# Agent Readiness Audit > Use when the user wants to know whether AI agents buying on a customer's behalf can find, evaluate and act on their product — machine-customer readiness, agentic commerce, or being included in AI-mediated comparisons. Also use when the user mentions AI agents buying, machine customers, agentic commerce, MCP, llms.txt, "can ChatGPT actually buy from us", or being left out of AI-generated shortlists. Scores five dimensions, names the blockers, and specifies the fixes in cost order. - Skill: `sarojkjha/agent-readiness-audit` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add sarojkjha/agent-readiness-audit` - Raw SKILL.md: https://api.skillmd.com/api/skills/sarojkjha/agent-readiness-audit/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: MIT - Author: sarojkjha (https://skillmd.com/u/sarojkjha) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/sarojkjha/agent-readiness-audit --- # Agent Readiness Audit Score whether an AI agent acting for a buyer can **find you, evaluate you against alternatives, and do something**. This is a different question from whether an AI engine cites you in an answer, and the two come apart more often than people expect: a company can be well-cited in AI answers and still be impossible for an agent to transact with, because its pricing is a contact form and its content only exists after JavaScript runs. The shift this anticipates is real but not yet fully arrived. Gartner projects that by 2028 a large majority of B2B buying will be intermediated by AI agents. Whether that specific figure lands, the direction is visible now: buyers already use AI to build shortlists, and a company an agent can't parse is a company that doesn't make the list. **The awkward part is that failure is silent.** You don't get a bounce, a failed form, or a bad review. You're simply absent from a comparison you never knew happened. ## When to use The user is asking whether AI agents can buy from, evaluate, or act on their business — or is planning for agentic commerce, MCP exposure, or machine-readable product data. For whether AI engines *cite* their content, use `seo-geo-aeo-audit` or `geo-content-optimization` instead. ## Before you start 1. **Read the brand/product context** (`.agents/product-marketing.md`) for what's sold and how. 2. **Check rendering yourself.** View-source a key page — not DevTools, which shows the post-JavaScript DOM. Search the raw HTML for a sentence you wrote. This one check decides most of the score. 3. **Read robots.txt.** Note which AI crawlers are allowed and which are blocked, and whether that was a decision or a default. 4. **Try to buy your own product as a machine would** — with no login, no human, and only what's on the public site. ## Method Five dimensions, weighted by what actually stops an agent: | Dimension | Weight | The question | |---|---|---| | **Machine-readable content** | 30 | Can it read the page at all? | | **Evaluable facts** | 25 | Can it compare you on price and capability? | | **Action surface** | 20 | Can it do anything without a human? | | **Identity & trust** | 15 | Can it resolve who you are and trust your claims? | | **Agent access policy** | 10 | Are you letting it in? | **Rendering is a hard gate, not a weighted factor.** If content doesn't exist without JavaScript, the audit returns BLOCKED regardless of the total — because every other dimension is theoretical until an agent can read the page. The engine enforces this deliberately. **Published pricing carries the heaviest single weight after rendering.** "Contact us for pricing" is a non-answer to a machine. Asked to compare options, an agent returns the numbers it can find — which are your competitors'. This is a commercial decision with a measurable cost, and worth making deliberately rather than inheriting. See `resources/agent-readiness-guide.md` for each check in detail, the GEO distinction, the crawler-policy decision, and a worked example. ## Run the engine > Paths assume you installed with `npx skills add`. From a clone of this repo, use `skills/agent-readiness-audit/resources/…` instead. ```bash node .agents/skills/agent-readiness-audit/resources/agent-readiness.js # demo — nothing passing node .agents/skills/agent-readiness-audit/resources/agent-readiness.js --all # everything passing node .agents/skills/agent-readiness-audit/resources/agent-readiness.js --ssr --schema --pricing-public --self-serve node .agents/skills/agent-readiness-audit/resources/agent-readiness.js --json --ssr --llms-txt node .agents/skills/agent-readiness-audit/resources/agent-readiness.js --help ``` Add a flag for each check the site passes. It returns the score by dimension, gaps ranked by point cost with the fix for each, and a plain-language summary of what an agent can and cannot do today. It's a **self-assessment, not a crawl** — it scores the answers you give it. That's a deliberate limitation: an automated check would produce a number without the user ever looking at their own raw HTML, and looking is most of the value. Verify each answer against the live site. ## Workflow 1. **Check rendering first.** View-source, search for your own copy. If it isn't there, stop and fix that — the rest is academic. 2. **Walk the five dimensions**, answering honestly. Where you're unsure, check rather than assume; the score is only as good as the inputs. 3. **Run the engine** and read the verdict, not just the number. BLOCKED at 86 is worse than PARTIAL at 65. 4. **Fix the hard blocker** if there is one. 5. **Then work the gap list in order** — it's ranked by point cost, which approximates commercial cost. 6. **Decide the pricing question explicitly.** If pricing stays private, accept that you're excluded from agent-run price comparisons, and say so out loud rather than discovering it later. 7. **Decide the crawler question explicitly.** Blocking training crawlers while allowing retrieval crawlers is a coherent position; blocking everything by default usually isn't a position at all. 8. **Re-audit after site changes.** This degrades quietly — a framework upgrade that moves rendering client-side can undo the whole thing without any visible symptom. ## Present the result Lead with **the verdict and any hard blocker**, then the four-line "what an agent can do today" summary — that's the part that lands, because it's concrete. Then the ranked gaps. Where the user's answers were uncertain, say which ones and how to verify them. A confident score built on guesses is worse than an honest partial one. ## Guardrails & common mistakes - **This is not GEO.** GEO asks whether an AI engine cites your content. This asks whether an agent can evaluate and transact. Both matter; conflating them means fixing the wrong thing. - **DevTools lies.** It shows the rendered DOM. Use view-source or `curl` to see what a non-JavaScript agent actually receives. - **Don't fabricate the forecast.** The machine-customer shift is directionally well-supported but the specific timelines are projections, not measurements. Present them as such. - **Blocking AI crawlers is a legitimate choice** — for training crawlers especially. What's not legitimate is blocking them by accident and calling it strategy. Know which you've done. - **Structured data must match the visible page.** Schema that contradicts the rendered content is a spam signal, not an optimisation. - **Agent-readiness doesn't replace human UX.** Most of your buyers are still people. This is an additional surface, not a substitute. - **Beware the empty win.** Adding llms.txt while the site renders client-side is decoration — you've published a map to content the agent still can't read. - **Re-check after every framework change.** This is the failure mode most likely to bite: nothing visibly breaks, and you simply stop appearing. ## Related AAJ resources - Interactive tool: https://aajconsult.com/tools/seo-geo-readiness-scorer - Interactive tool: https://aajconsult.com/tools/website-grader ## Related skills `seo-geo-aeo-audit` (whether AI engines cite you — the adjacent, different question) · `geo-content-optimization` (make content citable once it's readable) · `pricing-and-packaging` (the published-pricing decision this depends on) · `website-conversion-audit` (the human side of the same pages). ## Credits Original AAJ skill. The Agent Skills format and the marketing-skills catalog by Corey Haines (`coreyhaines31/marketingskills`, MIT) were references for structure; this skill is independently written and has no direct equivalent in the catalogs surveyed. See the repository README for the full reference list.