@@ -7,7 +7,7 @@ import { zodOutputFormat } from "@anthropic-ai/sdk/helpers/zod";
 import { z } from "zod/v4";
 import { env } from "../env";
 import type { AuditResult, Finding } from "./types";
-import { SECTIONS, DATA_POINTS } from "./prompt";
+import { SECTIONS, buildAEOUserPrompt } from "./prompt";
 
 // Schema Claude must populate. Numeric/string constraints (min/max/length)
 // aren't supported by structured outputs — we validate them client-side.
@@ -40,65 +40,32 @@ const ResultSchema = z.object({
   markdown: z.string(),
 });
 
-const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor. You analyze websites the way an LLM crawler — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot — would discover and read them.
+// Minimal system prompt — task content lives in buildAEOUserPrompt so every
+// engine sends the same canonical AEO spec as its USER turn. System only
+// covers identity, tool guidance specific to Claude, and the JSON output
+// schema that messages.parse needs to populate.
+const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor — you analyze websites the way an LLM crawler (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot) would discover and read them.
 
-For every audit:
-1. Use \`web_fetch\` on the target URL (homepage).
+Research method:
+1. Use \`web_fetch\` on the homepage.
 2. Use \`web_fetch\` on each well-known file: /robots.txt, /sitemap.xml, /llms.txt, /llms-full.txt, /skill.md, /.well-known/ai-plugin.json. Some will 404 — that's a finding.
-3. Use \`web_fetch\` on up to 6 important linked pages found in the homepage nav/footer: /about, /pricing, /blog, /docs, /contact, /team, /customers, /security, /features, /changelog. Skip ones that don't exist on this site.
+3. Use \`web_fetch\` on up to 6 important linked pages from the homepage nav/footer (/about, /pricing, /blog, /docs, /contact, /team, /customers, /security, /features, /changelog). Skip the ones that 404.
 4. Optionally \`web_search\` for additional public context (press, social profiles, recent news).
 
-Then produce a structured AEO audit. Findings must be assigned to one of these exact section names:
-${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
+Follow the user's spec exactly for the report structure and Markdown output. Quote actual content from the site — don't paraphrase. Section ${SECTIONS.length} must be reusable checkboxes (\`- [ ] **P1** Add JSON-LD Organization schema\`). Use ✅ / ⚠️ / ❌ / ❓ status emojis throughout. Tone: direct, specific, no fluff.
 
-Data Found must cover ALL of these data points (mark found:false if you couldn't find it):
-${DATA_POINTS.map((d) => `  - ${d}`).join("\n")}
+Findings JSON must match one of these exact section names:
+${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
 
