whatshouldistudy

Find a programme at a specific university with a short questionnaire, optionally supplemented by your own data. The questionnaire covers RIASEC interests (four custom items per area plus an attention check), school subjects and values; the 20 Mini-IPIP Big Five items are optional. Results are available with at least two interest answers per area or imported data, without personality answers. Personality remains a questionnaire step and counts towards the overall progress (67 entries, including optional questions). The custom interest scale is not yet validated in a study. Results rank 80 fields to programmes in the available country datasets. Germany and Austria currently link to official programme searches; their own programme readers are pending. The paid report reveals the best field and programme and offers an optional AI comparison of programme pages with the person's profile.

Professional documentation: /fachinformationen on country sites (/schweiz/fachinformationen locally), /professional on the global site. The manual covers administration, scoring, item keys, evidence status, data flows and retention, adviser self-study, and a proposed validation protocol. It does not claim psychometric validation or institutional approval. No representative German-speaking norms have been established for this implementation. Big Five results therefore show reverse-keyed item means (1–5), requiring all 20 valid responses, rather than approximate percentiles. In questionnaire mode personality affects neither the field ranking nor hidden suggestions or their reasons. Old shared profiles without means are not converted into new scores. The personal AI endpoint requires consent: true in addition to the existing paid access check.

Separate questionnaire session: /beratung on country sites (/schweiz/beratung locally), /counselling globally. Available from the manual, teacher and organisation pages. It starts empty, uses tab memory only and shows descriptive profiles after all 45 activities and at least ten school subjects. It does not mount imports, personal AI, measurement or organisation capture, and does not expose the paid programme ranking. Entry uses a full page navigation to unload previously loaded scripts. Explicit print/JSON copies are personal data; ending a session does not delete downloaded files. Page requests and hosting logs still require review.

Research preparation: evidence log, pilot protocol, research decisions and unresolved rights, and institutional dossier. npm run validation:bundle records exact forms for each language, scoring/source checksums and an unreviewed field-profile template. Review version IDs before regenerating; npm run validation:check (also in tests) detects drift. Keep completed independent reviews separately from the generated template. Catalogue releases must be archived separately before a study. Result wire version 1 is unchanged; new method metadata identifies the procedure, while old shared links without it remain unversioned. Its locale is the current display language, not proof of the language used for older saved responses. No external test items, norms or restricted research datasets were adopted, and no study has been launched.

Four sites from one codebase: the global English site and three German-language country sites: wasstudieren for Switzerland (wasstudieren.ch), findemeinstudium for Germany (findemeinstudium.de) and wasstudieren for Austria (wasstudieren.at). The Swiss site's French and Italian versions are called whatshouldistudy and live at whatshouldistudy.ch/fr and /it. Each country site lists the degree programmes of its own universities and universities of applied sciences and leads with them. Without a domain set they live at /schweiz, /deutschland and /oesterreich.

The app serves at the domain root. WSIS_BASE_PATH mounts it under a path instead (see Deploy on Vercel). The Swiss site also speaks French (/fr) and Italian (/it) on its own domain; a switch in its header leads to the same page in German, French, Italian or, on the global site, English.

The field results are free. The full programme list (every matching programme with the fee for the student's citizenship, earnings, admission rates, filters, CSV) is a one-time payment.

Was arbeiten? — free Swiss job explorer. The results page links to /wasarbeiten (or /schweiz/wasarbeiten under the global domain). It reuses the same browser-local questionnaire and effective interest summaries, including an already enabled content analysis. No new upload, AI call, payment or profile transmission is needed. Shared results are not imported into the viewer's profile. The study paywall and its endpoints remain separate.

The page combines 25 editorial career directions with current Swiss job adverts. Career suggestions compare activity interests, enjoyed school subjects, work priorities and existing interest summaries; they do not assess qualifications. Employer fit compares the preferred city, work arrangement and sector with one specific opening. It neither invents a company culture nor combines conditions from different vacancies. Extra preferences live in the existing site store and are removed by its normal data deletion. Interface text is available in DE/EN/FR/IT.

Run npm run data:jobs to refresh data/jobs/current.json, or node scrapers/jobs.ts --previous previous/jobs.json --out output/jobs.json. The scraper reads public career pages for Swissquote, Swiss Medical Network and Nexthink, plus the public job feeds embedded in the Coop group, federal government and SBB career portals. It reads all pagination levels, observes robots.txt, waits at least 1.5 seconds between requests, and stores only job facts and original links, not descriptions or contact details. The central SmartRecruiters API is excluded because its robots.txt disallows this reader. Switzerland must be explicit in the work country or (for Coop apprenticeships without a country) a Swiss canton. Foreign postings, speculative applications and trial visits are excluded. Publications with the same requisition ID are merged. Each source must return a complete list; truncated, repeated or changing pages fail without deleting the previous source snapshot. These employer groups are still a selection of the market.

The 13 October 2026 snapshot contains 2,954 adverts across 93 employer names and organisational units, including 964 apprenticeships, internships and trainee roles. Coop contributes 2,235, the federal government 421, SBB 147, Swissquote 37, Swiss Medical Network 110 and Nexthink 4. Counts and training places are shown separately in the interface. Subsidiaries of a diversified group get their own sector; unknown employers are not assigned the sector of the parent group.

To refresh a subset, use e.g. --sources Coop,Bund,SBB. Other configured sources retain their original statuses and check dates. The loader accepts larger feeds with a bounded native HTTP request and hourly memory cache, avoiding Next's 2 MB fetch-cache limit. Employer cards are paginated as well as job cards.

.github/workflows/wsis-jobs.yml refreshes daily at 04:41 UTC and when scraper code changes on main. It publishes the separate whatshouldistudy-jobs branch without triggering a Vercel deployment. The app reads its jobs.json with hourly page revalidation and memory caching; WSIS_JOBS_URL can override the full URL. Until that branch exists, the bundled real snapshot is used with its original check dates. Complete reads remove closed jobs; failed or unexpectedly incomplete reads retain the previous source and its old timestamps. Expired jobs and jobs unchecked for seven days are hidden even if a refresh fails. Source status and check dates are visible on the page.

