{
 "blyg": "0.3",
 "id": "0vj1pw7wsjr259g77gd0mw6r76",
 "kind": "thread",
 "origin": "https://protocolsforbusiness.com/blyg/",
 "page": "t/0vj1pw7wsjr259g77gd0mw6r76/",
 "author": {
  "name": "Protocols for Business"
 },
 "created": "2026-11-18T10:45:39Z",
 "updated": "2026-11-18T10:51:23Z",
 "version": 2,
 "content_md": "# Agent readiness: what assistants read, trust and act on\n\n*Session of 16 November 2026.* 12 participants, 54 minutes. Guest: Navid Nouhi of [Credoble](https://credoble.com/).\n\nBuyers now reach businesses through AI assistants, and many will soon send software agents to compare, book and buy for them. A business in that position needs answers to three questions: what an assistant reads, what it trusts, and what an agent can do on the site. The guest, a growth practitioner with twenty years in paid search, SEO and growth, set out the current answers. The scheduled reading carries over to the next session.\n\n## The shift in discovery\n\n- **Zero-click search.** More than two thirds of Google searches end on Google, with the answer on the results page.\n- **Assistant use at scale.** 1.2 billion weekly ChatGPT users (September 2026 figures), before Claude and Perplexity. Between a quarter and a half of consumers use AI to find local businesses, depending on the survey.\n- **Prompts replace keywords.** An average of 23 words per prompt, with a recommendation as the expected output.\n- **Conversion.** Adobe: AI referral visits to US retail sites up 138% in a year, with a conversion rate 54% above other traffic. AI referrals remain a small share of total traffic.\n- **Non-human traffic.** Cloudflare: over half of internet traffic this year from bots and agents, and 1,700% growth in daily agents over one year.\n\n*Note: figures as cited by the guest, not checked against the original reports.*\n\n## A three-layer stack\n\n1. **SEO:** crawling, understanding and ranking of pages by a search engine.\n2. **AEO or GEO** (answer or generative engine optimization): a named, trusted place in an assistant's answer.\n3. **Agent readiness:** facts and actions (price, booking, purchase) open to software acting for a person. No official standard exists yet. The guest's focus: Google WebMCP, a way for a site to declare the actions on offer to an agent.\n\nThree differences from classic SEO:\n\n- **Rank no longer earns a citation.** 38% of cited pages sit outside the top ten results, and 62% of citations come from pages ranked 10 to 100 and beyond.\n- **Query fan-out.** An assistant splits one question into several background searches, in effect sub-agents, and narrows the answer from the results.\n- **Mentions over links.** Ahrefs, across 75 brands: brand mentions on the web correlate with AI visibility about three times more strongly than backlinks. Correlation, not proof.\n\n## Operating requirements\n\n- **Bot access by purpose.** Each vendor runs separate crawlers. OpenAI: GPTBot for training, OAI-SearchBot for its search index, ChatGPT-User for live user requests. A block on training and a block on live retrieval carry different consequences.\n- **Text first.** Assistants read plain HTML, not content rendered by JavaScript.\n- **Technical floor.** Sitemaps, clean links, indexable pages, page speed. Current agents browse through virtual machines, so load time still matters to them.\n- **Open trust pages.** In one audit, a company had blocked its terms and privacy pages from crawlers by mistake. After the fix, traffic rose: these pages signal a real company.\n- **One set of facts.** Address, phone and prices must match across the site and third-party profiles such as Google Business Profile. A mismatch carries a trust penalty. The guest found inconsistent prices on two pages of his own site.\n- **Proof over claims.** Assistants favour what generic content cannot fake: first-hand experience, proprietary data, a named expert. Google's E-E-A-T guidance applies, with trust as the top criterion. A new company with no third-party mentions gets no recommendation.\n- **Content from customer questions.** Meeting notes and FAQs mirror the prompts customers type into assistants. For long-form work such as an open book, quality and verifiable data outweigh format. The guest has no test results for the book route yet.\n\n## Discussion\n\n- **Gaming gets harder.** One participant: agents do more cognitive work per query and follow each user's own tuning, so a one-size ranking loses ground and gaming loses its payoff. The guest agreed and named a casualty: B2B pricing hidden behind \"book a meeting\" now works against the seller.\n- **Creators as a channel.