diff --git a/memory/README.md b/memory/README.md
index 1bdf1dd..5dd9c16 100644
--- a/memory/README.md
+++ b/memory/README.md
@@ -8,3 +8,4 @@ and link them below.
 - [ci-and-verification.md](ci-and-verification.md) — how the site builds/deploys and how to verify changes (incl. the rtk/prettier gotcha)
 - [csp-policy.md](csp-policy.md) — the meta-tag CSP: what it allows and why, and what can't be enforced this way
 - [seo-and-metadata.md](seo-and-metadata.md) — SEO/metadata decisions (German-only, no JSON-LD, canonical/feed)
+- [finding-unused-assets.md](finding-unused-assets.md) — how to check if an image/PDF/CSS class is really unused (and the srcset / URL-encoding traps)
diff --git a/memory/finding-unused-assets.md b/memory/finding-unused-assets.md
new file mode 100644
index 0000000..15f9c20
--- /dev/null
+++ b/memory/finding-unused-assets.md
@@ -0,0 +1,25 @@
+# Finding unused assets (images, PDFs, CSS classes)
+
+This is a Webflow export, so `assets/` carries a lot of dead scaffolding. When checking whether an asset is really unused, grep the **built `_site`** (the
+ground truth of what ships), not just the source — and scan HTML **plus CSS plus `site.webmanifest`**, because references live in all three.
+
+```bash
+bundle exec jekyll build >/dev/null
+grep -rhoE '[A-Za-z0-9._%ÖÄÜ+-]+\.(png|jpg|jpeg|svg|webp|gif)' _site \
+  --include='*.html' --include='*.css' --include='*.webmanifest' | sort -u
+```
+
+Two false-positive traps that made me nearly flag _used_ files as orphans:
+
+- **`srcset` responsive variants.** Webflow images come as `foo.png` plus `foo-p-500.png` / `-p-800` / `-p-1080`. Pages reference the variants via `srcset`,
+  often NOT the full-size original. (Real case: the index hero uses only the `InOEG_IRIS-One-Pager_Hero-p-*` variants; the 2 MB full-size
+  `InOEG_IRIS-One-Pager_Hero.png` is never referenced and never ships. So the earlier "optimise the 2 MB hero" idea was moot — PurgeCSS/the build already never
+  emit it.)
+- **URL-encoded filenames.** PDFs/images with spaces are linked with `%20` (e.g. `230829_inoeg_Halbzeit Ampel.pdf` →
+  `/documents/230829_inoeg_Halbzeit%20Ampel.pdf`). A plain `grep -F "Halbzeit Ampel.pdf"` misses the link. Cross-check by counting: `<pdf links in page>` vs
+  `<pdf files on disk>` — all 13 documents/ PDFs are in fact linked in `veroeffentlichungen.md`.
+
+Also note: dead Webflow CSS (unused `w-icon-*`, `w-lightbox-*` rules and their embedded `data:` fonts/images) is stripped from production by PurgeCSS, so it
+never reaches users even when left in the source `assets/css/`.
+
+Decision on file (2026-07): the maintainer chose NOT to delete orphaned image files despite ~4 MB being unused. Don't propose asset deletion again unless asked.