Free programme catalogue for search engines. Every programme can be found for free: /programmes lists the countries with programme data, /programmes/<country> has a search and links to one page per field, /programmes/<country>/<field> lists every programme of that field by level and institution. Each country site has the same for its own country at /studiengaenge and /studiengaenge/<fach>. The catalogue shows name, institution, place, degree, field, language, duration, mode and the links (CATALOGUE_KEYS in lib/programmes.ts, search in /api/catalogue). What stays paid is the ranking for the person, fees by citizenship, earnings and admission rates per programme, the filters across countries and the CSV. Countries appear once the data has programmes for them, the UK as soon as Discover Uni is in. Field pages with fewer than three programmes are not indexed.

Every programme also has a page of its own under its main field, /programmes/<country>/<field>/<name-institution-id> and /studiengaenge/<fach>/<slug> (about 177,000 addresses). Only the id at the end finds the programme; another field or an old slug redirects (308) to the programme's address. The pages show the catalogue facts, the programmes of the same field at the institution and nearby, and the source with its licence. They exist only for countries whose source has an open licence: not for Germany and Austria (LINK_ONLY); PROGRAMME_PAGES_PENDING can hold back others. Sources without an open licence are not used at all until their publisher agrees in writing (PERMISSION_PENDING in scrapers/build.ts): studyprogrammes.ch and the Studiensuche of the Bundesagentur für Arbeit. Switzerland then has the BFS data (free use with attribution), and for the largest universities their own programme lists (ch-hochschulen.ts), one page per field of study at an institution; Germany and Austria get pages per field that link to the official search. Reports and school plans are on sale in both countries; the landing page, report checkout and school plans disclose that individual German and Austrian programmes are not included. Selecting only DE/AT returns no individual programmes. The paid report includes the top field with evidence and matching programmes from the covered countries. Field pages list the programmes as links to these pages, on the global site a few per country. They are built on the first request and cached for a day. Their sitemaps are separate files of at most 40,000 addresses, /sitemaps/<site>-<country>-<n>.xml, listed in robots.txt.

Open source, local analysis by default. The code is public under the MIT licence (LICENSE), and the site says so on the landing page, the start page, in the footer and in the privacy policy, with a link here (NEXT_PUBLIC_SOURCE_URL, default this repository on GitHub). There are no user accounts. Optional usage statistics store only aggregate counters, as described below.

Privacy model: the default analysis runs in the browser. OAuth tokens stay in the tab and exports are unpacked on the device. Parsed Google activity and YouTube export data stays in local IndexedDB for the optional AI step; other raw imports are discarded after analysis. Removing the YouTube source or deleting all data clears that record. Programme requests send field ids, scores, topics and filters to the server. Optional import statistics send only a recognised source category. The separate, explicit AI content step previews a bounded sample of titles and channel context from a user-selected Takeout export, then sends it through our server to Infomaniak (Swiss AI infrastructure). API-fetched YouTube metadata never enters this sample. We do not store or log these requests. The original and AI summaries stay locally.

Licence and paywall

The code is MIT licensed (LICENSE). The paid programme list, the Stripe checkout and the signed unlock token are the way this project finances itself. The MIT licence lets you run your own copy, but the hosted service's payment and unlock tokens are not yours to bypass: do not forge tokens, reuse another person's token or call the paid endpoints of the hosted site without paying. AI coding assistants are given the same rule in AGENTS.md.

Run it

cd whatshouldistudy
pnpm install
pnpm dev            # http://localhost:3000, demo data, free unlock
pnpm test           # matcher, connectors, scrapers (on fixtures)
pnpm typecheck
pnpm build

Without any configuration you get the questionnaire, file exports (Takeout, Spotify), GitHub, the demo dataset in data/sample and free unlocks in development. .env.example lists everything else.

Four sites

	Global	Switzerland	Germany	Austria
Path	/	/schweiz	/deutschland	/oesterreich
Language	English	Swiss Standard German	German (Germany)	German (Austria)
Default filter	anywhere	Switzerland, Swiss citizenship	Germany, EU citizenship	Austria, EU citizenship
Price	9 CHF/EUR/USD	CHF 9	EUR 9	EUR 9
Browser storage	wsis:v1	wsis:v1:ch	wsis:v1:de	wsis:v1:at

All four render the same views (app/views/, app/ui/). The global routes are in app/(global)/… (/start, /results, /fields, /how-it-works, …), the country routes once in app/(local)/[mount]/… (/start, /resultat, /faecher, /so-funktionierts, /datenschutz, /agb, /impressum, …). Configuration per site is in lib/site/config.ts.

Language. The German texts in lib/site/dict.ts, labels.ts and fields-de.ts are written once in Swiss Standard German. lib/site/regional.ts turns them into German and Austrian German: ß where it belongs, „…“ quotes, 10.000 instead of 10’000, and a short word list (Lohn → Gehalt, Resultat → Ergebnis, innert → innerhalb von, Spital → Krankenhaus, Doktorat → Promotion). What differs in substance (data sources, institution types, admission, consumer law) is in DE_OVERRIDES and AT_OVERRIDES in dict.ts and in lib/institutions.ts.

Domains. Each country site is mounted twice: publicly at /schweiz, /deutschland, /oesterreich, and internally at /ch-site, /de-site, /at-site. proxy.ts rewrites every request on a country host to its internal mount, so a country domain serves the pages at its root with plain links. To give a country site its own domain:

1. Add the domain (e.g. wasstudieren.ch, and www.) to the whatshouldistudy Vercel project.
2. Set NEXT_PUBLIC_CH_URL=https://wasstudieren.ch (_DE_URL, _AT_URL for the others). Its host with and without www. is then that country's. Other hosts, such as a former domain, go in WSIS_CH_HOSTS / WSIS_DE_HOSTS / WSIS_AT_HOSTS and redirect to it with the same path. Old /schweiz/… links redirect to the new domain as well.
3. Optionally rename it: NEXT_PUBLIC_CH_NAME / _DE_NAME / _AT_NAME. The Swiss French and Italian versions can have their own domain and name: NEXT_PUBLIC_CH_FR_IT_URL=https://whatshouldistudy.ch (and _FR_IT_NAME, default whatshouldistudy). That host serves /fr, /it and the sign-in callbacks, sends its root to /fr or /it by browser language (German to the Swiss domain) and everything else to the Swiss domain.
4. Register https://wasstudieren.ch/callback/google (and spotify) with the OAuth providers. Reddit takes a single redirect URI per app, so Reddit sign-in works on one domain only (or with a second Reddit app).

