diff --git a/benchmarks/page_template.md b/benchmarks/page_template.md
index 4964e0e..edb7ca8 100644
--- a/benchmarks/page_template.md
+++ b/benchmarks/page_template.md
@@ -1,7 +1,7 @@
 # Benchmarks
 
 adata-cli measured against anndata, and against scanpy wherever scanpy has a real
-equivalent. The [`Benchmark`](../.github/workflows/benchmark.yml) workflow rewrites
+equivalent. The [`Benchmark`](https://github.com/cellgeni/adata-cli/blob/main/.github/workflows/benchmark.yml) workflow rewrites
 this page on every tag and keeps each run's raw numbers in `docs/benchmarks/`.
 
 **Peak RSS is the headline, not wall time.** This tool exists so that memory is set
diff --git a/benchmarks/report.py b/benchmarks/report.py
index 9596dae..c49652f 100644
--- a/benchmarks/report.py
+++ b/benchmarks/report.py
@@ -20,6 +20,10 @@
 
 from benchmarks.cases import by_name
 
+#: Links out of `docs/` cannot be relative: GitHub Pages serves that
+#: directory as the site root, so `../` and bare directory paths 404.
+REPO = "https://github.com/cellgeni/adata-cli"
+
 
 def _bytes(n: int) -> str:
     if not n:
@@ -186,7 +190,7 @@ def history_table(docs: Path) -> str:
         "### History",
         "",
         "`concat-inner` on adata-cli, run by run. Full results for each are "
-        "in [`docs/benchmarks/`](benchmarks/).",
+        f"in [`docs/benchmarks/`]({REPO}/tree/main/docs/benchmarks).",
         "",
         "| Run | Ref | Tier | Wall time | Peak RSS | Raw |",
         "|---|---|---|---|---|---|",
diff --git a/docs/BENCHMARKS.md b/docs/BENCHMARKS.md
index c327df0..d4d7086 100644
--- a/docs/BENCHMARKS.md
+++ b/docs/BENCHMARKS.md
@@ -1,7 +1,7 @@
 # Benchmarks
 
 adata-cli measured against anndata, and against scanpy wherever scanpy has a real
-equivalent. The [`Benchmark`](../.github/workflows/benchmark.yml) workflow rewrites
+equivalent. The [`Benchmark`](https://github.com/cellgeni/adata-cli/blob/main/.github/workflows/benchmark.yml) workflow rewrites
 this page on every tag and keeps each run's raw numbers in `docs/benchmarks/`.
 
 **Peak RSS is the headline, not wall time.** This tool exists so that memory is set
@@ -218,7 +218,7 @@ Rows marked `n/a` are operations the baseline does not offer; that is a result,
 
 ### History
 
-`concat-inner` on adata-cli, run by run. Full results for each are in [`docs/benchmarks/`](benchmarks/).
+`concat-inner` on adata-cli, run by run. Full results for each are in [`docs/benchmarks/`](https://github.com/cellgeni/adata-cli/tree/main/docs/benchmarks).
 
 | Run | Ref | Tier | Wall time | Peak RSS | Raw |
 |---|---|---|---|---|---|
diff --git a/tests/test_docs_are_accurate.py b/tests/test_docs_are_accurate.py
index f6f4608..a3455e3 100644
--- a/tests/test_docs_are_accurate.py
+++ b/tests/test_docs_are_accurate.py
@@ -211,3 +211,57 @@ def test_options_claimed_in_prose_exist(option, command, source):
         f"{source}: prose says {command} supports {option}, "
         f"but the command does not accept it"
     )
+
+
+# ---------------------------------------------------------------------------
+# links that have to work on the published site
+#
+# GitHub Pages serves `docs/` as the site root, so a relative link out of it
+# cannot resolve: `../.github/workflows/benchmark.yml` and a bare
+# `benchmarks/` both went live as 404s on the 0.6.0 release page. Checked
+# offline against the filesystem, so it costs nothing and needs no network.
+
+
+def _relative_links(text: str):
+    """(link, target) for every relative markdown link, fragments stripped."""
+    for match in re.finditer(r"\]\(([^)\s]+)\)", text):
+        link = match.group(1)
+        if link.startswith(("http://", "https://", "mailto:", "#")):
+            continue
+        target = link.split("#", 1)[0]
+        if target:
+            yield link, target
+
+
+DOC_PAGES = sorted((REPO / "docs").glob("*.md"))
+
+
+@pytest.mark.parametrize("page", DOC_PAGES, ids=lambda p: p.name)
+def test_every_relative_link_resolves_inside_the_published_site(page):
+    """A relative link must point at something Pages actually serves.
+
+    Jekyll rewrites `TESTING.md` to `/TESTING.html`, so a `.md` target is
+    fine; anything reached with `../`, or a bare directory with no index,
+    is not, and has to be an absolute URL to GitHub instead.
+    """
+    broken = []
+    for link, target in _relative_links(page.read_text()):
+        resolved = (page.parent / target).resolve()
+        try:
+            inside = resolved.is_relative_to((REPO / "docs").resolve())
+        except AttributeError:  # pragma: no cover - Python < 3.9
+            inside = str(resolved).startswith(str((REPO / "docs").resolve()))
+
+        if not inside:
+            broken.append(f"{link} -- leaves docs/, so Pages cannot serve it")
+        elif not resolved.exists():
+            broken.append(f"{link} -- no such file")
+        elif resolved.is_dir() and not (resolved / "index.md").exists():
+            broken.append(f"{link} -- a directory with no index page")
+
+    assert not broken, (
+        f"{page.name} has links that 404 on the published site:\n  "
+        + "\n  ".join(broken)
+        + "\nUse an absolute https://github.com/... URL for anything outside "
+        "docs/."
+    )