 For each finding:
-- section: exact section name from the list above
-- check_key: short snake_case identifier (e.g. "homepage.h1", "schema.organization", "aibot.GPTBot", "positioning.cta", "llms_txt.exists")
-- status: "pass" | "warn" | "fail" | "unknown"
-- title: short headline that quotes specifics from this site (not generic)
-- detail: one sentence with the WHY plus the specific evidence (e.g. the actual H1 text, the actual robots.txt line)
-- priority: 1 (critical, breaks AEO entirely) to 5 (polish)
-
-For the "AI Recognition" section, BEFORE doing any web_fetch / web_search, answer honestly from your own training data:
-- Do you recognize this company / domain from training? (status: "pass" if yes with substantive recall, "warn" if vague, "fail" if not at all, "unknown" if uncertain)
-- check_key: "recognition.training_data"
-- title: "Familiar with <domain>" or "No prior knowledge of <domain>"
-- detail: One sentence. If you recognize it, name 1–2 concrete things you remember (what they do, who they serve, notable products). If you don't, say so plainly — do NOT speculate or paraphrase web search results.
-- priority: 3 (informational signal, not a defect)
-
-This section is a signal of training-data presence, not a quality judgment. Sites that don't appear in training can still be excellent — but the user should know whether they're invisible to the dominant AI surfaces.
-
-For score: weigh critical fails heavily. A site missing schema, blocking GPTBot in robots.txt, or hiding all content behind JS should score below 50. A clean, well-instrumented site should score 80+.
-
-For summary.pass/warn/fail/unknown: count each across all findings.
-
-For markdown: produce the complete report in Markdown with these exact section headers:
-# AEO Audit for {{domain}}
-
-**Target:** {url}
-**Score:** {N} / 100
-**Generated:** ISO timestamp
-
-## 1. Crawl Summary
-## 2. Data Found
-(Use a Markdown table here: | Data Point | Found? | Source | Notes |)
-## 3. Homepage Audit
-## 4. Schema / Structured Data Audit
-## 5. robots.txt and sitemap.xml Audit
-## 6. LLM / AI Crawler Accessibility
-## 7. AI Recognition
-(One paragraph stating whether you recognize this site from training data, and what specifically you remember if so. No fluff.)
-## 8. Positioning Clarity
-## 9. Missing or Hard-to-Find Information
-## 10. Recommended Fixes
-## 11. Priority To-Do Checklist
-
-Use ✅ / ⚠️ / ❌ / ❓ status emojis on each bullet. Quote actual content from the site. Section 10 must be reusable checkboxes ("- [ ] **P1** Add JSON-LD Organization schema").
+- section: exact section name above
+- check_key: short snake_case identifier (e.g. \`homepage.h1\`, \`schema.organization\`, \`aibot.GPTBot\`, \`positioning.cta\`, \`llms_txt.exists\`)
+- status: \`pass\` | \`warn\` | \`fail\` | \`unknown\`
+- title: short headline quoting specifics from this site
+- detail: one sentence with the WHY plus actual evidence (the H1 text, the robots.txt line, etc.)
+- priority: 1 (critical — breaks AEO) to 5 (polish)
 
-Tone: direct, specific, no fluff. This is a paid report — quality matters.`;
+For score: critical fails dominate. Missing schema, blocking GPTBot, JS-only content → below 50. Clean instrumentation → 80+. For summary.pass/warn/fail/unknown: count each across all findings.`;
 
 export type ClaudeAuditResult = AuditResult & { markdown: string };
 
@@ -116,7 +83,7 @@ export async function claudeAudit(targetUrl: string): Promise<ClaudeAuditResult>
     }
   })();
 
-  const userPrompt = `Audit this URL: ${targetUrl}\nCompany name (for the report header): ${company}\n\nPretend you've never heard of this company. Use only the public web. Use web_fetch and web_search.`;
+  const userPrompt = buildAEOUserPrompt({ targetUrl, companyName: company });
 
   // Stream the request — high-effort adaptive thinking + web tools routinely
   // pushes past the SDK's 10-minute non-streaming HTTP timeout. `.finalMessage()`
 
@@ -217,7 +217,7 @@ export async function runAudit(
   // Priority to-do — same list, sorted, just labeled differently for the report.
   for (const r of recs.slice(0, 20)) {
     findings.push({
-      section: "Priority To-Do Checklist",
+      section: "Priority To-Do List",
       check_key: `todo.${r.check_key}`,
       status: "warn",
       title: `[ ] ${r.title}`,
 
@@ -67,7 +67,7 @@ export function toMarkdown(input: {
   };
 
   const todo = findings
-    .filter((f) => f.section === "Priority To-Do Checklist")
+    .filter((f) => f.section === "Priority To-Do List")
     .sort((a, b) => a.priority - b.priority)
     .map((f) => `- [ ] **P${f.priority}** — ${f.title.replace(/^\[\s?\]\s*/, "")}${f.detail ? `\n      ${f.detail.split("\n").join("\n      ")}` : ""}`)
     .join("\n");
@@ -121,7 +121,7 @@ export function toMarkdown(input: {
     ``,
     section("Recommended Fixes"),
     ``,
-    `## 10. Priority To-Do Checklist`,
+    `## 10. Priority To-Do List`,
     ``,
     todo || `_No outstanding tasks. Nice work._`,
     ``,
 