Sitemap and robots.txt answer per host.

Design. Cream paper, ink outlines, hard offset shadows, sticker colours (pink, lime, yellow, sky, orange, violet) and Bricolage Grotesque (app/fonts/, SIL OFL). All four sites share the violet accent and the sparkle logo; only name and language differ. Tokens and components (.card, .btn, .chip, .sticker, .hl) are in app/globals.css; light and dark mode.

How the matching works

lib/engine/ (documented in detail on /how-it-works):

1. Connectors (lib/sources/) turn each source into SignalItems: a subscription, a liked video, a watched video, a search, a podcast, a subreddit, a repository. Each has a base weight (a subscription 3, a view 0.3, your own repo 3–5) and, where known, a time.
2. Classifier (classifier.ts): a compiled multilingual lexicon (lib/taxonomy/lexicon.ts, ~6,400 terms in EN/DE/FR/ES/IT including creator and channel names) scores every item against every field; 100,000 titles take under a second. Separate term lists detect learning vs. entertainment.
3. Accumulator (accumulate.ts): channel context (an item inherits half its channel's profile), diminishing returns per channel and month (log₂(1+n)), learning weighting (0.7×–1.3×), music kept apart. Output is a small SourceSummary per source, the only thing stored (localStorage).
4. Scoring (scoring.ts): per field, the log-lift of your share against a popularity baseline, shrunk by evidence (n/(n+4)), weighted by persistence across months, fused over sources by reliability and volume, with a bonus for cross-source agreement. Combined with RIASEC congruence, Big Five tilts, subject fit and value fit; weights adapt to what data exists. Also produces hidden matches, insights and a yearly interest timeline.

The taxonomy (lib/taxonomy/fields.ts) gives every field a RIASEC profile, Big Five tilts, subject demands, what it offers, careers, and codes for the data sources (CIP for the US, CAH for the UK, ISCED-F, OpenAlex subfields).

Optional semantic content analysis

The unlocked result has a separate «Understand my viewing with AI» control. It prepares the existing YouTube import locally, including after a reload or checkout return. Review the sample, then explicitly agree to send it; there is no separate file picker. Parsed export signals are saved in IndexedDB, separately for each site, and replaced or deleted with the source. If browser storage is unavailable, the import remains usable in memory for the current visit. Imports from older versions only have summaries and need one import through the start page. Existing AI summaries remain usable without it. A local before/after table shows the top ten field ranks and points, including when the order is unchanged.

contentPlan selects at most 240 distinct records, including up to 60 subscriptions and 60 recurring channels, with recent videos spread across channels. Each includes only its title, channel name and up to four example titles. Account identifiers, URLs, dates, watch counts, searches, comments, music-listening entries and API metadata are excluded from the request. /api/content-analysis requires the existing unlock and consent flag, processes 40 records per batch, and charges the existing AI and purchase budgets. Gemma at Infomaniak interprets the content against the known subject taxonomy, can abstain, and must not infer political beliefs or other sensitive traits of the viewer.

Each direct reading replaces the keyword classification of that title. A recurring channel subject provides half-strength context to other videos in that channel, explicitly marked as such in the evidence; it cannot override an individually reviewed video. A virtual channel record never adds activity. Repetition, dates, subscriptions and the usual questionnaire weighting remain in the local scoring. Semantic subjects also reach programme-content matching. This is a sample-based interpretation, not a transcript analysis or validated measure of study suitability. The original summary remains available; disabling or deleting the AI summary restores it. Failed or cancelled batches do not replace a previous result.

Data sources and their setup

Source	What	Setup
YouTube Data API	subscriptions, likes, playlists, uploads, channel details	Google Cloud project, enable YouTube Data API v3, OAuth client (web), consent screen with scope youtube.readonly. Until Google verifies the app (sensitive scope, a few weeks, needs the privacy page and a demo video), only listed test users can sign in. Default quota is 10,000 units/day ≈ 50–60 full scans; request more via the quota form.
Google Takeout	YouTube history, Google searches and readable Google News article titles from My Activity; separate YouTube exports remain supported	none: one main link preselects My Activity (custom/my_activity); upload the .zip (all split parts together). Other imports are collapsed under optional sources.
Spotify Web API	podcasts, episodes, audiobooks, artists/genres	app on developer.spotify.com. Development mode allows 25 allow-listed users; public use needs Spotify's extended quota approval, which Spotify currently grants mainly to organisations. The Spotify data export works for everyone.
Instagram data download (JSON)	topics Instagram files you under, accounts you follow, likes, saved posts, searches, posts seen	none, the user requests it in Accounts Center and drops the .zip; direct messages are skipped. Instagram's API only covers business accounts and their own posts, so there is no sign-in.
TikTok data download (JSON or TXT)	searches, accounts you follow, hashtags, own comments, how many videos watched and liked	none, the user requests it in the app (Settings → Account → Download your data). TikTok's Login Kit gives only profile and own videos.
Reddit API	communities, saved, upvoted, own posts	app of type installed app on reddit.com/prefs/apps
GitHub	public repos and stars	none (public API, 60 requests/hour per visitor IP)

My Activity is the default import. YouTube leads; Google searches and News are supplementary (source weight 0.15 versus 0.85 for the YouTube export). URL-only and unavailable video titles are excluded, including from optional AI context. News titles are classified locally. The first AI selection (up to 240 entries) uses only YouTube; the explicit maximum run (up to 1,000 entries) can include up to 200 readable News titles after preview and consent. Google searches are never sent to this content analysis. AI News results replace the corresponding Google summary, keeping its lower source weight; disabling AI restores both originals. Activity weight halves every two calendar years; activity six or more calendar years old remains historical but no longer affects current interests. This is a product heuristic, not a validated psychological norm. Reimport older saved summaries to apply the new title filtering.

