diff --git a/backend/apps/web/search_quality_bench.py b/backend/apps/web/search_quality_bench.py
new file mode 100644
index 00000000..347f5d3f
--- /dev/null
+++ b/backend/apps/web/search_quality_bench.py
@@ -0,0 +1,112 @@
+"""Objective quality benchmark for keyless web search, with no third-party reference (ENG-233).
+
+The original plan was "50 queries vs Google SERP, judged blind". Measured 2026-08-13, that
+reference cannot be obtained: google.com/search returns HTTP 200 with ~92KB of JavaScript shell,
+zero `<h3>`, zero `/url?q=`, zero `data-ved`. Not a captcha, not a consent wall, just no results in
+the HTML. A scraper checking status codes would report success and produce an EMPTY reference arm,
+which then makes our results "match Google" by construction. That is worse than no benchmark.
+
+So this measures what can be measured without asking a competitor anything: for a corpus of fact
+queries whose answers are known and stable, does our search surface the answer at all, how fast,
+and from how many distinct sources.
+
+    answered      the known answer string appears in the returned text
+    latency       wall clock per query
+    diversity     distinct registrable domains in the result text
+    engine mix    which rung actually won, so a silent single-engine dependency is visible
+
+Run it, do not import it:
+    PYTHONPATH=$PWD backend/.venv/bin/python backend/apps/web/search_quality_bench.py
+"""
+import asyncio
+import json
+import re
+import statistics
+import sys
+import time
+from typing import Dict, List, Optional, Tuple
+
+from typeguard import typechecked
+
+# Fact queries with answers that do not drift. Kept boring on purpose: a benchmark whose expected
+# answer changes with the news measures the news, not the search.
+CORPUS: List[Tuple[str, str]] = [
+    ("what year was the python programming language first released", "1991"),
+    ("who wrote the book the hobbit", "Tolkien"),
+    ("what is the chemical symbol for tungsten", "W"),
+    ("how many bones are in the adult human body", "206"),
+    ("what is the capital city of australia", "Canberra"),
+    ("what does HTTP status code 418 mean", "teapot"),
+    ("who created the linux kernel", "Torvalds"),
+    ("what is the boiling point of water in fahrenheit at sea level", "212"),
+    ("what year did the berlin wall fall", "1989"),
+    ("what is the largest planet in the solar system", "Jupiter"),
+]
+
+P_DOMAIN = re.compile(r"https?://([^/\s)]+)", re.I)
+
+
+@typechecked
+def p_domains(text: str) -> int:
+    """Distinct registrable-ish domains in a result blob; a proxy for source diversity."""
+    hosts = {h.lower().lstrip("www.") for h in P_DOMAIN.findall(text or "")}
+    return len(hosts)
+
+
+@typechecked
+async def p_one(query: str, expected: str) -> Dict:
+    from backend.apps.web.web import SearchBody, search
+
+    t0 = time.perf_counter()
+    try:
+        out = await search(SearchBody(query=query, num_results=5))
+    except Exception as exc:
+        return {"query": query, "ok": False, "error": f"{type(exc).__name__}: {exc}"[:90]}
+    dt = time.perf_counter() - t0
+    text = str((out or {}).get("results") or "")
+    return {
+        "query": query,
+        "ok": True,
+        "seconds": round(dt, 3),
+        "backend": (out or {}).get("backend") or "?",
+        "answered": expected.lower() in text.lower(),
+        "domains": p_domains(text),
+        "chars": len(text),
+    }
+
+
+@typechecked
+async def run(limit: Optional[int] = None) -> Dict:
+    rows: List[Dict] = []
+    for query, expected in CORPUS[: limit or len(CORPUS)]:
+        row = await p_one(query, expected)
+        rows.append(row)
+        mark = "ok " if row.get("answered") else ("ERR" if not row.get("ok") else "MISS")
+        print(f"  {mark} {row.get('seconds', '-'):>6}s  {row.get('backend', '-'):<10} {query[:44]}")
+    good = [r for r in rows if r.get("ok")]
+    answered = [r for r in good if r.get("answered")]
+    lat = [r["seconds"] for r in good]
+    return {
+        "n": len(rows),
+        "usable": len(good),
+        "answered": len(answered),
+        "answer_rate": round(len(answered) / len(rows), 3) if rows else 0.0,
+        "median_seconds": round(statistics.median(lat), 3) if lat else None,
+        "median_domains": round(statistics.median([r["domains"] for r in good]), 1) if good else None,
+        "engines": sorted({r.get("backend", "?") for r in good}),
+    }
+
+
+if __name__ == "__main__":
+    summary = asyncio.run(run())
+    print()
+    print(json.dumps(summary, indent=2))
+    # A benchmark that cannot fail is not a benchmark. These are the thresholds; N is stated so a
+    # shrunken corpus cannot pass by measuring less.
+    if summary["n"] < 10:
+        print("FAIL: corpus shrank below 10 queries; the rates would not mean anything")
+        sys.exit(1)
+    if summary["answer_rate"] < 0.7:
+        print(f"FAIL: answer rate {summary['answer_rate']} below the 0.70 threshold")
+        sys.exit(1)
+    print(f"PASS: {summary['answered']}/{summary['n']} answered, median {summary['median_seconds']}s")