diff --git a/DEPLOY.md b/DEPLOY.md
index 9a84ba1..7031a34 100644
--- a/DEPLOY.md
+++ b/DEPLOY.md
@@ -118,8 +118,10 @@ shared secret is needed.
    https://floatchat.vercel.app,http://localhost:3000
    ```
 
-5. Open the site and ask a question. The run log should fill in step by step and the answer should stream in.
-   If you later add a custom domain on Vercel, add it to `ALLOWED_ORIGINS` too.
+5. In **Vercel → your project → Settings → Environment Variables**, add `NEXT_PUBLIC_SITE_URL` = that URL and
+   redeploy, so shared links show your site's preview image (the default is the original demo's address).
+6. Open the site and ask a question. The run log should fill in step by step and the answer should stream in.
+   If you later add a custom domain on Vercel, add it to `ALLOWED_ORIGINS` and `NEXT_PUBLIC_SITE_URL` too.
 
 ## 6. Streamlit Community Cloud: the old interim UI (optional)
 
diff --git a/docs/screenshots/web-chat-sql.png b/docs/screenshots/web-chat-sql.png
index 4615cc1..2f74c62 100644
Binary files a/docs/screenshots/web-chat-sql.png and b/docs/screenshots/web-chat-sql.png differ
diff --git a/eval/latest.json b/eval/latest.json
index 0ce7896..2f5ff82 100644
--- a/eval/latest.json
+++ b/eval/latest.json
@@ -1,15 +1,15 @@
 {
   "api": "https://floatchat-ai-svjw.onrender.com",
-  "ran_at": "2026-11-08 18:16 UTC",
+  "ran_at": "2026-11-09 05:16 UTC",
   "summary": {
     "cases": 30,
-    "passed": 29,
-    "routing_accuracy": 0.967,
+    "passed": 30,
+    "routing_accuracy": 1.0,
     "filter_accuracy": 1.0,
     "answer_pass_rate": 1.0,
-    "median_latency_ms": 2595,
-    "llm_routed": 28,
-    "llm_wrote_answer": 30,
+    "median_latency_ms": 1929,
+    "llm_routed": 30,
+    "llm_wrote_answer": 28,
     "failed_requests": 0
   },
   "results": [
@@ -31,7 +31,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2701,
+      "latency_ms": 1249,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -61,7 +61,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1498,
+      "latency_ms": 1783,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -95,7 +95,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1712,
+      "latency_ms": 1839,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -121,7 +121,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1957,
+      "latency_ms": 1772,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -147,7 +147,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1836,
+      "latency_ms": 1976,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -173,7 +173,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1807,
+      "latency_ms": 1572,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -203,7 +203,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1931,
+      "latency_ms": 1764,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -229,7 +229,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1818,
+      "latency_ms": 1782,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -259,7 +259,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2003,
+      "latency_ms": 1883,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -285,7 +285,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 3129,
+      "latency_ms": 1999,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -315,7 +315,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2463,
+      "latency_ms": 2010,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -343,7 +343,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1996,
+      "latency_ms": 1492,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -374,8 +374,8 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2165,
-      "router": "keywords",
+      "latency_ms": 1477,
+      "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
       "filters": true,
@@ -402,7 +402,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 1672,
+      "latency_ms": 1625,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -432,7 +432,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 181,
+      "latency_ms": 2420,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -464,7 +464,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2882,
+      "latency_ms": 2727,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -491,7 +491,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2162,
+      "latency_ms": 1777,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -523,7 +523,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 7575,
+      "latency_ms": 7085,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -553,7 +553,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 3701,
+      "latency_ms": 16268,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -581,7 +581,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 2894,
+      "latency_ms": 10319,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -611,7 +611,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 10737,
+      "latency_ms": 12675,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -643,9 +643,9 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 10217,
-      "router": "keywords",
-      "llm_wrote_answer": true,
+      "latency_ms": 5667,
+      "router": "llm",
+      "llm_wrote_answer": false,
       "routing": true,
       "filters": true,
       "answer": true,
@@ -671,7 +671,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 4217,
+      "latency_ms": 2374,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -701,7 +701,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 11665,
+      "latency_ms": 8044,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -725,9 +725,9 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 7732,
+      "latency_ms": 9042,
       "router": "llm",
-      "llm_wrote_answer": true,
+      "llm_wrote_answer": false,
       "routing": true,
       "filters": true,
       "answer": true,
@@ -747,23 +747,23 @@
         }
       },
       "got": {
-        "path": "analytical",
-        "intent": "custom",
+        "path": "semantic",
+        "intent": null,
         "filters": {
           "region": "Bay of Bengal",
           "month_num": 9,
           "year": 2026
         },
-        "answered_by": "analytical"
+        "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 4907,
+      "latency_ms": 193,
       "router": "llm",
       "llm_wrote_answer": true,
-      "routing": false,
+      "routing": true,
       "filters": true,
       "answer": true,
-      "note": "routed to analytical/custom"
+      "note": null
     },
     {
       "id": "follow-region",
@@ -787,7 +787,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 3397,
+      "latency_ms": 1843,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -819,7 +819,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 5795,
+      "latency_ms": 2524,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -849,7 +849,7 @@
         "answered_by": "analytical"
       },
       "status": 200,
-      "latency_ms": 2489,
+      "latency_ms": 1783,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
@@ -879,7 +879,7 @@
         "answered_by": "semantic"
       },
       "status": 200,
-      "latency_ms": 13054,
+      "latency_ms": 3235,
       "router": "llm",
       "llm_wrote_answer": true,
       "routing": true,
diff --git a/eval/latest.md b/eval/latest.md
index 8ae8df8..5d745d8 100644
--- a/eval/latest.md
+++ b/eval/latest.md
@@ -1,15 +1,15 @@
-## FloatChat evaluation — 29/30 questions fully correct
+## FloatChat evaluation — 30/30 questions fully correct
 