Further sources worth adding later (account access exists): Google "My Activity" search history is already read from Takeout; Goodreads and Letterboxd CSV exports (books and films); Steam (owned games and playtime via Steam Web API); Strava (sports).

Programme database

scrapers/ with a weekly GitHub Action (.github/workflows/wsis-daten.yml):

Scraper	Coverage	Licence
us-scorecard.ts	every US college: programmes by CIP and level, in-/out-of-state tuition, median earnings and debt per programme, admission rate	public domain
uk-discoveruni.ts	every UK undergraduate course at all ~460 providers, from the search API behind discoveruni.gov.uk's course finder: award, length, mode, campus, subjects (CAH, names mapped with the OfS lookup in scrapers/lib/cah.ts), provider website from its Discover Uni page	CC BY 4.0
fr-parcoursup.ts	every French first-year programme on Parcoursup, statutory fees, capacity, admission rate	Licence Ouverte 2.0
ch-bfs.ts	every Swiss Bachelor and Master: students per institution, subject (Fachrichtung) and level, from the BFS PXWeb tables (universities, FH, PH); institution type, fees, languages and admission from lib/ch-institutions.ts	open use, «Quelle: BFS»
ch-hochschulen.ts	the degree programmes as ETH Zürich, EPFL, the universities of Zürich, Bern, Basel, Luzern, St. Gallen, Fribourg, Geneva and Lausanne, ZHAW, FHNW, BFH and HSLU list them on their own websites: name, level, teaching language, credits, duration, specialisations, link. Facts only, no descriptions; one request every 1.5 s. Replaces the BFS row for the same institution, level and field	facts from public pages, each with a link to the source
ch-weiterbildung.ts	Swiss continuing education (MAS, DAS, CAS; executive MBAs count as MAS) as ZHAW, FHNW, BFH and the universities of Basel, Geneva and Lucerne list it: name, type, link, and where the list has them ECTS, language, campus and the fee for the whole programme. Lists only, no programme pages; one request every 1.5 s. ETH, UZH, HSG, UNIL/EPFL, Fribourg, Bern and HSLU load their lists in the browser only and are not read yet. Own source and own levels mas, das, cas: counted apart from the degree programmes (byCountry.continuing), listed after them on field pages, ranked only when the person picks «Weiterbildung», never read by the AI facts run	facts from public pages, each with a link to the source
ch-studyprogrammes.ts	not run until swissuniversities agrees: its legal notice reserves public or commercial use of its content. The Swiss catalogue uses the BFS data only	none stated

build.ts merges everything into programmes/<field>.json, research/<field>.json, directory/<CC>.json, catalogue/<CC>.json (the free catalogue, without the paid fields), meta.json and stats.json, keeps the date each programme was first seen (so new programmes are flagged), and keeps last week's data for any source that failed. The result is force-pushed as a single commit to the branch whatshouldistudy-data; the site reads it from raw.githubusercontent.com (configurable with WSIS_DATA_URL). Runs on other branches only upload the result as an artifact.

Secrets for the workflow: SCORECARD_API_KEY (free at api.data.gov/signup, strongly recommended), optionally OPENALEX_API_KEY and OPENALEX_EMAIL.

State of the first live runs (October 2026):

Source	Result
Switzerland (BFS)	1,069 programmes for 2025/26: 536 university, 75 ETH/EPFL, 373 FH, 85 PH, at 43 institutions
Germany (Studiensuche)	20,094 programmes from 21,585 offers: 12,467 university, 5,651 HAW/FH, 1,182 private, 726 art and music colleges, 37 dual, 31 public administration
Austria (BMFWF open data, data.gv.at)	2,556 records, 2,364 programmes (1,047 university, 593 FH, 253 private, 356 arts, 115 PH). Doctorates and continuing education are left out.
France (Parcoursup cartographie)	25,805 programmes read, ~22,300+ classified
United States	with DEMO_KEY only ~1,000 of ~2,700 institutions (≈45,000 programmes) before the rate limit; complete with a free API key
OpenAlex	14,657 universities in 65 countries
Directory	10,268 universities in 200 countries
United Kingdom (Discover Uni)	30,461 programmes at 461 providers (26,433 bachelor, 2,145 integrated master's, 1,796 foundation degrees and HNDs, 87 medicine/dentistry/vet). HESA's .zip download sits behind a Cloudflare challenge, so the scraper reads the public search API of the course finder instead.

The scrapers read column names defensively and log the columns they find, so when a publisher renames something the run log shows it. pnpm data:all runs everything locally; pnpm data:sample rebuilds the demo dataset from scrapers/fixtures (CI checks it is up to date).

Germany stays in link mode until a publisher agrees in writing. Reading the universities' own websites, as in Switzerland, is ruled out there: unlike Switzerland, both countries have the database maker's right (§ 87a ff. UrhG in Germany, § 76c ff. UrhG in Austria). The Studiensuche of the Bundesagentur für Arbeit takes its programmes from the HRK's Hochschulkompass, so the permission needed is the HRK's: it gives a data export to cooperation partners with a nationwide programme search.

Austria reads the open dataset «Studiengänge» of the BMFWF on data.gv.at, licensed under CC BY 4.0. studienwahl.at confirmed in writing that the same data is published under that licence. For Austria, unidata.gv.at publishes student numbers per public university and ISCED field under CC BY; it lacks programme names but could add student counts.

Next countries with official open data to add: Netherlands (Studiekeuzedatabase, licence on request), Italy (Universitaly), Spain (QEDU), Australia (QILT), Canada.

Measurement

lib/measure.ts and app/ui/measurement.tsx, configured with the variables under Measurement in .env.example; without them nothing loads.

