browser. Part of the traffic has always come from crawlers and automated tools; what changed is that a fraction of it is now also used to compose responses in AI-powered search tools and, still experimentally, to act directly on a user's behalf. Chrome for Developers describes this shift as two stages of the \"agentic web\": agents that search (searching, where SEO fundamentals still apply) and agents that use the site directly (using, where the accessibility tree, layout stability, and, at the limit, WebMCP come into play). [[1]](#ref-1) The W3C Technical Architecture Group discussed, at a March 2026 meeting, how to distinguish humans, bots, and agents acting on behalf of users within the Web's architecture — an ongoing debate, with no closed conclusion published in the post. [[2]](#ref-2)\n\nOne number tends to circulate to illustrate this shift: Imperva's 2025 Bad Bot Report, based on the vendor's own network in 2024, recorded automated traffic (51%) surpassing human traffic (49%), with 37% classified as malicious bots and 14% as \"good\" bots. This is telemetry from a security vendor about its own customer base — not a census of the web — and \"automated\" is not synonymous with \"AI agent.\" Use this number to justify crawl policy and bot monitoring, not to budget for \"agent presence.\" [[3]](#ref-3)\n\n<a id=\"tres-camadas\"></a>\n\n## Three layers: discover, cite, execute\n\nIt's useful to separate three jobs that tend to get bundled together under \"SEO for AI\":\n\n- **Discover** — being crawled, indexed, and eligible to appear in search features, including AI-generated ones.\n- **Cite** — being used as evidence, with a link and attribution, inside a generative response. This is not the same as holding a stable ranking position.\n- **Execute** — an agent calling actions directly on the site (filling out a form, buying, scheduling). Today this is experimental.\n\nFor the Discover layer, Google Search Central's guide to generative AI features states that these features are built on the same core Search quality and ranking systems, and that there is no special AI markup nor an `llms.txt` file requirement — optimizing for them is still optimizing for Search. The \"AI features and your website\" documentation reinforces that there is no technical requirement beyond being indexed and eligible for a snippet. [[4]](#ref-4) [[5]](#ref-5)\n\nFor the Cite layer, generative engines retrieve content and synthesize responses with attribution — OpenAI describes ChatGPT search this way, with a sidebar of sources. [[6]](#ref-6) There is no public \"scoreboard\" today for citation position equivalent to search ranking.\n\nThe Execute layer depends on WebMCP, a proposal still in origin trial in Chrome (from version 149) and documented by the W3C Web Machine Learning Community Group as a Draft Community Group Report — explicitly, it is not a W3C standard nor on track to become one. It has relevant limitations: it requires an open browser context, does not work headless, and capability discovery requires visiting the site directly. [[7]](#ref-7) [[8]](#ref-8)\n\n<a id=\"por-que-nao-e-so-seo-com-schema\"></a>\n\n## Why it's not just \"SEO with more schema\"\n\nThe same Google guide lists, among the misconceptions to avoid, betting on `llms.txt`, special AI markup, splitting content into \"chunks\" for AI, or rewriting text only for agents. Structured data remains useful when it matches the visible content — not as a citation shortcut. [[4]](#ref-4) The \"Succeeding in AI search\" article, from the Search Central Blog, extends the same people-first advice and the same longstanding technical requirements (crawling, indexing, snippet eligibility) to AI-generated responses. [[9]](#ref-9)\n\nThis doesn't invalidate separating discover, cite, and execute — it just avoids turning the map into proven science. Chrome's \"searching vs. using\" language and the TAG's \"humans, bots, agents\" language support the vocabulary, but it is implementer and architecture language, not proof that this model is the best way to describe 2026. We treat the three layers as a working hypothesis tested against the available evidence — not as a closed framework. [[1]](#ref-1) [[2]](#ref-2)\n\n<a id=\"o-que-priorizar-primeiro\"></a>\n\n## What to prioritize first\n\n- **SEO foundation** — crawling, indexing, snippet eligibility, and people-first content. This is the documented baseline for appearing in any AI-powered search feature. [[4]](#ref-4) [[5]](#ref-5)\n- **Clean representations** — semantic HTML and main content clearly distinguishable from menus and footers; negotiating Markdown (`Accept: text/markdown`) is a complementary capability when the stack supports it, not a Search requirement.\n- **Crawl policy** — separate training bots from search/citation bots. OpenAI distinguishes GPTBot (training), OAI-SearchBot (search indexing), and ChatGPT-User (real-time user-initiated search); Anthropic documents how to block its crawler. [[10]](#ref-10) [[11]](#ref-11) Edge vendors report that AI crawlers already account for a relevant share of the bot traffic observed on their own networks — Fastly cited nearly 80% of the AI bot traffic identified in its analysis; use this kind of figure to size bandwidth and origin costs, not as a global web statistic. [[12]](#ref-12)\n- **Real metrics** — Search Console (including impression reports for gen-AI features, where they exist), sessions, and conversion — not `llms.txt` files published \"because everyone has one.\" An Ahrefs study of about 137 thousand sites found that 97% of published `llms.txt` files received no requests at all in the analyzed month. [[13]](#ref-13)\n- **Execution (WebMCP)** — watch, without committing budget; it is still an origin trial and a community draft.