diff --git a/sql/2025/page-weight/bytes_per_type.sql b/sql/2025/page-weight/bytes_per_type.sql
index d06962d28cf..668da00672d 100644
--- a/sql/2025/page-weight/bytes_per_type.sql
+++ b/sql/2025/page-weight/bytes_per_type.sql
@@ -1,25 +1,57 @@
+-- Because summary bytesHtmlDoc is a little wrong and always zero in the pages table (see: https://github.com/HTTPArchive/wptagent/issues/47)
+-- we need to recalculate it by joining back to the requests table.
+-- To optimize performance, we first create a CTE to pre-filter and extract the necessary data
+-- from the requests table, then join that CTE in the main query.
+-- when fixed, we can revert to just using summary.bytesHtmlDoc directly and remove the join and CTE.
+WITH HtmlDocRequests AS (
+  SELECT
+    page,
+    client,
+    date,
+    -- Extract and cast the response size once inside the CTE
+    CAST(JSON_VALUE(summary.respSize) AS INT64) AS respSize_bytes
+  FROM
+    `httparchive.crawl.requests`
+  WHERE
+    -- Pre-filter the requests table for efficiency.
+    -- This MUST match the date in the main query's WHERE clause.
+    date = '2025-07-01' AND
+    type = 'html' AND
+    is_main_document
+)
+
+-- Main query now joins the CTE
 SELECT
   percentile,
-  client,
-  is_root_page,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesTotal) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS total_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesHtml) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS html_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesJS) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS js_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesCss) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS css_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesImg) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS img_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesOther) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS other_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesHtmlDoc) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS html_doc_kbytes,
-  APPROX_QUANTILES(CAST(JSON_VALUE(summary.bytesFont) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS font_kbytes
+  p.client,
+  p.is_root_page,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesTotal) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS total_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesHtml) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS html_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesJS) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS js_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesCss) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS css_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesImg) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS img_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesOther) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS other_kbytes,
+  -- Use the pre-calculated column from the CTE
+  APPROX_QUANTILES(r.respSize_bytes / 1024, 1000)[OFFSET(percentile * 10)] AS html_doc_kbytes,
+  APPROX_QUANTILES(CAST(JSON_VALUE(p.summary.bytesFont) AS INT64) / 1024, 1000)[OFFSET(percentile * 10)] AS font_kbytes
 FROM
-  `httparchive.crawl.pages`,
+  `httparchive.crawl.pages` p
+LEFT JOIN
+  -- Join the new CTE on the matching keys
+  HtmlDocRequests r
+ON
+  p.page = r.page AND
+  p.client = r.client AND
+  p.date = r.date,
   UNNEST([10, 25, 50, 75, 90, 100]) AS percentile
 WHERE
-  date = '2025-07-01'
+  -- This filter on the main 'pages' table is still required
+  p.date = '2025-07-01'
 GROUP BY
   percentile,
-  client,
-  is_root_page
+  p.client,
+  p.is_root_page
 ORDER BY
-  client,
-  is_root_page,
+  p.client,
+  p.is_root_page,
   percentile