-Run against `https://floatchat-ai-svjw.onrender.com` on 2026-11-08 18:16 UTC.
+Run against `https://floatchat-ai-svjw.onrender.com` on 2026-11-09 05:16 UTC.
 
 | Metric | Score |
 |---|---|
-| Routing accuracy (SQL vs search, and which analysis) | 97% |
+| Routing accuracy (SQL vs search, and which analysis) | 100% |
 | Filter accuracy (region, month, year, sensor) | 100% |
 | Answer checks passed | 100% |
-| Median latency | 2595 ms |
-| Understood by the LLM router | 28/30 |
-| Answers written by the LLM | 30/30 |
+| Median latency | 1929 ms |
+| Understood by the LLM router | 30/30 |
+| Answers written by the LLM | 28/30 |
 
 | Case | Question | Routing | Filters | Answer | Notes |
 |---|---|---|---|---|---|
@@ -38,7 +38,7 @@ Run against `https://floatchat-ai-svjw.onrender.com` on 2026-11-08 18:16 UTC.
 | `no-data-red-sea` | Describe recent profiles from the Red Sea | ✅ | ✅ | ✅ |  |
 | `find-nitrate-arabian` | Find nitrate measurements in the Arabian Sea | ✅ | ✅ | ✅ |  |
 | `find-thermocline` | Show me profiles with a strong thermocline | ✅ | ✅ | ✅ |  |
-| `find-bengal-september` | Profiles in the Bay of Bengal in September 2026 | ❌ | ✅ | ✅ | routed to analytical/custom |
+| `find-bengal-september` | Profiles in the Bay of Bengal in September 2026 | ✅ | ✅ | ✅ |  |
 | `follow-region` | now only the Bay of Bengal *(follow-up to depth-avg)* | ✅ | ✅ | ✅ |  |
 | `follow-month` | and just in March *(follow-up to depth-avg-arabian)* | ✅ | ✅ | ✅ |  |
 | `follow-monthly-arabian` | now for the Arabian Sea *(follow-up to monthly)* | ✅ | ✅ | ✅ |  |
diff --git a/readme.md b/readme.md
index daf63f0..6025560 100644
--- a/readme.md
+++ b/readme.md
@@ -3,10 +3,12 @@
 **Ask questions about a year of real ocean data in plain English. FloatChat writes the SQL or searches the
 profiles, shows every step it took, and draws the charts.**
 
-![FloatChat landing page](docs/screenshots/web-landing.png)
+### ▶ Try it live: [float-chat-ai-teal.vercel.app](https://float-chat-ai-teal.vercel.app)
+
+[![FloatChat landing page](docs/screenshots/web-landing.png)](https://float-chat-ai-teal.vercel.app)
 
 **Live API:** [floatchat-ai-svjw.onrender.com/docs](https://floatchat-ai-svjw.onrender.com/docs) ·
-**Benchmark:** 29/30 questions fully correct ([details](eval/latest.md)) · **Deploy your own:** [DEPLOY.md](DEPLOY.md)
+**Benchmark:** 30/30 questions fully correct ([details](eval/latest.md)) · **Deploy your own:** [DEPLOY.md](DEPLOY.md)
 