** On creator-driven demand, the guest classed it as paid acquisition, separate from organic work. YouTube transcripts are indexable, though, and rank among the more cited sources in his audits. No measure of impact yet.\n- **Legibility versus capture.** The sharpest objection: \"your incentive is to be fully transparent, but at the end of the day, it's just cash that wins. So what's the move in that game?\" A transparent site hands its knowledge to well-funded rivals and to the labs, and an `agents.txt` claim offers no defence against a competitor's ad budget. With assistants as black boxes, a business has no means to confirm the absence of ad-driven steering. The guest's answer: publish the recipe and keep the method private. The questioner's reply: an assistant can reverse-engineer the ratios from the recipe. On the balance of paid and organic answers in ChatGPT ads, the guest had no answer: \"time will tell. Because, ultimately, you gotta pay for those GPUs.\"\n- **Capital has limits.** Another participant: even a well-funded player has finite capital and a choice of markets. Product architecture and positioning trade-offs leave niches, and organic reach costs less than ads over the medium term.\n- **Monetization unsettled.** The guest's read: subscriptions and agent fees ahead of ads in the near term, with Google best placed through decades of Gmail and search data, its own cloud and its own chips.\n- **Who the buyer is.** The session lead asked whether to sell to people or to agents. The guest: \"Ultimately, it's a people business,\" with agents as the filter on the options people see. Hence \"build for clankers.\"\n\n## What this means for protocol watching\n\nAn interpretation from this write-up, not stated in the session. Each layer of the stack runs on a protocol with a different owner:\n\n- crawler rules: robots.txt and per-purpose bots, set by each vendor;\n- trust signals: consistency checks across sources, set by each assistant's ranking;\n- action interfaces: WebMCP and similar proposals, still in draft.\n\nA business controls its own side of each protocol and holds no seat in the rule-setting. The objection on capture is a question about that asymmetry: who sets the protocol, and who pays for exceptions to it.\n\n## Open questions\n\n1. Does transparency to agents pay, or does it transfer value to the best-funded party?\n2. Which monetization model wins for assistants: ads, subscriptions or a hybrid?\n3. Which agent-readiness standard becomes the default?\n4. How can a business audit an assistant's recommendations for paid influence?\n\n## Next\n\n- **The deferred reading,** Anthropic's [Patterns and problems in multiagent systems](https://www.anthropic.com/research/multiagent-systems), carries over to the next session.\n- **Guest suggestions.** The litmus test: a person who runs an AI-native business. Suggestions go to the channel or through [offer a talk](https://protocolsforbusiness.com/research/#speak).\n- **Protocol watch** continues: one paragraph per member in the channel every two weeks. Starting questions: the [protocol watching guide](https://protocolsforbusiness.com/play/watching/).\n- **Free audit.** The guest offers members a free site audit through [Credoble](https://credoble.com/).\n- **Next session, 30 November:** Will Binns on Mru, an operating system for a thousand years of autonomous operation.\n\n## References\n\nAnthropic. \"Patterns and problems in multiagent systems.\" 2026. https://www.anthropic.com/research/multiagent-systems.\n\nGoogle. \"Creating helpful, reliable, people-first content.\" Google Search Central. https://developers.google.com/search/docs/fundamentals/creating-helpful-content.\n\nSource: [recording notes](https://pub-fb4a559f683a4e3b876823eb2bfe12a3.r2.dev/recordings/SIG-P4B/2026-11-16/1082444651946049567_1382801113224581240_1794843217216/summary.md) from OpenRecapper, the recording bot on the Protocol Institute Discord, and the session transcript, both in the [group's repository](https://github.com/protocolvision/sig-p4b/tree/main/ops/sessions/2026-11-16). Quotes come from the transcript, minus repeated words. The session has a recording.\n",
