Stichting Join Us
Nobody signed up after 200,000+ ad views. I found out why and redesigned the page: from 0 to 20 sign-ups.
Ad views 203,613
Clicks 1,309
Sign-ups 0 → 20
Vacaturescan
A free job-ad checker that marks problems while you type. Two test rounds showed free AI does the same job, so I stopped it. This case is about that decision.
- Built in
- 1 day
- Test rounds
- 2
- Stopped on
- Day 3
- Developers
- 0
Live tool · Wed 15 Oct, 19:56
‘creatives’ is a normal design word
Stopped · Wed 15 Oct, 20:05
/vacaturescan → /afhaakcheck (301)
In short
The case in three steps
Each part keeps its colour further down the page, so you can jump straight to the one you need.
Part 1 · Mon 13 Oct
1 day
to research it, build it and put it live on a hidden page. It only flagged what has proof.
- 11:25 Research
- 12:29 Build
- 14:55 Test
- 17:25 Version 2
Four steps on day 1
Part 2 · Mon to Wed
4 of 6
English design words it marked as mistakes. Testers said they’d just use AI.
- creatives
- ads
- mockups
- identity
- skilled
- fast-paced
Wavy: normal design words · dotted: fair notes
Part 3 · Wed 15 Oct
57 hours
from brief to stop. Every fix made it equal to free AI, not better.
My 3 stop rules: all met
Stopped at 20:05, the same evening
01 · Context
A free tool to bring in employers
I run Talentbezorgd: I find out why people stop applying. The free tool had to bring in employers.
- Step 1 · free Paste a job ad No email, no account.
- Step 2 · free See what to fix Vacaturescan marks the text while you type.
- Step 3 · what I offer Book the Afhaakcheck What the scan can’t see: the route after the click.
0 free Dutch tools like this since 2020. The last one, VacatureVerbeteraar, went paid.
The page as it went live on talentbezorgd.com/vacaturescan: free, no email, hidden from Google and not linked anywhere. Your text never left your device.
02 · Process
57 hours from brief to stop
Tue 14
Wed 15 Oct
Research
Build
Test
Decide
Live,
hidden
11
12
13
14
15
16
17
18
11
12
13
14
15
16
17
18
19
20
Brief 11:25Mon 13 Oct
Hand test · time not loggedStopped 20:05
Times in CEST. Tuesday is collapsed: version 2 stayed live on a hidden page.
- Mon 11:25 Research 6 tracks, 5 ideas ranked
- Mon 12:29 Version 1 12 checks, live at 14:55
- Mon 14:55 Test round 1 4 testers from HR
- Mon 17:25 Version 2 Tap to fix, 20 checks
- Wed 11:18 Test round 2 Clumsy next to free AI
- Wed Hand test Local market: no better
- Wed 20:05 Stopped Link now goes to the Afhaakcheck
Part 1 of 3 · Mon 13 Oct
What did I build in a day?
Short answer
A checker that marks job-ad problems while you type. It only flags what has proof from real hiring, and shows how strong that proof is.
03 · Research
Only four things in a job ad have proof
Before building, I checked which job-ad advice has proof from real hiring. Most of it has none.
Proof from real hiring
4 · these became the checks
- Show the pay
- Fewer optional requirements ≈ +7% applications, Uber test
- Say what is really required
- Say exactly what is flexible
No proven effect
3 · left out
- Readability scores
- Ad length
- “Gendered words” 0 effect in 2,201 real ads
Myths
3 · never repeated
- “Women only apply at 100%” men 52%, women 56%
- “400 ms Doherty threshold” not in the 1982 IBM paper
- “8-second attention span”
Every check shows how strong its proof is
No total score and no “+X% applicants”. The labels from the live tool:
- wet Strength 4 of 4
- veldproef Strength 3 of 4
- labproef Strength 2 of 4
- cijfers Strength 2 of 4
- enquête Strength 1 of 4
- eigen regel Strength 0 of 4
The finding that predicted the end
AI-written job ads are normal now.
