diff --git a/CLAUDE.md b/CLAUDE.md
index 0175e09..873bcc5 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -143,9 +143,15 @@ face value:
   unique, which `SubscriberQuery.spec.js` asserts by entry count so a future collision fails loudly
   instead of dropping a column.
 - **The server chooses the column order**, not the caller. Parse by header name, never by position.
-- **Two columns cannot be exported and are dropped in silence:** `tag_ids` and `group_membership`.
-  Asking for all 48 returns 46 with no error, so the tool diffs what it asked for against the header
-  and reports `missing_columns`. This is the same silent-drop hazard as `columnView`.
+- **One column cannot be exported and is dropped in silence:** `tag_ids`. Asking for all 48 returns
+  **47** with no error, so the tool diffs what it asked for against the header and reports
+  `missing_columns`. This is the same silent-drop hazard as `columnView`. This file said *two*
+  columns until 2026-09-03 and named `group_membership` as the second; measured that day, all 48
+  came back as 47 and `group_membership` carried the value `None`. Whether Substack changed or the
+  original reading was wrong cannot be told apart after the fact — which is why the tool diffs the
+  header instead of carrying a list of undeliverable columns, and why the count above is the part
+  to distrust first. The suite's CSV fixture omitted that header, so the error had a test agreeing
+  with it: the same shape as the pending-poll mock above.
 - **The download url is relative to the publication host and cookie-authenticated** (403 without it,
   not pre-signed) and answers **CSV, not JSON**. `handleResponse` unconditionally `JSON.parse`s, so
   `readBody` was split out of it and `requestUrl({parse: 'text'})` is the raw path. `requestUrl` also
diff --git a/src/server.js b/src/server.js
index 0694f47..a7bd970 100644
--- a/src/server.js
+++ b/src/server.js
@@ -93,8 +93,8 @@ export const tools = {
       "seen, post views, unique posts seen, comments, shares, links clicked, days active and " +
       "activity rating. Takes the same filters as list_subscribers and covers the whole matching " +
       "set — there is no paging. Substack generates the file asynchronously, so this waits for it " +
-      "and returns the parsed records. Two columns cannot be exported and are reported in " +
-      "`missing_columns` rather than failing: tag_ids and group_membership.",
+      "and returns the parsed records. One column cannot be exported and is reported in " +
+      "`missing_columns` rather than failing: tag_ids.",
     schema: exportSubscribersSchema,
     handler: exportSubscribersHandler,
   },
diff --git a/src/tools/export_subscribers.js b/src/tools/export_subscribers.js
index e189961..98f4e36 100644
--- a/src/tools/export_subscribers.js
+++ b/src/tools/export_subscribers.js
@@ -165,8 +165,11 @@ export const exportSubscribersHandler = async (args, {sleep = sleepSeconds} = {}
   const csv = await substack_api.downloadExport(url);
   const {columns: returned, unmapped, subscribers} = recordsFromCsv(csv);
 
-  // Unsupported columns are dropped by the API with no error at all — `group_membership` and
-  // `tag_ids` never come back. Reporting the difference is the only way the caller learns it.
+  // An unsupported column is dropped by the API with no error at all — `tag_ids` never comes back.
+  // Reporting the difference is the only way the caller learns it. The check stays a diff rather
+  // than a hardcoded list of one: which columns the export refuses is not ours to know statically.
+  // `group_membership` was recorded here as undeliverable and measurably is not — whether Substack
+  // changed or the original reading was wrong cannot be told apart now, which is the argument.
   const missing_columns = columns.filter((column) => !returned.includes(column));
 
   if (missing_columns.length > 0) {
diff --git a/src/tools/export_subscribers.spec.js b/src/tools/export_subscribers.spec.js
index 3bee63b..71a5f5c 100644
--- a/src/tools/export_subscribers.spec.js
+++ b/src/tools/export_subscribers.spec.js
@@ -194,16 +194,18 @@ describe('exportSubscribersHandler — parsing the CSV back', () => {
     assert.deepEqual(result.columns, [
       'user_email_address', 'user_name', 'subscription_created_at',
       'num_email_opens_last_30d', 'num_web_post_views', 'total_revenue_generated',
-      'activity_rating', 'country',
+      'activity_rating', 'country', 'group_membership',
     ]);
   });
 
-  // Verified against the live API: asking for all 48 returns 46. `group_membership` and `tag_ids`
-  // are dropped with no error at all, so a caller told only "success" would believe it had them.
+  // Measured against the live API 2026-09-03: asking for all 48 returns 47, and the only column
+  // that never comes back is `tag_ids`. It is dropped with no error at all, so a caller told only
+  // "success" would believe it had the tags. `group_membership` exports fine — this test and the
+  // CSV fixture both used to claim otherwise.
   test('names the requested columns that never came back', async () => {
-    const result = await run({columns: ['user_email_address', 'user_name', 'tag_ids', 'group_membership']});
+    const result = await run({columns: ['user_email_address', 'tag_ids', 'group_membership']});
 
-    assert.deepEqual(result.missing_columns, ['tag_ids', 'group_membership']);
+    assert.deepEqual(result.missing_columns, ['tag_ids']);
   });
 
   test('missing_columns is empty when everything asked for arrived', async () => {
@@ -389,7 +391,8 @@ describe('exportSubscribersHandler — logging', () => {
 
     const done = find(lines, 'export_subscribers.done');
     assert.equal(done.count, 2);
-    assert.equal(done.columns, 8);
+    // The nine headers EXPORT_CSV carries, not the number requested.
+    assert.equal(done.columns, 9);
   });
 
   test('records each poll that found the export unfinished', async () => {
diff --git a/test/helpers/msw-server.js b/test/helpers/msw-server.js
index 661f467..c75e6cf 100644
--- a/test/helpers/msw-server.js
+++ b/test/helpers/msw-server.js
@@ -42,11 +42,16 @@ export const EXPORT_FILE_URL = `${TEST_ENV.SUBSTACK_PUBLICATION_URL}${EXPORT_FIL
  * A CSV shaped exactly like a real export: the header carries human LABELS rather than column keys,
  * the server's own column order (not the requested one), a quoted currency value instead of a
  * number, and a name containing a comma — the case a `split(',')` gets wrong.
+ *
+ * `Group membership` is in the header because the live API returns it: measured 2026-09-03, all 48
+ * columns requested came back as 47 with only `tag_ids` missing, and the value is the word "None"
+ * rather than an empty cell. This fixture omitted the header, which is how the repo came to claim
+ * the column was undeliverable and had a passing test to prove it.
  */
 export const EXPORT_CSV = [
-  'Email,Name,Start date,Emails opened (30d),Post views,Revenue,Activity,Country',
-  'one@example.com,One,2026-07-29T22:07:50.299Z,2,1,"€0.00",5,BR',
-  'two@example.com,"Two, Junior",2026-06-01T10:00:00.000Z,0,7,"€50.00",3,IT',
+  'Email,Name,Start date,Emails opened (30d),Post views,Revenue,Activity,Country,Group membership',
+  'one@example.com,One,2026-07-29T22:07:50.299Z,2,1,"€0.00",5,BR,None',
+  'two@example.com,"Two, Junior",2026-06-01T10:00:00.000Z,0,7,"€50.00",3,IT,None',
 ].join('\n');
 
 export const DRAFT_RESPONSE = {