• Google tag (gtag.js) with GA4 and Google Ads, Consent Mode v2. The country comes from /api/config (x-vercel-ip-country): EU/EEA and UK visitors are asked first (opt-in), everyone else, Switzerland included, gets a notice and runs until they refuse (opt-out, Art. 45c FMG). Where consent is missing, NEXT_PUBLIC_CONSENT_MODE=advanced (default) still loads the tag denied for cookieless pings; basic loads nothing. «Cookie settings» in the footer reopens the choice.
• Events: page_view by hand on each route with a cleaned address (only utm_* and click ids survive, OAuth callbacks are never measured), connect_source (kind only), generate_lead (own result shown, once per session), begin_checkout, purchase (Stripe session id as transaction id, amount and currency from Stripe). Each of the last three can also fire a Google Ads conversion by label. Purchases carry a SHA-256 of the buyer's email for enhanced conversions, hashed in /api/unlock.
• Never sent to Google measurement: anything derived from the sources or the questionnaire. YouTube API data is under Google's Limited Use rules and may not reach advertising, not even as a field id.
• Vercel Web Analytics for cookieless page counts of every visitor (NEXT_PUBLIC_VERCEL_ANALYTICS=1).

The landing hero recognises Google Ads arrivals using nonempty gclid, gbraid, wbraid, or utm_source=google with a paid medium (cpc, ppc, paidsearch, paid_search, paid-search, cpm, display). Other arrivals keep the free-start hero unless they are returning participants. Sales-paused sites retain their normal entry. Preview locally at /schweiz?utm_source=google&utm_medium=cpc.

Google Ads paywall A/B test

google-ads-paywall-v1 randomly assigns eligible unpaid Ads visitors 50/50: direct shows the buy-first hero (local one-time price, AI analysis and TWINT logo), late shows the usual free-start hero with payment on the results page. Only the hero/badge change; other free entry links remain available. Prices, Stripe payment and signed unlock checks remain the same.

Each site runs for 14 days from its first participant. On the first eligible visit after that deadline, one atomic Redis operation chooses the higher unique buyer / assigned browser rate, using purchases made during that 14-day test and confirmed by webhook before selection. On the owner's explicit request, selection happens even with small samples: this is an automatic business rule, not a statistically established winner. Ties, including zero sales, select direct. The saved winner is permanent for this experiment and new eligible Ads visitors receive it; returning participants retain their original assignment until its 14-day expiry. Rollout visitors do not enter the randomised test totals. Later purchases cannot change the saved winner.

Assignment requires the existing measurement choice to permit it, a configured Google tag, working browser storage and the production usage Redis connection. A random browser UUID and variant persist for up to 14 days; the server stores only its hash, site, variant, assignment time and checkout/purchase flags. Checkout sends the validated hash/variant as Stripe metadata. Only verified, paid Stripe webhook sessions count as purchases, deduplicated atomically; repeat payments add revenue but not a second buyer. Attribution expires after 14 days per participant. Later purchases inside that window appear separately from buyers within the fixed test window. Refusing measurement clears local assignment and stops future attribution; existing aggregate counts remain. Individual Redis identifiers expire within 90 days; the aggregate winner and decision snapshot persist. Privacy notices describe this in DE/EN/FR/IT.

Without measurement permission or storage, tagged arrivals retain the direct hero and are outside the experiment. Existing buyers are not newly enrolled. The denominator is assigned browsers, not unique people or button clicks. Cleared storage, bots, request failures and blockers affect coverage. Conversion latency can favour the faster flow near the deadline; this fixed-window rule does not wait an additional 14 days before choosing. Compare sites separately. Revenue is gross paid revenue before refunds.

pnpm experiment:report

Uses the same USAGE_REPORT_URL, USAGE_ADMIN_TOKEN and optional Vercel bypass as usage:report. The protected GET /api/paywall-experiment reports each site's start/deadline, saved winner and decision snapshot, assigned visitors, unique checkouts/buyers, payments and revenue. No GA4 custom dimensions are required. Do not use synthetic production arrivals to test assignment: they affect conversion denominators. The actual Redis/Lua integration test runs against a disposable local Redis with REDIS_TEST_CLI and REDIS_TEST_PORT.

Free aggregate usage statistics

No Vercel Custom Events subscription is needed. The counters live in a dedicated Free Upstash Redis database from the Vercel Marketplace, connected to Production only so preview deployments don't count:

vercel integration add upstash/upstash-kv --plan free -m primaryRegion=fra1 \
  -e production --prefix USAGE_ --no-env-pull

The integration sets USAGE_KV_REST_API_URL and USAGE_KV_REST_API_TOKEN (plus a few unused ones). --no-env-pull keeps it from overwriting .env.local. Then add the switch and the admin secret:

NEXT_PUBLIC_USAGE_STATS=1
USAGE_ADMIN_TOKEN=YOUR-RANDOM-ADMIN-SECRET

A database created by hand on upstash.com works too, with USAGE_REDIS_REST_URL and USAGE_REDIS_REST_TOKEN instead.

Generate the admin secret with openssl rand -hex 32. Only the first setting is public; never prefix the other variables with NEXT_PUBLIC_. Redeploy after setting them. No new packages are required. The free tier has quotas; collection can stop when those are reached. Stay on Free to avoid paid overages. See Upstash pricing. This feature still uses normal hosting requests and function execution within your hosting allowance.

Read a report locally (set USAGE_REPORT_URL to your deployed site and USAGE_ADMIN_TOKEN in your local environment or .env.local):

pnpm usage:report 2026-10
pnpm usage:report 2026-10-20
pnpm usage:report 2026-10 --min=10

If Vercel Bot Protection is enabled, also set VERCEL_AUTOMATION_BYPASS_SECRET in .env.local to the project's Settings → Deployment Protection → Protection Bypass for Automation secret. The script sends it in the x-vercel-protection-bypass header, alongside the separate admin authorization. Vercel's browser challenge otherwise returns HTTP 429 before /api/usage runs; retrying cannot solve that challenge. See Vercel's automation access documentation. Keep both secrets private and out of URLs. This requires no app deployment.

