diff --git a/src/tools/web.ts b/src/tools/web.ts
index 6a059fd..d046fc2 100644
--- a/src/tools/web.ts
+++ b/src/tools/web.ts
@@ -1093,6 +1093,50 @@ async function extractAnswerFromResults(query: string, rawResults: string): Prom
   }
 }
 
+/**
+ * Text that actually contains specifications, as opposed to a page that merely talks about them.
+ * Measured 2026-08-11 against the three sources this path surfaces for a GPU comparison:
+ *   nvidia.com product page  → 6016 chars, ZERO matches (JS-rendered spec table; the fetched text
+ *                              is marketing copy plus "This site requires Javascript")
+ *   techpowerup.com          → 275 chars, ZERO matches (bot wall)
+ *   nanoreview.net           → matches CUDA/Boost Clock/Bandwidth/TFLOPS/GB-s/Base Clock
+ * So the manufacturer's own page is the worst of the three here, and this filter is what keeps
+ * 6KB of "Game Changer" out of the model's context.
+ */
+const SPEC_SIGNAL = /\d[\d,.]*\s*(GB\/s|GB|MB|MHz|GHz|TFLOPS|nm)\b|\b(cuda\s*cores?|boost\s*clock|base\s*clock|memory\s*(bus|interface|bandwidth)|cores|tmus)\b\s*[:：]?\s*\d/i;
+
+/** Does this fetched page body actually carry specifications? See SPEC_SIGNAL for the measurements. */
+export function hasSpecContent(text: string): boolean {
+  return SPEC_SIGNAL.test(String(text || ''));
+}
+
+/**
+ * Opens the top results until one of them actually contains specification text.
+ *
+ * web_search returns titles and snippets; the figures a spec question needs live in the page body.
+ * The tool description has told the model to web_fetch result URLs all along, and it does not:
+ * across a 3,300-line production log (2026-08-11) web_fetch was called exactly once, for a weather
+ * page, never for any of the GPU spec questions in that same log. So the answer was always written
+ * off snippets alone, which is why it stayed at the level of "80 라인업이 60 라인업보다 빠르다".
+ *
+ * Stops at the first page with real spec content: one good source answers the question, and each
+ * extra page is context the model has to read past.
+ */
+async function fetchFirstSpecPage(results: any[], maxTries = 2): Promise<string | null> {
+  const urls = (Array.isArray(results) ? results : [])
+    .map(r => String(r?.url || '').trim())
+    .filter(u => /^https?:\/\//i.test(u))
+    .slice(0, maxTries);
+  for (const url of urls) {
+    try {
+      const r = await executeWebFetch({ url, max_chars: 3000 });
+      const text = String(r.stdout || '').trim();
+      if (r.success && hasSpecContent(text)) return `[본문: ${url}]\n${text}`;
+    } catch { /* a dead or walled URL is the normal case here, not an error worth surfacing */ }
+  }
+  return null;
+}
+
 /**
  * Runs the combined query plus one search per entity, and hands the model all three labelled.
  * Partial failure is fine — any section that came back with results is still grounding the model
@@ -1119,10 +1163,31 @@ async function runComparisonSearch(
 
   if (!sections.length) return combined;
   console.log(`[v2] web_search comparison split: "${args.query}" → "${split.a}" + "${split.b}"`);
+
+  // Snippets name the products; only the page body has the numbers. Fetch one real spec page per
+  // entity, in parallel, since the model will not do it itself.
+  const [pageA, pageB] = await Promise.all([
+    fetchFirstSpecPage((a.data as any)?.results || []),
+    fetchFirstSpecPage((b.data as any)?.results || []),
+  ]);
+  const fetched = [pageA, pageB].filter(Boolean) as string[];
+  if (fetched.length) console.log(`[v2] web_search comparison split: auto-fetched ${fetched.length} spec page(s)`);
+
+  const body = fetched.length
+    ? `${sections.join('\n\n')}\n\n${fetched.join('\n\n')}`
+    : sections.join('\n\n');
+  const lead = fetched.length
+    ? '비교 질문이라 각 대상을 따로 검색하고, 스펙이 실제로 실린 페이지 본문까지 가져왔습니다. 아래 [본문] 블록의 수치를 근거로 답하세요.'
+    : '비교 질문이라 각 대상을 따로 검색했습니다. 아래 세 검색 결과를 모두 근거로 쓰세요.';
+
   return {
     success: true,
-    data: { ...(combined.data as any || {}), comparison_split: [split.a, split.b] },
-    stdout: `비교 질문이라 각 대상을 따로 검색했습니다. 아래 세 검색 결과를 모두 근거로 쓰세요.\n\n${sections.join('\n\n')}`,
+    data: {
+      ...(combined.data as any || {}),
+      comparison_split: [split.a, split.b],
+      auto_fetched_pages: fetched.length,
+    },
+    stdout: `${lead}\n\n${body}`,
   };
 }
 