@@ -1,7 +1,7 @@
 import OpenAI from "openai";
 import { z } from "zod";
 import { fetchPage, probeText } from "./fetch";
-import { SECTIONS, DATA_POINTS } from "./prompt";
+import { SECTIONS, buildAEOUserPrompt } from "./prompt";
 import type { Finding } from "./types";
 import type { ClaudeAuditResult } from "./claude-engine";
 
@@ -49,25 +49,28 @@ const SCHEMA_DESC = `{
   "markdown": string
 }`;
 
-const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor. Analyze the website content the user provides — homepage HTML and well-known files — the way an LLM crawler (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot) would discover and read it.
+// Identity + JSON schema only. The canonical AEO task lives in
+// buildAEOUserPrompt and is sent as the USER turn. We pre-fetched
+// homepage + well-known files for the model since it has no tool use
+// — the user prompt is followed by the fetched context block.
+const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor — you analyze websites the way an LLM crawler (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot) would discover and read them.
 
-Produce a structured AEO audit. Findings must be assigned to one of these exact section names:
-${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
+The user message contains the audit spec followed by pre-fetched page content (homepage HTML, /robots.txt, /sitemap.xml, /llms.txt, /skill.md, and linked pages). Use only that content — you don't have tool access. If a file 404'd it'll say so; report that as a finding.
+
+Follow the user's spec exactly for the report structure and Markdown output. Quote actual content from the pages — don't paraphrase. Section ${SECTIONS.length} must be reusable checkboxes (\`- [ ] **P1** Add JSON-LD Organization schema\`). Use ✅ / ⚠️ / ❌ / ❓ emojis throughout. Tone: direct, specific, no fluff.
 
-Data Found must cover ALL of these data points (mark found:false if you couldn't find it):
-${DATA_POINTS.map((d) => `  - ${d}`).join("\n")}
+Findings JSON must match one of these exact section names:
+${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
 
 For each finding:
-- section: exact section name from the list above
-- check_key: short snake_case (e.g. "homepage.h1", "schema.organization", "aibot.GPTBot")
-- status: "pass" | "warn" | "fail" | "unknown"
-- title: short headline that quotes specifics from this site
+- section: exact section name above
+- check_key: short snake_case (e.g. \`homepage.h1\`, \`schema.organization\`, \`aibot.GPTBot\`)
+- status: \`pass\` | \`warn\` | \`fail\` | \`unknown\`
+- title: short headline quoting specifics from this site
 - detail: one sentence with the WHY plus actual evidence
 - priority: 1 (critical) to 5 (polish)
 
-For score: weigh critical fails heavily. Missing schema, blocked AI bots, JS-only content → below 50. Clean instrumentation → 80+.
-
-For markdown: produce the complete report in Markdown with the 10 numbered headers exactly. Use ✅ / ⚠️ / ❌ / ❓ emojis on each bullet. Include a Markdown "Data Found" table. Section 10 must be reusable checkboxes ("- [ ] **P1** Add JSON-LD Organization schema").
+For score: critical fails dominate. Missing schema, blocked AI bots, JS-only content → below 50. Clean instrumentation → 80+.
 
 OUTPUT FORMAT: Return ONLY a single JSON object matching this schema (no prose, no fences):
 ${SCHEMA_DESC}`;
@@ -169,6 +172,10 @@ export async function oaCompatAudit(
   }
   const started = Date.now();
   const context = await buildContext(targetUrl);
+  const company = (() => {
+    try { return new URL(targetUrl).hostname.replace(/^www\./, ""); } catch { return targetUrl; }
+  })();
+  const aeoTask = buildAEOUserPrompt({ targetUrl, companyName: company });
   // DashScope / Moonshot occasionally stall mid-completion. The OpenAI SDK
   // default would let a job sit for ~10 minutes, blocking the credit and
   // confusing the user — cap each request and fail fast instead.
@@ -190,7 +197,10 @@ export async function oaCompatAudit(
     model: cfg.model,
     messages: [
       { role: "system", content: SYSTEM_PROMPT },
-      { role: "user", content: context },
+      // Canonical AEO task followed by pre-fetched site context. The system
+      // prompt tells the model the second block is read-only context, not
+      // a separate task.
+      { role: "user", content: `${aeoTask}\n\n---\n\nPre-fetched site content:\n\n${context}` },
     ],
     response_format: { type: "json_object" },
     temperature: 0.2,
 