The protected GET /api/usage?month=YYYY-MM returns UTC monthly counts; GET /api/usage?day=YYYY-MM-DD returns daily counts for Europe/Zurich (including daylight saving time), with collectingSince recording the first day counted. Daily history starts with this rollout, without backfilling. The CLI labels the first day as partial and distinguishes unavailable or expired history from an empty report. Monthly counters remain unchanged; summing Swiss days can differ from a UTC month around its boundaries. Both reports are ordered by frequency and require the admin secret as a Bearer header; it must not be placed in URLs or browser code. The script prints event, source and programme tables, plus profile codes and their first fields. --min=N (default 5) hides profile combinations counted fewer than N times. The same counters can be inspected in the Upstash console.

• Imports: count recognised source types only after a successful upload, sign-in or GitHub import. No filenames, contents, questionnaire answers or error text are sent.
• Events: result counts each view of an own result (not shared links), unlock each paid Stripe session exchanged for a token in /api/unlock. Both count page views, not people: a reload counts again.
• Programmes: count catalogue IDs, names and institutions returned by the existing teaser and paid programme endpoints. Counts distinguish preview and unlocked. Only authorised responses count; bulk CSV requests over 100 items are excluded. Counts describe API output, not confirmed screen views or unique people. Reloads, retries and filters can count again.
• Profiles: with each result event the browser sends the result's coarse profile class and its first three fields, to check whether interests lead to plausible fields. The class is whether the profile rests on the questionnaire, imported data or both, and the three-letter Holland code, or flat/careless instead of the code. The server accepts only these allowlisted values (parseProfile in lib/usage-stats.ts) and counts one {kind: 'profile', basis, code, rank, field} counter per place. No answers, scores, source types or programmes are attached. Counts describe page views of results, not people, and carry no empirical validity on their own (see docs/validation/PILOT_PROTOCOL.md). For the same check without any visitor data, pnpm validation:profiles runs synthetic profiles through the scoring.
• Storage: one aggregate hash per UTC month and one per Swiss calendar day. Monthly counters expire 90 days after month end; daily counters expire at 00:00 UTC 91 days after the day's date (at least 90 days after local day end). One persistent calendar date marks when daily collection began. No individual event records, exact timestamps, IPs, user agents, cookies, visitor IDs or links of source types or programmes to single results are stored. Upstash receives server-side counter commands, not visitor requests.
• No new consent prompt, browser storage access, or advertising integration. Existing Google measurement remains separately controlled. Hosting providers still process connection metadata: configure access-log retention separately. Cookie-free counters alone do not guarantee a consent exemption in every jurisdiction; assess the deployment's full processing and privacy notice.
• Daily and monthly writes share one atomic operation, are bounded to 20,000 categories per period, and fail without breaking imports or recommendations. There is no visitor fingerprinting for deduplication or abuse controls: bots can distort these approximate totals. Storage errors produce only a generic warning, without event payloads.

The privacy page describes exactly what is switched on.

Paywall

/api/checkout creates a Stripe Checkout session (plain REST, no SDK); /unlocked exchanges the paid session for an HMAC-signed token (lib/server/token.ts, valid 12 months) stored in the browser; /api/programmes and /api/research require it. No database. Free tier: /api/teaser (aggregate counts and fields only, excluding the top field; no programme records). The best programme appears directly below the top field in results; its details are fetched through /api/programmes only after unlock.

After payment, /api/stripe/webhook emails the buyer a personal access link using the Checkout email address and the existing Infomaniak SMTP settings. This runs independently of the browser redirect. The email does not contain questionnaire answers: it explains how to reopen the report in the original browser or complete the questionnaire if no profile is present. Checkout records the interface locale for German, English, French or Italian email copy.

Register a Stripe webhook for checkout.session.completed and checkout.session.async_payment_succeeded, pointing to the canonical HTTPS /api/stripe/webhook URL. Set its STRIPE_WEBHOOK_SECRET in the deployment. The endpoint verifies the raw-body signature and timestamp, ignores unpaid, subscription and unrelated-site sessions, and preserves the existing paid-session check on the access link. SMTP or Redis failures return 503 so Stripe retries. The existing USAGE Redis stores only a SHA-256 session identifier and a sent marker for 400 days, plus a short concurrent-delivery lock. SMTP cannot guarantee exactly once delivery if a process stops after the mail is accepted but before the marker is saved; retries prefer a possible duplicate over a missing access email. Ensure Vercel Bot Protection allows Stripe to reach this signed endpoint. Test deliveries must reach the handler, not a browser challenge.

Right of withdrawal: every site and every buyer gets 14 days to withdraw without giving a reason, also after opening the report, so checkout asks for no waiver. lib/site/withdrawal-text.ts holds the notice and model form in German (regionalised for Germany and Austria), English, French and Italian. The terms show it in full, the buy buttons and Stripe Checkout (custom_text[submit]) show a one-sentence note, and the payment email carries the notice and form as the confirmation on a durable medium. Refunds are issued by hand in the Stripe dashboard.

Owner recovery: POST /api/stripe/recover-email with the existing Authorization: Bearer <USAGE_ADMIN_TOKEN> and JSON { "session_id": "cs_…", "locale": "it-CH" } rechecks the actual session in Stripe and sends to its recorded buyer only. The optional locale corrects old Swiss language redirects. It uses the same duplicate protection; it cannot grant access to an unpaid session.

Price: country sites always use their own currency (Germany/Austria EUR 9, Switzerland CHF 9), for both the displayed price and checkout, including organisation plans. The global site uses the visitor's currency (lib/pricing.ts, country from Vercel's x-vercel-ip-country): CHF 9, EUR 9, USD 9, GBP 15, CAD 23, AUD 25, SEK 189, NOK 189, DKK 125, PLN 75, JPY 2,600, INR 1,099 and more. Checkout uses the Stripe price with lookup key wsis_report (or STRIPE_PRICE_ID) only when its amount matches the displayed price in the selected currency. Otherwise the amount is sent as price_data, so an old Stripe price cannot override the current price. The one-time report has a fixed price with no discount, struck-through price or promotion codes. Run pnpm stripe:setup to update the stored Stripe prices as well. For TWINT on the Swiss site, enable it in the Stripe dashboard (payment methods); Checkout shows it automatically for CHF.