 "content_html": "<h1>Agent readiness: what assistants read, trust and act on</h1>\n<p><em>Session of 16 November 2026.</em> 12 participants, 54 minutes. Guest: Navid Nouhi of <a href=\"https://credoble.com/\">Credoble</a>.</p>\n<div class=\"blyg-tk-gen\"><p>Buyers now reach businesses through AI assistants, and many will soon send software agents to compare, book and buy for them. A business in that position needs answers to three questions: what an assistant reads, what it trusts, and what an agent can do on the site. The guest, a growth practitioner with twenty years in paid search, SEO and growth, set out the current answers. The scheduled reading carries over to the next session.</p>\n<h2>The shift in discovery</h2>\n<ul>\n<li><strong>Zero-click search.</strong> More than two thirds of Google searches end on Google, with the answer on the results page.</li>\n<li><strong>Assistant use at scale.</strong> 1.2 billion weekly ChatGPT users (September 2026 figures), before Claude and Perplexity. Between a quarter and a half of consumers use AI to find local businesses, depending on the survey.</li>\n<li><strong>Prompts replace keywords.</strong> An average of 23 words per prompt, with a recommendation as the expected output.</li>\n<li><strong>Conversion.</strong> Adobe: AI referral visits to US retail sites up 138% in a year, with a conversion rate 54% above other traffic. AI referrals remain a small share of total traffic.</li>\n<li><strong>Non-human traffic.</strong> Cloudflare: over half of internet traffic this year from bots and agents, and 1,700% growth in daily agents over one year.</li>\n</ul>\n<p><em>Note: figures as cited by the guest, not checked against the original reports.</em></p>\n<h2>A three-layer stack</h2>\n<ol>\n<li><strong>SEO:</strong> crawling, understanding and ranking of pages by a search engine.</li>\n<li><strong>AEO or GEO</strong> (answer or generative engine optimization): a named, trusted place in an assistant's answer.</li>\n<li><strong>Agent readiness:</strong> facts and actions (price, booking, purchase) open to software acting for a person. No official standard exists yet. The guest's focus: Google WebMCP, a way for a site to declare the actions on offer to an agent.</li>\n</ol>\n<p>Three differences from classic SEO:</p>\n<ul>\n<li><strong>Rank no longer earns a citation.</strong> 38% of cited pages sit outside the top ten results, and 62% of citations come from pages ranked 10 to 100 and beyond.</li>\n<li><strong>Query fan-out.</strong> An assistant splits one question into several background searches, in effect sub-agents, and narrows the answer from the results.</li>\n<li><strong>Mentions over links.</strong> Ahrefs, across 75 brands: brand mentions on the web correlate with AI visibility about three times more strongly than backlinks. Correlation, not proof.</li>\n</ul>\n<h2>Operating requirements</h2>\n<ul>\n<li><strong>Bot access by purpose.</strong> Each vendor runs separate crawlers. OpenAI: GPTBot for training, OAI-SearchBot for its search index, ChatGPT-User for live user requests. A block on training and a block on live retrieval carry different consequences.</li>\n<li><strong>Text first.</strong> Assistants read plain HTML, not content rendered by JavaScript.</li>\n<li><strong>Technical floor.</strong> Sitemaps, clean links, indexable pages, page speed. Current agents browse through virtual machines, so load time still matters to them.</li>\n<li><strong>Open trust pages.</strong> In one audit, a company had blocked its terms and privacy pages from crawlers by mistake. After the fix, traffic rose: these pages signal a real company.</li>\n<li><strong>One set of facts.</strong> Address, phone and prices must match across the site and third-party profiles such as Google Business Profile. A mismatch carries a trust penalty. The guest found inconsistent prices on two pages of his own site.</li>\n<li><strong>Proof over claims.</strong> Assistants favour what generic content cannot fake: first-hand experience, proprietary data, a named expert. Google's E-E-A-T guidance applies, with trust as the top criterion. A new company with no third-party mentions gets no recommendation.</li>\n<li><strong>Content from customer questions.</strong> Meeting notes and FAQs mirror the prompts customers type into assistants. For long-form work such as an open book, quality and verifiable data outweigh format. The guest has no test results for the book route yet.</li>\n</ul>\n<h2>Discussion</h2>\n<ul>\n<li><strong>Gaming gets harder.</strong> One participant: agents do more cognitive work per query and follow each user's own tuning, so a one-size ranking loses ground and gaming loses its payoff. The guest agreed and named a casualty: B2B pricing hidden behind \"book a meeting\" now works against the seller.</li>\n<li><strong>Creators as a channel.</strong> On creator-driven demand, the guest classed it as paid acquisition, separate from organic work. YouTube transcripts are indexable, though, and rank among the more cited sources in his audits. No measure of impact yet.</li>\n<li><strong>Legibility versus capture.</strong> The sharpest objection: \"your incentive is to be fully transparent, but at the end of the day, it's just cash that wins. So what's the move in that game?