I run a visa agency in Bali. The other half of my week goes into Apify Actors, and one of them reads ImmoScout24 — Germany's biggest property portal. You give it a city, a district, a postcode or a pasted search URL; it gives you one row per listing with the rent as a number, €/m², living space, rooms, the full address with coordinates, energy class, the cost breakdown and the realtor.

Because the Actor health-checks itself every night with the same prefilled example, I am not looking at a snapshot. I have a panel: 45 runs between 22 October and 17 November with byte-identical input — locations: ["Berlin"], propertyType: "apartmentrent", maxItems: 50, sortBy: "relevance", details on. 2,172 listing rows, 256 distinct apartments. (A forty-sixth run with the same input stopped at four rows; that one is a budget story, at the end.)

Here is the thing that made me write this up. For each run I took the median of its 50 pricePerSqm values — the simplest possible "what does Berlin cost" number:

run rows median €/m²
22 Oct 14:46 50 11.51
27 Oct 20:08 47 32.40
31 Oct 20:14 47 11.51
5 Nov 20:06 47 33.63
13 Nov 20:10 50 18.20
17 Nov 21:01 50 24.13

Same search. Same Actor. Nothing in the input changed. Across the 45 runs the per-run median ranges from €11.51 to €33.63 per m² — a factor of 2.9. Seventeen runs said under €15, twelve said over €25.

The variable is who is allowed to rent the flat

The 256 distinct apartments split into two populations, and German housing policy is the dividing line.

46 of the 256 require a WBS — a Wohnberechtigungsschein, the certificate Berlin issues to households under an income ceiling, without which you cannot sign the lease. Their median asking rent is €11.50/m². The other 210 listings sit at €19.93/m².

Those 46 come from exactly two landlords: Gewobag (25) and GESOBAU (21), both municipal housing companies. They are 667 of my 2,172 rows — 31% of everything the search delivered.

Run the per-run share of WBS flats against the per-run median and you get a Spearman correlation of −0.926 over 45 runs. The five runs on 5 November that contained zero WBS flats are the €33.63 runs. The three runs on 31 October that contained 31 of them are the €11.51 runs.

The cheap half is not a cheaper version of the same product. One address makes it concrete — three doors on the same street, same landlord, same floor plan:

Hellersdorfer Straße 91 51.7 m² 2 rooms 362 € 7.00 €/m² "mit WBS-140-Förderung" Hellersdorfer Straße 87 51.85 m² 2 rooms 760 € 14.66 €/m² "freifinanziert" Hellersdorfer Straße 89 51.94 m² 2 rooms 761 € 14.65 €/m² "freifinanziert" 

Identical flats, 2.1× the rent, and the difference is a piece of paper. Average those together and you are averaging two different legal markets.

You cannot filter them out by keyword

My first instinct was description.includes("WBS"). That is wrong in both directions.

82 of the 256 listings contain the token WBS or Wohnberechtigungsschein somewhere in their texts. Only 46 of those actually require one. Ten of them say the exact opposite and still match, because every unit in a subsidised complex carries the same marketing boilerplate:

Rund 60 Prozent der Wohnungen werden gefördert und an Interessenten mit Wohnberechtigungsschein vergeben.
("About 60 percent of the flats are subsidised and allocated to applicants with a WBS.")

That sentence is in all of them. What the unit actually is appears one paragraph later, and only there:

Der WBS 140 ist zwingend erforderlich und im Original bei Anmietung vorzulegen. → needs one
Die Wohnung ist freifinanziert … deren Jahres-Netto-Einkommen über den Einkommensgrenzen für einen Wohnberechtigungsschein (WBS) … → needs the opposite

So I tried negation-aware matching — flag the row if a window around the token contains ohne or kein. That handed me a listing classified as "no WBS needed" whose text reads "Der Gebäudekomplex verfügt über **keine* Mieterkeller. Sie sind im Besitz eines gültigen WBS (160, 180 und 220 %)…"* — the negation was about the basement.