diff --git a/tests/web-comparison-split.test.ts b/tests/web-comparison-split.test.ts
index 0e15999..0f26161 100644
--- a/tests/web-comparison-split.test.ts
+++ b/tests/web-comparison-split.test.ts
@@ -13,7 +13,7 @@
 
 import { test, describe } from 'node:test';
 import assert from 'node:assert/strict';
-import { splitComparisonQuery } from '../src/tools/web';
+import { splitComparisonQuery, hasSpecContent } from '../src/tools/web';
 
 describe('쪼개야 하는 쿼리', () => {
   test('실제 사고 쿼리 — 비교 동사는 버리고 속성어는 남긴다', () => {
@@ -71,3 +71,47 @@ describe('쪼개면 안 되는 쿼리', () => {
     assert.equal(splitComparisonQuery(null as any), null);
   });
 });
+
+/**
+ * hasSpecContent — 가져온 본문에 진짜 스펙이 실려 있는지
+ *
+ * 이 판정이 필요한 이유는 2026-08-11 실측이다. 비교 검색이 물어오는 상위 세 출처를 실제로
+ * 가져와 봤더니, 제조사 공식 페이지가 셋 중 가장 쓸모없었다:
+ *   nvidia.com   6016자, 스펙 키워드 0개 (스펙 표가 JS 렌더링, 본문은 홍보 문구)
+ *   techpowerup   275자, 스펙 키워드 0개 (봇 차단)
+ *   nanoreview   Cores/TMUs/Boost Clock/Bandwidth/TFLOPS 전부 있음
+ * 이 필터가 없으면 "Game Changer" 6KB가 모델 컨텍스트를 차지한다.
+ */
+describe('hasSpecContent — 본문에 수치가 실렸는지 판정', () => {
+  test('NVIDIA 공식 페이지 본문은 통과시키지 않는다 (실측 발췌)', () => {
+    assert.equal(hasSpecContent(
+      'GeForce RTX 5060 Family Graphics Cards | NVIDIA. This site requires Javascript in order to '
+      + 'view all its content. Game Changer. The NVIDIA GeForce RTX 5060 Ti and RTX 5060, powered by '
+      + 'the NVIDIA Blackwell architecture, enables game-changing AI capabilities.',
+    ), false);
+  });
+
+  test('봇 차단 페이지도 통과시키지 않는다 (실측 발췌)', () => {
+    assert.equal(hasSpecContent(
+      'NVIDIA GeForce RTX 5060 Specs | TechPowerUp GPU Database. Automated bot check in progress. '
+      + 'Your browser must support Javascript.',
+    ), false);
+  });
+
+  test('nanoreview 본문은 통과시킨다 (실측 발췌)', () => {
+    assert.equal(hasSpecContent(
+      'GeForce RTX 5060. Graphics Processor: GB206-250. Cores: 3840. TMUs / ROPs: 120 / 48.',
+    ), true);
+  });
+
+  test('대역폭·클럭 표기도 스펙으로 본다', () => {
+    assert.equal(hasSpecContent('Memory Bandwidth: 936 GB/s'), true);
+    assert.equal(hasSpecContent('Boost Clock: 2505 MHz'), true);
+    assert.equal(hasSpecContent('CUDA Cores: 9728'), true);
+  });
+
+  test('숫자 없는 일반 문장은 스펙이 아니다', () => {
+    assert.equal(hasSpecContent('이 카드는 게이밍 성능이 매우 뛰어납니다.'), false);
+    assert.equal(hasSpecContent(''), false);
+  });
+});