\" A transparent site hands its knowledge to well-funded rivals and to the labs, and an <code>agents.txt</code> claim offers no defence against a competitor's ad budget. With assistants as black boxes, a business has no means to confirm the absence of ad-driven steering. The guest's answer: publish the recipe and keep the method private. The questioner's reply: an assistant can reverse-engineer the ratios from the recipe. On the balance of paid and organic answers in ChatGPT ads, the guest had no answer: \"time will tell. Because, ultimately, you gotta pay for those GPUs.\"</li>\n<li><strong>Capital has limits.</strong> Another participant: even a well-funded player has finite capital and a choice of markets. Product architecture and positioning trade-offs leave niches, and organic reach costs less than ads over the medium term.</li>\n<li><strong>Monetization unsettled.</strong> The guest's read: subscriptions and agent fees ahead of ads in the near term, with Google best placed through decades of Gmail and search data, its own cloud and its own chips.</li>\n<li><strong>Who the buyer is.</strong> The session lead asked whether to sell to people or to agents. The guest: \"Ultimately, it's a people business,\" with agents as the filter on the options people see. Hence \"build for clankers.\"</li>\n</ul>\n<h2>What this means for protocol watching</h2>\n<p>An interpretation from this write-up, not stated in the session. Each layer of the stack runs on a protocol with a different owner:</p>\n<ul>\n<li>crawler rules: robots.txt and per-purpose bots, set by each vendor;</li>\n<li>trust signals: consistency checks across sources, set by each assistant's ranking;</li>\n<li>action interfaces: WebMCP and similar proposals, still in draft.</li>\n</ul>\n<p>A business controls its own side of each protocol and holds no seat in the rule-setting. The objection on capture is a question about that asymmetry: who sets the protocol, and who pays for exceptions to it.</p>\n<h2>Open questions</h2>\n<ol>\n<li>Does transparency to agents pay, or does it transfer value to the best-funded party?</li>\n<li>Which monetization model wins for assistants: ads, subscriptions or a hybrid?</li>\n<li>Which agent-readiness standard becomes the default?</li>\n<li>How can a business audit an assistant's recommendations for paid influence?</li>\n</ol>\n<h2>Next</h2>\n<ul>\n<li><strong>The deferred reading,</strong> Anthropic's <a href=\"https://www.anthropic.com/research/multiagent-systems\">Patterns and problems in multiagent systems</a>, carries over to the next session.</li>\n<li><strong>Guest suggestions.</strong> The litmus test: a person who runs an AI-native business. Suggestions go to the channel or through <a href=\"https://protocolsforbusiness.com/research/#speak\">offer a talk</a>.</li>\n<li><strong>Protocol watch</strong> continues: one paragraph per member in the channel every two weeks. Starting questions: the <a href=\"https://protocolsforbusiness.com/play/watching/\">protocol watching guide</a>.</li>\n<li><strong>Free audit.</strong> The guest offers members a free site audit through <a href=\"https://credoble.com/\">Credoble</a>.</li>\n<li><strong>Next session, 30 November:</strong> Will Binns on Mru, an operating system for a thousand years of autonomous operation.</li>\n</ul>\n<h2>References</h2>\n<p>Anthropic. \"Patterns and problems in multiagent systems.\" 2026. https://www.anthropic.com/research/multiagent-systems.</p>\n<p>Google. \"Creating helpful, reliable, people-first content.\" Google Search Central. https://developers.google.com/search/docs/fundamentals/creating-helpful-content.</p></div>\n<p>Source: <a href=\"https://pub-fb4a559f683a4e3b876823eb2bfe12a3.r2.dev/recordings/SIG-P4B/2026-11-16/1082444651946049567_1382801113224581240_1794843217216/summary.md\">recording notes</a> from OpenRecapper, the recording bot on the Protocol Institute Discord, and the session transcript, both in the <a href=\"https://github.com/protocolvision/sig-p4b/tree/main/ops/sessions/2026-11-16\">group's repository</a>. Quotes come from the transcript, minus repeated words. The session has a recording.</p>",
 "content_hash": "sha256:5b3b7cd2095ebb69d9a501caac0f7a6b588621dbd3773aab6828fdd67f840db6",
 "media": [],
 "transclusions": [],
 "generated": [
  {
   "sources": [],
   "model": "Claude Code",
   "at": "2026-11-18T10:51:23Z"
  }
 ],
 "changelog": [
  {
   "version": 1,
   "at": "2026-11-18T10:45:39Z",
   "note": "Session notes: 16 November, agent readiness with Credoble"
  },
  {
   "version": 2,
   "at": "2026-11-18T10:51:23Z",
   "note": "Session notes, 16 November: source note on the figures moves below the list"
  }
 ]
}