Wiles & Horton 2026 · randomised test with 180,000+ employers
- AI-written posts More generic
- Applications Slightly more
- Hires Fewer, later
04 · Design
A spell checker for hiring
Every key updates the marks. It runs in the browser, so the text never leaves your device.
Version 1 · Mon 14:55
The hint sat in a grey box under the editor, and only for the word your cursor was on. Nothing on the page said so. Testers found it late.
Version 2 · Mon 17:25
Vaag over loon veldproef Strength 3 of 4
Zet er een bedrag in.
marktconform salaris
salaris van € 3.550 tot € 5.100 bruto per maand
Dan ziet iemand meteen of het past. Dit is wat de middelste helft van de elektriciens en elektromonteurs verdient (CBS). Pas het aan naar jouw loon.
Tap a mark and a card opens with the fix. On a phone it docks above the keyboard (the Speak pattern). After a fix, the next card opens.
The marks
Six ways to mark a word, plus where one phone screen ends
- competitief salaris Hier haakt iemand af
- flexibel inzetbaar Let op
- € 3.200 tot € 3.800 Dit helpt
- goede secundaire Zegt niets
- moedertal Spelling of taal
- Ben jij die gedreven doorpakker Zin te lang
- Tot hier: 1 telefoonscherm
Labels from the live tool. Red marks: a reader drops out or should watch out. Green: this helps. The blue line is my own estimate of one phone screen, about 80 words, and the tool said so.
Borrowed from (Mobbin)
Patterns people already know
- Grammarly The old word struck out next to the new one, then Accept or Dismiss.
- WRITER Fixes grouped with counts, and one rule under each fix.
- Speak Wrong line, right line, the reason underneath. Fits a phone.
Build facts, all measured
- 0.30 ms per scan of the example ad
- 124.6 KiB the whole core, compressed
- 4,937 job titles with CBS pay data
- 51 of 51 automated browser checks passed
Nobody noticed the 0.3 ms. Being right mattered; being fast didn’t.
Part 2 of 3 · Mon to Wed
What did testers say?
Short answer
Round 1 asked how it works, so I fixed it that evening. Round 2 asked why they’d use it instead of free AI. A list of rules can’t read context.
05 · Testing
Round 1 asked how. Round 2 asked why.
Round 1 · Mon 14:55 · 4 testers from HR
“How does it work?”
Fixed the same evening. I keep going.
Round 2 · Wed 11:18 · the same 4 testers
“Why use it at all?”
No quick fix. I stop.
-
“People were clicking on the dotted lines, expecting it to suggest fixes easily.”
Fixed in v2
A card with the fix opens when you tap a mark
-
“People were intentionally making typos, and the vacaturescan would then ignore text like ‘moedertaal’…”
Fixed in v2
Typo-proof word lists: ‘moedertal’ now counts as ‘moedertaal’
-
“People only later found out that the grey box underneath the scanbox acts like a tooltip, which is bad ux.”
Fixed in v2
The card opens at the word, and docks above the keyboard on a phone
-
“People do like that the vacaturescan automatically fills and guesses the job in the ‘loon’ component.”
Kept in v2
“Vul in” now writes the CBS pay range straight into the ad
Feedback in at 15:14. Version 2 built by 17:25 and live, still hidden, at 17:30.
“The Vacaturescan doesn’t work pleasantly compared to what people are used to these days from autocorrect from Apple or Grammarly… it doesn’t read context well… I also feel like people would just use AI.”
What it was compared to
- Grammarly
- Apple autocorrect
- ChatGPT
Not a bug list any more. The question had moved from “does it work?” to “why use this instead of what I already have?”
06 · Why it failed
Rules see letters, not meaning
Design ads mix Dutch and English. On a sample design ad, the tool marked 6 English words as mistakes. 4 of them are normal design words.
Je moet creatives maken voor ads op Meta en TikTok. Je werkt samen met onze copywriter en de accountmanager. Je maakt mockups in Figma en je levert assets aan in de juiste formats. Ook denk je mee over de brand identity van klanten en geef je feedback op de briefing.