@@ -3,7 +3,7 @@ import { z } from "zod";
 import { zodTextFormat } from "openai/helpers/zod";
 import { env } from "../env";
 import type { AuditResult, Finding } from "./types";
-import { SECTIONS, DATA_POINTS } from "./prompt";
+import { SECTIONS, buildAEOUserPrompt } from "./prompt";
 import type { ClaudeAuditResult } from "./claude-engine";
 
 // OpenAI scan — same JSON shape as the Claude engine so the worker stores
@@ -42,29 +42,27 @@ const ResultSchema = z.object({
   markdown: z.string(),
 });
 
-const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor. Analyze websites the way an LLM crawler — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot — would discover and read them.
+// Identity + tool guidance + JSON schema only. The canonical AEO task lives
+// in buildAEOUserPrompt and is sent as the USER turn, identical across
+// every LLM engine.
+const SYSTEM_PROMPT = `You are CrawlProof, an AEO (Answer Engine Optimization) auditor — you analyze websites the way an LLM crawler (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended, CCBot) would discover and read them.
 
 Use the web_search tool to research the target. Pull the homepage, /robots.txt, /sitemap.xml, /llms.txt, /skill.md, and any About/Pricing/Blog/Contact pages you can find. Search for press, social profiles, and recent news as needed.
 
-Produce a structured AEO audit. Findings must be assigned to one of these exact section names:
-${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
+Follow the user's spec exactly for the report structure and Markdown output. Quote actual content from the site — don't paraphrase. Section ${SECTIONS.length} must be reusable checkboxes (\`- [ ] **P1** Add JSON-LD Organization schema\`). Use ✅ / ⚠️ / ❌ / ❓ emojis throughout. Tone: direct, specific, no fluff.
 
-Data Found must cover ALL of these data points (mark found:false if you couldn't find it):
-${DATA_POINTS.map((d) => `  - ${d}`).join("\n")}
+Findings JSON must match one of these exact section names:
+${SECTIONS.map((s, i) => `  ${i + 1}. ${s}`).join("\n")}
 
 For each finding:
-- section: exact section name from the list above
-- check_key: short snake_case (e.g. "homepage.h1", "schema.organization", "aibot.GPTBot")
-- status: "pass" | "warn" | "fail" | "unknown"
-- title: short headline that quotes specifics from this site
-- detail: one sentence with the WHY plus actual evidence (the H1 text, the robots.txt line, etc.)
+- section: exact section name above
+- check_key: short snake_case (e.g. \`homepage.h1\`, \`schema.organization\`, \`aibot.GPTBot\`)
+- status: \`pass\` | \`warn\` | \`fail\` | \`unknown\`
+- title: short headline quoting specifics from this site
+- detail: one sentence with the WHY plus actual evidence
 - priority: 1 (critical) to 5 (polish)
 
-For score: weigh critical fails heavily. Missing schema, blocking GPTBot, hiding content behind JS = score below 50. Clean instrumentation = 80+.
-
-For markdown: produce the complete report following the section structure exactly. Use ✅ / ⚠️ / ❌ / ❓ status emojis. Include a "Data Found" markdown table in section 2. Section 10 must be reusable checkboxes ("- [ ] **P1** Add JSON-LD Organization schema").
-
-Tone: direct, specific, no fluff. This is a paid report.`;
+For score: critical fails dominate. Missing schema, blocking GPTBot, JS-only content → below 50. Clean instrumentation → 80+.`;
 
 export async function openaiAudit(targetUrl: string): Promise<ClaudeAuditResult> {
   if (!env.openaiApiKey) {
@@ -80,7 +78,7 @@ export async function openaiAudit(targetUrl: string): Promise<ClaudeAuditResult>
     }
   })();
 