 **Contents:** [What FloatChat is](#-what-floatchat-is) · [How it's built](#-how-its-built) ·
 [How well it works](#-how-well-it-works) · [Run it on your machine](#-run-it-on-your-machine) ·
@@ -68,7 +70,7 @@ Each float has its own page: trajectory, latest profile, a depth-time section an
 All data are **real Argo observations** from the **Indian Ocean**, downloaded from the
 [Ifremer ERDDAP](https://erddap.ifremer.fr/erddap/) server, keeping only values that passed Argo's quality control:
 
-| | Live demo | Bundled snapshot (local setup) |
+| | [Live demo](https://float-chat-ai-teal.vercel.app) | Bundled snapshot (local setup) |
 |---|---|---|
 | Period | rolling last 12 months, refreshed nightly | January to March 2024 |
 | Floats | about 660 (107 measuring oxygen) | 437 (52 measuring oxygen) |
@@ -229,14 +231,15 @@ every Monday (`eval.yml`) and shows the table in the job summary. Without an LLM
 template questions and simple follow-ups; the free-phrasing and custom-SQL questions are where the LLM earns its keep.
 
 <!-- eval-score -->
-**Latest score (live API, 8 Nov 2026): 29/30 questions fully correct.**
+**Latest score (live API, 9 Nov 2026): 30/30 questions fully correct.**
 
 | Routing | Filters | Answer checks | Median latency | Answers written by the LLM |
 |---|---|---|---|---|
-| 97% | 100% | 100% | 2.6 s | 30/30 |
+| 100% | 100% | 100% | 1.9 s | 28/30 |
 
-The one miss: *"Profiles in the Bay of Bengal in September 2026"* went to custom SQL instead of profile search. The
-full per-question table is in [`eval/latest.md`](eval/latest.md).
+The benchmark asks its questions back to back, so Groq's free tier sometimes hits its tokens-per-minute limit. When
+that happens FloatChat still answers with a plain summary of the data, and the run log marks it *Answer built from
+the data alone*. The full per-question table is in [`eval/latest.md`](eval/latest.md).
 <!-- /eval-score -->
 
 ---
diff --git a/web/.env.example b/web/.env.example
index d24b712..a71c04d 100644
--- a/web/.env.example
+++ b/web/.env.example
@@ -1,2 +1,5 @@
 # The FloatChat API the site talks to (the browser calls it directly)
 NEXT_PUBLIC_API_URL=https://floatchat-ai-svjw.onrender.com
+
+# This site's public address, for absolute links such as the social preview image
+NEXT_PUBLIC_SITE_URL=https://float-chat-ai-teal.vercel.app
diff --git a/web/app/layout.tsx b/web/app/layout.tsx
index 79b0c9f..6c41565 100644
--- a/web/app/layout.tsx
+++ b/web/app/layout.tsx
@@ -2,10 +2,11 @@ import type { Metadata, Viewport } from "next";
 import { GeistMono } from "geist/font/mono";
 import { GeistSans } from "geist/font/sans";
 import { SiteHeader } from "@/components/site-header";
+import { SITE_URL } from "@/lib/config";
 import "./globals.css";
 
 export const metadata: Metadata = {
-  metadataBase: new URL("https://floatchat.vercel.app"),
+  metadataBase: new URL(SITE_URL),
   title: { default: "FloatChat: ask the Indian Ocean", template: "%s · FloatChat" },
   description:
     "Ask questions about a year of real Argo ocean-float data in plain English. FloatChat writes the SQL or searches " +
diff --git a/web/lib/config.ts b/web/lib/config.ts
index e3fc648..ea79937 100644
--- a/web/lib/config.ts
+++ b/web/lib/config.ts
@@ -7,6 +7,9 @@ export const API_URL = (process.env.NEXT_PUBLIC_API_URL || "https://floatchat-ai
  */
 export const SERVER_API_URL = (process.env.API_URL_INTERNAL || API_URL).replace(/\/$/, "");
 
+/** The public address of this site, used for absolute links such as the social preview image. */
+export const SITE_URL = (process.env.NEXT_PUBLIC_SITE_URL || "https://float-chat-ai-teal.vercel.app").replace(/\/$/, "");
+
 export const REPO_URL = "https://github.com/CodexManvik/FloatChat-AI";
 export const BENCHMARK_URL = "https://raw.githubusercontent.com/CodexManvik/FloatChat-AI/main/eval/latest.json";
 export const BENCHMARK_TABLE_URL = `${REPO_URL}/blob/main/eval/latest.md`;