Plans for organisations (/organisations, /organisationen): 100, 1,000 or unlimited full reports a month, CHF/EUR/USD 49, 174 or 474 a month billed yearly, 59, 209 or 569 billed monthly (about 17% more), and the same price levels in every other currency (ORG_PRICES). After Checkout the organisation gets a student link (/start?org=…) and an admin link to its dashboard (usage, Stripe customer portal). Both are the subscription id signed with WSIS_TOKEN_SECRET under different prefixes; nothing is stored on our side. A student's unlock through the link sends a Stripe meter event (wsis_report); the meter's count for the calendar month enforces the limit.

pnpm stripe:setup creates the products, prices (with a currency option per currency) and the meter from lib/pricing.ts, and moves a lookup key to a new price when an amount changes. Run it with STRIPE_SECRET_KEY=… or STRIPE_CLI=live (a logged-in Stripe CLI whose key may write Products, Prices and Meters). The site's own key needs: Checkout Sessions write, Prices read, Subscriptions read, Billing Meters read, Meter Events write, Customer portal write. Enable the customer portal in the Stripe dashboard (Settings → Billing → Customer portal) so organisations can change plans, see invoices and cancel.

Every organisation plan starts with a 14-day free trial (TRIAL_DAYS, Stripe trial on Checkout, card required). During the trial, the plan's normal monthly report limit applies, including unlimited reports for the Institution plan. Usage is counted from the start of the calendar month. The first invoice comes after the trial; cancelling before then means no payment.

For teachers (/teachers, /lehrpersonen): a free 45-minute lesson with a printable worksheet (the print styles reduce the page to the sheet) and a ready-made text and HTML link for school websites. That link points to the home page; the student link of a plan unlocks reports and belongs on intranets and in emails, never on a public page.

Facts database (international programme pages)

.github/workflows/wsis-fakten.yml (after each successful data refresh or by hand, on main only) has Claude Haiku 5.5 at Anthropic read public programme pages for CH, FR, GB, NL, NO, FI, AU, KR and US (scrapers/programme-facts.ts, rules and prompt in scrapers/lib/facts.ts). robots.txt is respected, one request goes out every 1.5 seconds, including robots and redirects. A reading is refreshed when the page text or the reader version changes: the hash and version are kept, the page text is not. Named specialisations and distinctive statements require literal source evidence; incomplete or duplicate response batches are retried once and otherwise retain previous readings. Spending is capped per run (--max-usd).

Distinctive points are translated into DE/FR/IT/EN, and /api/facts selects the reader's language. Original statements remain private alongside their translations. Legacy points without translations are omitted from the display until backfilled; they never silently fall back to another language. Deploy the updated ingestion code before running the facts workflow with translate_only: true (start with country: CH, limit: 0, max_usd: 1). This translates stored statements without fetching university pages or changing scores, source hashes or reading dates. The same per-country spending cap applies; rerun until the log reports zero remaining. Failed batches keep their previous readings, and successful translations are skipped on subsequent runs. Apply to the other countries as well. The local equivalent, using a private readings file, is:

pnpm data:facts --translate-only --previous facts-prev/facts.json --out facts-out/facts.json --country CH --limit 0 --max-usd 1

Do not commit either readings file. The existing authenticated workflow stores the output privately and reads it back to verify all fields and translations. Its Vercel automation header uses the repository secret VERCEL_AUTOMATION_BYPASS_SECRET; GitHub OIDC authentication remains required. Other running server instances may retain cached readings for an hour; paid API responses are cached in the browser for up to ten minutes.

The readings as a whole are never published. Public pages show only the part that describes the programme itself (specialisationNames, publicReading in lib/facts-store.ts): field pages list the specialisation names, Swiss programme pages add a card «Von der offiziellen Seite» with the translated distinctive points, the specialisations, the reading date and a link to the page that was read. Interest profile, topics and the check stay in the paid comparison. The job writes the readings to the site's private database (lib/facts-store.ts: FACTS_REDIS_REST_*, or the counter database USAGE_* if that is all there is) through /api/facts/ingest. The job proves who it is with its GitHub Actions token (lib/actions-oidc.ts: this repository, main, this workflow); the site refuses everything else. No database secret is kept in GitHub. The site reads the readings through /api/facts (paid) and ranks the list in the browser against the answers. The optional step «include my answers» (off by default, after consent) sends the interest profile, weighted lexical topics and favourite school subjects with up to 60 matches to /api/personal-fit, in requests of up to 20 matches each, including smaller lists and programmes without readings. The server computes the numbers by fixed rules on the same calibrated scale as the browser. With a sufficient reading, the AI can adjust that content score by at most 30 points and explain the connections in up to three sentences. Without one, it uses the server's catalogue name, assigned fields and listed focus areas to explain the limited connection; these explanations never change the points. They appear in the programme details with their basis identified. Missing or unusable readings do not disable the comparison. Lexical topics are mapped to the field IDs used in stored readings. Missing readings have neutral content (50) in a list with readings; a list without readings retains the questionnaire scores. Programme ordering respects the person’s strongest fields first, using the programme’s primary field rather than an incidental secondary tag. Within a field, the profile counts 80 % and the content reading 20 %. A generic activity or school-subject overlap is shown as a limited resemblance, never as proof of interest in the programme’s subject; duplicate explanations and the standalone AI fit grade are omitted.

BFH offers are read through the official public search with all result pages, instead of the slow sitemap. Empty or sharply shrinking university/level groups retain their previous catalogue entries while other universities refresh. Programme languages and study modes remain unknown unless their source states them.