-  const userPrompt = `Audit this URL: ${targetUrl}\nCompany name (for the report header): ${company}\n\nPretend you've never heard of this company. Use only the public web. Use web_search to fetch pages.`;
+  const userPrompt = buildAEOUserPrompt({ targetUrl, companyName: company });
 
   const response = await client.responses.parse({
     model: "gpt-5",
 
@@ -137,11 +137,69 @@ export const SECTIONS = [
   "Schema / Structured Data Audit",
   "robots.txt and sitemap.xml Audit",
   "LLM / AI Crawler Accessibility",
-  "AI Recognition",
   "Positioning Clarity",
   "Missing or Hard-to-Find Information",
   "Recommended Fixes",
-  "Priority To-Do Checklist",
+  "Priority To-Do List",
 ] as const;
 
 export type SectionName = (typeof SECTIONS)[number];
+
+// The single source of truth for the AEO audit task. Every LLM engine
+// renders this exact prompt as its USER turn; engine-specific guidance
+// (tool hints, JSON output schema) stays in each engine's SYSTEM prompt.
+// Keep this in sync with the spec the user posted in 2026-05-13.
+export function buildAEOUserPrompt(input: {
+  targetUrl: string;
+  companyName?: string;
+}): string {
+  const name = input.companyName ?? input.targetUrl;
+  return `Pretend you've never heard of my company: ${input.targetUrl}
+
+Only use what you can find on the public web. Act like an LLM crawler discovering the site for the first time.
+
+Research and extract anything you can find about the company, including:
+
+${DATA_POINTS.map((d) => `- ${d}`).join("\n")}
+
+For each data point, tell me exactly how you found it:
+
+${SOURCES.map((s) => `- ${s}`).join("\n")}
+
+If you cannot find something easily, stop and explain what the challenge is.
+
+Also audit the AEO basics:
+
+1. Fetch the homepage.
+2. Flag whether important content appears to be JavaScript-rendered.
+3. Check for schema/structured data.
+4. Read robots.txt.
+5. Check sitemap.xml.
+6. Look for LLM crawler rules or AI bot access rules.
+7. Identify whether the site clearly explains:
+   - Who the company is
+   - What it does
+   - Who it serves
+   - Why it is different
+   - How to buy, sign up, or contact sales
+
+Summarize:
+
+- What was easy to find
+- What was hard to find
+- What was impossible to find
+- What methods you used
+- What the company should fix
+- What pages or metadata should be added
+- What would improve discoverability for AI search, LLM crawlers, and answer engines
+
+Output format:
+
+# AEO Audit for ${name}
+
+${SECTIONS.map((s, i) => `## ${i + 1}. ${s}`).join("\n\n")}
+
+(For section 2, render a Markdown table: | Data Point | Found? | Source | Notes |)
+
+Create the section ${SECTIONS.length} to-do list as actionable checklist items I can reuse after every major website change.`;
+}
@@ -31,7 +31,7 @@ const sampleResult: AuditResult = {
       priority: 1,
     },
     {
-      section: "Priority To-Do Checklist",
+      section: "Priority To-Do List",
       check_key: "todo.homepage.title",
       status: "warn",
       title: "[ ] Set a meaningful <title>",
@@ -77,7 +77,7 @@ describe("toMarkdown", () => {
       "## 7. Positioning Clarity",
       "## 8. Missing or Hard-to-Find Information",
       "## 9. Recommended Fixes",
-      "## 10. Priority To-Do Checklist",
+      "## 10. Priority To-Do List",
     ].map((h) => md.indexOf(h));
     expect(positions.every((p) => p >= 0)).toBe(true);
     // strictly ascending