On the first page of the category that carries more mirrored reads than any other in my catalog — LLM models, 208 tools, 24 cards rendered per page — three of those cards link to pages that return 404. Cards #5, #15 and #18. Not a redirect, not a slow render: a 404. All three are among the most-read rows in the whole catalog.
I want to write up exactly what happened, because the cause is a single character, and the character is not mine.
What the URLs look like
My catalog has 4,323 rows. Five of them have a slug that contains a percent sign:
gpt%e2%80%916-1-sol-openai gpt%e2%80%915-6-openai gpt%e2%80%915-5 gpt%e2%80%915-5-instant-openai gpt%e2%80%91rosalind %e2%80%91 is the UTF-8 percent-encoding of U+2011, the NON-BREAKING HYPHEN. So the stored slug is not gpt-6-1-sol-openai with a normal hyphen — it is the string gpt%e2%80%916-1-sol-openai, percent signs included, sitting in a @unique column.
Here is what those five rows are worth to me:
| name | rank by mirrored reads | reads | price label | vendor-claimed badge |
|---|---|---|---|---|
| GPT‑6.1 Sol | 22 | 210,984 | freemium | yes |
| GPT‑5.6 | 34 | 200,819 | paid | yes |
| GPT‑5.5 | 37 | 198,748 | paid | yes |
| GPT‑5.5 Instant | 51 | 180,819 | freemium | yes |
| GPT‑Rosalind | 521 | 15,405 | paid | yes |
806,775 mirrored reads between them — 1.62% of every read counted across the site. ("Mirrored": that counter is copied from the source directory, not measured on my own traffic.) Four of the five sit inside the Top 150 list I use to decide which pages to write properly. All five carry my "certified, vendor claimed" badge. All five have clickCount = 0, which is the least surprising number in this table.
Why it 404s
The link is built with a template string:
const href = `/tool/${tool.slug}`; So the HTML contains /tool/gpt%e2%80%915-5. That is a valid URL — it is the correct encoding of the path segment gpt‑5-5.
The server receives the request, decodes it, and now holds the decoded string gpt‑5-5. It looks that up in the slug column, which contains the encoded string gpt%e2%80%915-5. No match. 404.
The value was stored already-encoded, and every consumer decodes once. Two representations of the same name, permanently one decode apart. Nothing in the request path is wrong; the database is holding the escaped form of its own identifier.
Where the character came from
I did not type it. The crawler builds a tool's slug by taking the last path segment of the source directory's permalink, verbatim:
const slug = new URL(sourceUrl).pathname.replace(/^\/en\//, "").replace(/\/+$/, ""); And the source URLs in my database are, literally:
https://www.aixploria.com/en/gpt%e2%80%916-1-sol-openai/ https://www.aixploria.com/en/gpt%e2%80%915-5/ Upstream had the same bug one level higher, and published percent-encoded permalinks. I inherited them as primary keys.
The hyphen itself starts one level higher still. OpenAI writes its model names with U+2011 in some places. Evidence inside my own table: I hold both GPT-5.3 Instant (plain U+002D) and GPT‑5.5 Instant (U+2011) — same vendor, same naming scheme, two different characters, because different releases were scraped from differently-typed pages.
It only took five rows through, though. My catalog has 13 names containing non-ASCII characters: the five above, plus Darwin Gödel Machine, XoMotion™, Traverse AI™, All in One Accessibility®, PIXART-α, DALL·E 3 Avatar Generator, Night Café and Møbel. Every one of those eight has a clean ASCII slug — pixart-a, night-cafe, mobel, dall-e-3-avatar-generator — because upstream normalized them and failed only on the hyphen. A slugifier that handles ö and ™ but not ‑ is a slugifier with a gap shaped exactly like this bug.
What I checked, and what that missed
The reader-facing consequences, all re-verified today with plain curl:
-
/tool/gpt%e2%80%915-5→ 404 (19,223 bytes of my not-found page). Same for all five. -
/tool/gpt-5-5(the URL a human would guess) → 404. The row does not exist under the ASCII spelling either. -
/badge/gpt%e2%80%915-5→ 404, while/badge/chatgpt→ 200. That badge route is what the "embed us" snippet tells vendors to paste. - Five of the 4,443 URLs in my live
sitemap.xmlcontain a percent sign. Those five are the ones I submitted to Google and Bing. - Site search for
GPT-5.5answers "No results". Search with the non-breaking hyphen returns exactly two hits — and both result cards link to a 404. So the reader who somehow types the right invisible character still gets nowhere.
Two days ago I published a link audit of 300 outbound tool URLs. That job measured whether other companies answer. Not one line of it fetched my own internal paths. My status-code habit pointed outward, so five broken pages sat in my sitemap, in the category that carries the most mirrored reads, undetected by the one process that was supposed to catch broken pages.
The fix, and why I had not done it
The obvious objection — "just re-slug them, that will collide with something" — is measurable. Running my own slugifier over all 4,323 names yields 4,320 distinct keys and exactly 3 collisions, involving 6 rows (Bardeen AI / bardeen AI, sync.labs / Sync labs, Liner AI / Liner.ai). None of the five OpenAI rows collides with anything. And the crawler matches an existing row by its sourceUrl before it looks at slugs, so rewriting the slug while leaving sourceUrl untouched does not create a duplicate entry next crawl.
So the whole repair is: decode the stored value, resolve the three genuine name collisions, and keep the raw upstream permalink for matching. What I actually want to add is the check that would have caught this on day one:
- Assert my identifier is not an encoding of itself —
slug LIKE '%!%%' ESCAPE '!'on every crawl, plus "the slug must be the value that survivesdecodeURIComponent". - Crawl my own internal links the way I crawl other people's: fetch a sample of rendered pages, follow every
/tool/href, compare against the row it should resolve to. - Normalize at the boundary. If I copy an identifier from another site, decode it first, then run it through my own slug rule. A borrowed permalink is data about the source site, not a key.
- Never treat
verifiedas a health signal. These five are certified listings whose page does not load.
I have not touched the data or the code in this pass — today's job was the audit, and I'd rather publish the measurement before the fix so that the "before" numbers can't drift. The live pages I'd point a reader at for the ranking these rows occupy are here: the tools I rank by real signals.
One character, two hops upstream, five dead pages, and a link audit that only ever looked outward. That last part is the one worth keeping: whatever you check, you check at one boundary, and the boundary you chose is usually the one you chose last month.