Secret: ANTHROPIC_API_KEY (GitHub Actions), already configured in this repository. Only public programme information goes to this offline reader, never user profiles.

The workflow accepts ALL or one country (UK is normalized to GB). Each country has its own job, cached page index and reading cursor; one failure does not stop the others. Defaults: 150 reading attempts and at most USD 1 per country (USD 9 for ALL), after the weekly catalogue refresh and after scraper changes. Missing pages are found through institution sitemaps, at most five institutions and 100 pages per institution per run; successive runs rotate institutions and page batches. This builds coverage incrementally, not a promise that every listed programme already has a reading. DE/AT remain excluded.

Discovery stores URLs only when the title identifies the programme and the degree level, and rechecks that evidence when reading. Ambiguous matches remain unknown. Official Discover Uni and Parcoursup programme records are also accepted; Studyinfo uses its public education JSON record because its HTML is a JavaScript shell. Norwegian institution codes are joined to HK-dir’s checked website list because DBH omits the documented website column. DUO website origins support Dutch discovery.

Local example: pnpm data:pages -- --data scrapers/dist --country US, then pnpm data:facts -- --data scrapers/dist --country US --pages scrapers/out/pages-US.json --out scrapers/out/facts-US.json --limit 150 --max-usd 1. Readings stay private; the page index contains only public URLs, titles and progress timestamps.

Infomaniak AI and data processing

Interactive AI calls use google/gemma-4-31B-it at Infomaniak: export content classification, personal programme comparisons and public description translations. The offline reader for public programme pages uses Anthropic on the owner’s request; see the facts database section above. There is no runtime fallback to another provider. Gemma is a Google-developed model; the inference endpoint is Infomaniak, not Google. The separate private blog automation described below is outside this migration.

Set INFOMANIAK_API_TOKEN and INFOMANIAK_PRODUCT_ID in .env.local and in each Vercel deployment environment that should run interactive AI. Local credentials do not configure Vercel or GitHub automatically. Never use NEXT_PUBLIC_ for these values. ANTHROPIC_API_KEY is used by the offline public-page reader and the explicitly invoked benchmark; it is not used for personal analyses. Daily and per-purchase dollar budgets remain in place; Gemma costs are reserved at a deliberately conservative 2 USD/CHF, not represented as an actual exchange rate. The offline reader reserves conservative $1/$5-per-million-token estimates (above Haiku 5.5’s current prices), retaining reservations on failed requests. Personal comparison groups get up to 40 seconds, with any retry bounded by a shared 50-second deadline. Offline facts use groups of three and a 120-second timeout.

Privacy review, 2026-10-13:

• Infomaniak's DPA/AVV states that it forms part of the general terms. Retain the terms applicable to the organisation's account; this code change does not sign a new agreement or verify the account's contractual records. Its general transfer clause alone is not an exclusive Switzerland guarantee.
• The AI product commitment specifies Swiss infrastructure. Advertising is scoped to AI processing at Infomaniak, not all website processing (Vercel, Stripe and other processors).
• LLM API terms, article 6 exclude training with prompts/responses but allow temporary processing and operational/billing metadata. No documented zero-retention request setting was found. cache: 'no-store' controls our fetch cache, not provider retention.
• Requests are stateless, contain no conversation/user IDs, forbid redirects, and are not cached or logged with their content. Personal results stay in the browser. Cost counters contain usage amounts and hashed purchase references, not prompts. Public programme readings and translations are stored for reuse.
• Consent text names Infomaniak; the remembered consent key changes. Existing content analyses remain local, but cannot automatically trigger a comparison with the new provider until the person agrees. Pending content batches from the old provider are not resumed as part of a new analysis.

The repeatable synthetic comparison is scripts/ai-comparison.mjs. An explicit --run --with-gemma --provider infomaniak-gemma --repeats 1 --out <new-folder> uses the production Infomaniak transport. Reports contain invented data only. Existing reports retain their original provider settings and prompts.

Blog

Every site has a blog at /blog. Every other day the Swiss, German and Austrian site each get their own German article, researched from that country's sources and written in that country's German; once a week the global site gets an English one. Such a post has a site and appears on that site only. The first post was written for every site at once, in English and Swiss German, which Germany and Austria get through regionalize(). French and Italian have no blog. Each post is a JSON file in content/blog (content/blog/index.ts imports them all, pnpm blog:index rewrites it), written in a small Markdown subset (app/ui/markdown.tsx) with links to field pages (field:<id>) and the questionnaire (start). Every article says that it was written with AI and lists its sources.

The posts are researched and written with AI in a GitHub Action of the private repository, which checks out this one, adds the posts, runs typecheck, tests and a production build and pushes to main. That separate automation configures its own provider and is not part of the website's Infomaniak migration. The comparison script's provider SDK is a development-only dependency. checkPost() in lib/blog-check.ts is the bar a post has to clear: free slugs, known fields, at least three sources, the right language for its site, no dashes as sentence connectors and the spelling of its country (Swiss: no ß, «», 1’250, 14.7; German and Austrian: „“, 1.250; everywhere 83% and a capital after a colon). test/blog.test.ts applies it to every published post.

To take a post down, delete its JSON file and run pnpm blog:index.

Deploy on Vercel

whatshouldistudy is its own Vercel project and serves at the root of its domain:

1. New Vercel project from this repository, any project name. vercel.json skips builds when nothing changed and never deploys the data and bot branches.
2. Add the domain under Settings → Domains.
3. Set the variables from .env.example, at least NEXT_PUBLIC_SITE_URL (the public origin, e.g. https://whatshouldistudy.com).

With the system environment variables exposed (Vercel default), the data branch is found automatically, also after moving the repository.

To run it as a multi-zone app inside another site, set WSIS_BASE_PATH=/whatshouldistudy, switch Vercel Authentication off for production and let the host site rewrite that path to this deployment. The deployment's root then redirects to the base path.

Next: whatshouldiwork

The same engine works for careers: fields already carry careers, RIASEC and values; occupations would come from O*NET (US, public domain) and ESCO (EU), salaries from BLS / Eurostat.