- Je bent skilled in Adobe Creative Cloud (Photoshop, Illustrator, After Effects)
- Je hebt een goed oog voor detail en je bent een echte teamplayer
- Ervaring met motion design is een pré
- Je word blij van deadlines en een fast-paced omgeving
- Een competitief salaris
Underlines: orange = wrong flag · grey = fair note · green = caught correctly
What the tool marked as spelling
6 English words. 4 of them were wrong.
- creatives read as a typo of ‘creatieve’
- ads normal design word
- mockups normal design word
- identity normal design word
- skilled worth a note
- fast-paced worth a note
Caught correctly
- je word → je wordt
- competitief salaris vague pay
- is een pré a nice-to-have, said clearly
- teamplayer says nothing
| Why would people switch? | Vacaturescan rules only | With an AI model the fix | What they already use ChatGPT · Grammarly · Apple |
|---|---|---|---|
| Reads context | No, and this is why it failed | Yes | Yes |
| Free to use | Yes | Yes | Yes |
| Already part of their day | No | No | Yes |
| A proof label on every check | Yes | Yes | Not relevant |
| Reason to switch? | No: it misreads design words | No: equal to free AI | Already there |
Five ways to save it
The fix was also the problem
An AI model inside would fix the language. Then the tool is as good as free AI, not better.
- Fix the rules Patches symptoms
- Add an AI model Level with ChatGPT
- Lean on privacy Not a reason to switch
- A loader that feels like work Needs real work
- Compare with the local market Not better than AI
Part 3 of 3 · Wed 15 Oct
Why did I stop?
Short answer
Every fix made it equal to free AI, not better. The one idea AI can’t copy failed a test by hand. So I stopped it the same evening.
07 · The decision
How I decided to stop
I asked six questions, in this order. Every answer pointed the same way. The call came on day 3, the same evening as the hand test.
6 questions, asked in order
- What do people compare it to? Not to “nothing”: to Grammarly, Apple autocorrect and ChatGPT. It has to beat those
- Is the gap a bug or the design? The design. Rules can’t read context. Fixing it = rebuild on AI
- Does the fix make it better, or only equal? Equal to free AI. No reason to switch
- Is there one thing only I can offer? Maybe: a comparison with the local job market. Test it by hand first
- Did the hand test win? No. It didn’t feel better than AI. Stop
- What does keeping it cost? Wrong marks make me look clumsy right before I offer the Afhaakcheck. And time away from booked checks. Stop now, not later
- Stop the project Wed 15 Oct, 19:52. “Take it down, but I still want it as part of my portfolio.” Stopped
My rules for stopping
3 of 3 met · Wed 15 Oct
- The gap is in the core idea, not in the details. Met by question 2
- The best fix only makes me equal to the free alternative. Met by question 3
- The last idea fails a cheap test by hand. Met by question 5
Same evening, nothing lost
How I stopped
- Decision The hand test didn’t beat AI. Stop.
- Screens saved The live tool, just before the takedown.
- Page removed The old link sends people to the free Afhaakcheck (301 redirect).
commit 852132e - Tests pass, code kept All site tests pass. The code is tagged, so it can come back.
tag vacaturescan-last
08 · Learnings
What I learned
-
Rules see letters, not meaning. To check language well you need a language model. Then you are as good as free AI, not better.
-
People compare to what they already use. Testers compared it to Grammarly and ChatGPT, not to “nothing”. Ask “why use this instead of ChatGPT?” in the first test.
-
Test the idea by hand before you build. I built first and tested second. The market idea I tested by hand first, and that saved days of building.
-
Fast isn’t the same as useful. 0.3 ms per key didn’t matter to anyone. Being right did.
-
Stopping is a result. Three days to a clear no is cheaper than three months of a tool nobody needs.
-
What I kept
- The code, tagged vacaturescan-last
- Two rounds of research and sources
- 19 screenshots, live and build
- This write-up
Result
Stopped on day 3, 57 hours after the brief. The old link now sends employers to the free Afhaakcheck.
More work
More case studies
LAPAS
Garden firms lose hours on requests that never fit. I designed and built a six-question request flow that tells them the budget, the garden and the wishes before the first call.
Questions 6
Live in 5 days
Developers 0