What worked on this sample: the title, plus a requirement sentence that puts erforderlich / benötigt / zwingend / im Besitz eines gültigen within one sentence of the token, minus anything self-declaring freifinanziert. 40 of the 46 carry it in the title, which is the honest signal — these landlords put "WBS 220 erforderlich" in the headline on purpose, and the free-financed units in the same building never do. ImmoScout24 has no field for it, so no Actor can hand you a boolean; it can only hand you description and realtorCompany in the same row as the price.

78% of the daily churn is not churn

The second trap in a relevance-sorted panel. Compare each run with the previous day's run by listing id:

gap between runs pairs ids kept out of ~48
under 1 hour 17 47–50
1–10 hours 6 25–49
12–30 hours 18 median 7.5

Day-apart runs share seven rows out of forty-eight. That looks like a market with brutal turnover. It is not. Across those 18 daily pairs, 710 rows were "new since yesterday" — and 557 of them were listings I had already seen earlier in the panel, coming back. Only 153 had never appeared before, about 8 a day. 171 of the 256 listings left the top 50 and returned at least once. One flat ranged from position 1 to position 47 across its 29 appearances.

So relevance is not a stable ordering of a stable pool; it is a reshuffle of roughly the same 256 flats, stable within a session — five runs inside one hour on 5 November returned the identical 47 ids in the identical order — and different between days. Diff two daily snapshots by id and you will report 40 new listings in a market that produced 8. The real number was in searchTotalResults all along: 4,466 on 22 October, 5,274 on 17 November, +18% inventory in 26 days.

For freshness, sortBy: "newest" behaves completely differently — a 25-row Berlin run came back with a median listing age of 0 days and top rows stamped "vor 6 Minuten", exactly one of which wanted a WBS.

Going deeper is the cheap fix

One 500-row run of the same Berlin search, split by delivery position (details were off here, so the WBS count is title-only and undercounts):

positions rows "WBS" in title median €/m² median age
1–50 50 29 11.50 30 days
51–100 50 2 31.72 60 days
101–300 200 0 17.91 120 days
301–500 200 0 22.99 30 days

The subsidised block is a page-one phenomenon. Pooled over all 500 rows the median is €20.61/m², interquartile range €16.34–25.01 — a number I would actually publish. It took 6.6 seconds and $0.0029 of platform usage, because the Actor reads ImmoScout24's own app API: no browser, no proxy.

Two smaller things from the same panel:

  • publishedDaysAgo is quantised. ImmoScout24 publishes only a relative German string, which the Actor converts: across 256 listings the field takes exactly 30 distinct values — 0, 1, 4 … 25, 30, 60, 90 … 1460. One flat walked vor 15 Tagen → vor 25 Tagen → vor einem Monat, then froze at 30 for the panel's remaining two weeks. Median age on page one: 30 days; ten listings were over a year old. It is a bucket, not a date.
  • Berlin page one has no private landlords at all. isPrivateOffer was false for all 256, and 250 of them published a full street address. A 300-row Munich run, same Actor: 168 of 300 private sellers, 218 with a complete address, and realtorPhone filled on 59 — private landlords never expose one.

Calling it

curl -s -X POST "https://api.apify.com/v2/acts/lergassy~immoscout24-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \ -H 'Content-Type: application/json' \ -d '{"locations":["Berlin"],"propertyType":"apartmentrent", "maxItems":500,"sortBy":"newest","includeDetails":true}' 

Filter type == "listing" — error rows ("ImmoScout24 does not know this place") and the closing run-summary row share the dataset and are free. One run option is worth knowing because it is enforced to the row: with maxTotalChargeUsd at $1 the search delivered 50 rows, at $0.05 the same search delivered 47, at $0.005 it delivered 4 — each time finishing green with "Stopped at the maximum cost per run you set — everything written is paid for." That is also the explanation for the 47s scattered through my panel: not a scraping failure, a five-cent budget.

When this is the wrong tool. If you want the German rent index, buy the index — portal asking prices are not signed contracts, and page one is not even a stable sample of asking prices. This reads immobilienscout24.de only, so no Austria or Switzerland. And private landlords' names and numbers are not published by the portal, which means they are not in the data either.

Next on my list: a requiresWbs field built from the title rule above, shipped as a boolean with the matched sentence beside it, so nobody else has to rediscover that keine Mieterkeller means "no basement".