\n\n<a id=\"onde-fica-o-hype\"></a>\n\n## Where the hype lies\n\nThe paper that coined the term GEO (Generative Engine Optimization) reported a relative visibility gain \"of up to 40%\" — but in the authors' own experimental setting, with a specific pipeline (Google's top-5 reprocessed by a GPT-3.5-style model), not in a public ranking product. This is method evidence, not a prediction of business return. [[14]](#ref-14)\n\nEven when a piece of content is cited, that does not equal a visit. The Pew Research Center found, in a correlational analysis, an organic-link click-through rate of about 8% when an AI summary appeared in Google results, versus about 15% when it did not; only around 1% of users clicked a link inside the summary itself. [[15]](#ref-15) Digiday reporting, citing Digital Content Next data, reported a median of approximately -10% year-over-year in referral traffic from Google among publishers (with wide variation between cases) associated with AI Overviews; another report, citing Similarweb, put zero-click news searches rising from about 56% to 69%, with traffic coming from AI platforms growing without offsetting that loss. These are trade-press figures about specific cohorts — useful as a signal of economic pressure, not as universal causal proof. [[16]](#ref-16) [[17]](#ref-17)\n\nThe quality of the citation itself is also fragile: a 2023 study (Liu et al.) measured per-sentence citation recall at around 51,5% across the generative search engines evaluated — nearly half of the generated sentences lacked an adequate verifiable citation in the studied snapshot. [[18]](#ref-18)\n\n<a id=\"incertezas\"></a>\n\n## Uncertainties\n\n- It is not known whether the three-layer model actually helps make better budget decisions, or merely organizes the narrative without changing outcomes.\n- There is no public list of the exact criteria Google uses to choose which links appear as support in AI Overviews.\n- It is not known when (or whether) WebMCP will move past origin trial / community draft status and become a stable cross-browser standard.\n- It is not possible to separate, within Imperva-style \"automated\" traffic figures, what fraction is an agent acting on a user's behalf versus a common crawler or scraper.\n- There is no public study that converts synthetic \"AI visibility\" metrics into publishers' real impressions, clicks, or revenue.\n\n## References\n\n1. <a id=\"ref-1\"></a>[Chrome for Developers — A developer toolkit to make your website agent-ready](https://developer.chrome.com/blog/agent-ready-toolkit). Chrome for Developers. Accessed August 07, 2026.\n2. <a id=\"ref-2\"></a>[W3C — The W3C TAG Meeting in London, March 2026](https://www.w3.org/blog/2026/the-w3c-tag-meeting-in-london-march-2026/). W3C. Accessed August 07, 2026.\n3. <a id=\"ref-3\"></a>[Imperva — 2025 Bad Bot Report](https://www.imperva.com/resources/reports/2025-Bad-Bot-Report.pdf). Imperva. Accessed August 07, 2026.\n4. <a id=\"ref-4\"></a>[Google Search Central — Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide). Google. Accessed August 07, 2026.\n5. <a id=\"ref-5\"></a>[Google Search Central — AI features and your website](https://developers.google.com/search/docs/appearance/ai-features). Google. Accessed August 07, 2026.\n6. <a id=\"ref-6\"></a>[OpenAI — Introducing ChatGPT search](https://openai.com/index/introducing-chatgpt-search/). OpenAI. Accessed August 07, 2026.\n7. <a id=\"ref-7\"></a>[Chrome for Developers — WebMCP](https://developer.chrome.com/docs/ai/webmcp). Chrome for Developers. Accessed August 07, 2026.\n8. <a id=\"ref-8\"></a>[W3C Web Machine Learning Community Group — WebMCP API (Draft Community Group Report)](https://webmachinelearning.github.io/webmcp/). W3C WebML CG. Accessed August 07, 2026.\n9. <a id=\"ref-9\"></a>[Google Search Central Blog — Succeeding in AI search](https://developers.google.com/search/blog/2025/05/succeeding-in-ai-search). Google. Accessed August 07, 2026.\n10. <a id=\"ref-10\"></a>[OpenAI — Crawler / bot documentation](https://developers.openai.com/api/docs/bots). OpenAI. Accessed August 07, 2026.\n11. <a id=\"ref-11\"></a>[Anthropic — Does Anthropic crawl data from the web, and how can site owners block the crawler?](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler). Anthropic. Accessed August 07, 2026.\n12. <a id=\"ref-12\"></a>[Fastly — New Fastly threat research reveals AI crawlers make up almost 80% of AI bot traffic](https://www.fastly.com/press/press-releases/new-fastly-threat-research-reveals-ai-crawlers-make-up-almost-80-of-ai-bot). Fastly. Accessed August 07, 2026.\n13. <a id=\"ref-13\"></a>[Ahrefs — We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read](https://ahrefs.com/blog/llmstxt-study/). Ahrefs. Accessed August 07, 2026.\n14. <a id=\"ref-14\"></a>[Aggarwal et al. — GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735). arXiv. Accessed August 07, 2026.\n15. <a id=\"ref-15\"></a>[Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/). Pew Research Center. Accessed August 07, 2026.\n16. <a id=\"ref-16\"></a>[Digiday — Google AI Overviews linked to publisher referral traffic drops, new data shows](https://digiday.com/media/google-ai-overviews-linked-to-25-drop-in-publisher-referral-traffic-new-data-shows/). Digiday. Accessed August 07, 2026.\n17. <a id=\"ref-17\"></a>[Digiday — In graphic detail: AI platforms are driving more traffic, but not enough to offset zero-click search](https://digiday.com/media/in-graphic-detail-ai-platforms-are-driving-more-traffic-but-not-enough-to-offset-zero-click-search/). Digiday. Accessed August 07, 2026.\n18. <a id=\"ref-18\"></a>[Liu et al. — Evaluating Verifiability in Generative Search Engines](https://arxiv.org/abs/2304.09848). arXiv. Accessed August 07, 2026.\n"
}