8,530 listings, 197 clicks and no customers: building ArendaBE

ArendaBE currently holds 8,530 records marked active in its database. By 14 September 2026, four Telegram Ads campaigns had produced 28,426 impressions, 197 clicks and 60 “Start bot” actions. Confirmed customers acquired from that traffic: zero.

It would be easy to turn those numbers into a cheerful growth update. I find the harder question more useful: why might somebody see a promising home but decline to pay for help finding it? And can I sell a subscription at all when the freshness of the main data source is now in doubt?

This is not the story of a finished startup. It is a candid account of moving from a rough Telegram prototype to a catalogue, swipe cards, translations, payments and analytics—and of why the work is not finished.

The starting point was a concrete frustration. Looking for a long-term rental in Belgium means moving between websites, languages and many similar listings. For newcomers from Ukraine and Russian-speaking renters, Dutch or French forms, unfamiliar rental practices and contract clauses add another layer of uncertainty. I thought search, first screening and application support could live in one familiar place: Telegram.

My first ambition was stronger: focus on listings directly from owners. Without a large acquisition budget or my own stream of landlords, that turned out to be the hardest part. I stopped presenting agency listings as private-owner offers and focused on apartments, houses and studios for families and individual adults. Student rooms were not the core audience.

The business hypothesis was never “people will pay for another list of links”. It was that they might pay for a shorter, clearer route from a promising home to a confident application. That remains a hypothesis, not demonstrated willingness to pay.

The first ArendaBE commits date to 23 March 2026: documentation, source import, normalization of Belgian listings and an experimental search on Tweedehands. The goal was to test “find → show → choose”, not to declare a rental platform complete.

It quickly became clear why row count is a poor proxy for quality. Sale listings slipped into rental results, prices exceeded the target budget, and photos fetched from external hosts failed inside Telegram. Tapping “I like it” could show the same card again. City selection became an alphabetical list of hundreds of places. One notification run sent several messages at once when it was supposed to save the user’s attention.

In housing, these defects are more than polish. One wrong listing or stalled card can destroy trust. I added long-term-rental filtering up to €2,000 per month, swipe cards with photo galleries, a better city flow, on-demand description translation and compact alerts. Some of these journeys still need validation with real users.

I realized that polishing pricing screens while the bot was empty created no value. The first priority had to be a large catalogue that could be refreshed and trusted. We ruled out building on sites that explicitly prohibit automated collection and looked for an open technical interface.

Imino, a Belgian property index with a public MCP endpoint, became the main source. On 6 September, a full server run processed 23 pages and saved 11,205 rental listings. The test sample covered more than 1,300 agency domains. This gave ArendaBE genuine volume for the first time. It also meant an agency catalogue—not verified direct-owner inventory.

On 14 September, 8,530 records were still marked active in the database; 8,215 passed the technical display criteria: rent no higher than €2,000 and a presentation-ready card. “Active in the database” must not be confused with “confirmed fresh at the source”.

I chose not to push visitors straight to a third-party site. A card lets a renter inspect photos, a short description and key facts, swipe, save a promising option and only then open the original source. The bot now has Russian, English, Dutch, French and Ukrainian interface languages; descriptions are translated on demand and can be reused from the database.

Saved searches, Telegram Stars plans, admin analytics and an AI assistant for candidate profiles and lease reviews were added around that flow. Access to the original listing and help with an application are intended paid benefits. But a button existing in code does not prove that a document review is useful or that a customer can pay without friction. Documents, images, repeated alerts and navigation have all exposed real failures.

I increasingly judge the product by one question: can a person calmly assess a home, understand the terms, keep the card and take a useful next step?

Advertising inside Telegram looked like a natural test: the ad and the bot live in the same app, with no landing page or separate signup. I ran separate messages aimed at Russian-speaking, Ukrainian, French-speaking and Dutch-speaking audiences in Belgium. This tests entry into a bot; it does not prove active demand for a subscription.

By 14 September, the four campaigns had accumulated 28,426 impressions, 197 clicks and 60 “Start bot” actions. Total spend shown in Telegram Ads was 5.62 TON. I am not retroactively converting that into euros at today’s exchange rate: it would imply false precision. Server costs, development and my own time are not included. The advertising screenshots are earlier snapshots, not evidence for the later totals above. Impressions and clicks are diagnostic signals, not customers.

In-bot event tracking began on 9 September, so its measurement window does not match the full advertising history. In that period, the bot recorded 56 unique starts, 40 people opening the catalogue, 38 viewing a listing card, 11 liking a listing and 14 opening pricing. The recorded orders contain no completed, non-refunded purchase from an acquired customer; earlier payment events were owner tests that were refunded.

Dividing 14 pricing views by 197 ad clicks would not be a valid conversion rate: the windows and cohorts differ. Still, the direction is clear. Cheap clicks are not the missing ingredient. Either the value is not felt at the right moment, or the route to payment needs work. I cannot responsibly claim one exact cause without walking the entire customer journey and talking to users.

For an indie founder, that is the practical takeaway: “Start bot” looks impressive only until you measure catalogue opens, card views, pricing interest and completed purchases.

On 14 September, an automated policy check detected changed Imino terms. The new wording explicitly prohibits systematic automated collection or reproduction regardless of the tool used. The system moved the source to YELLOW and stopped scheduling new collection jobs. The last completed extraction was at 05:19 UTC; freshness after that point is unverified.

I had previously treated an open MCP endpoint as sufficient grounds for integration. The revised terms exposed the weakness of that assumption: a public interface is not, by itself, a licence to build a commercial index. I will not describe the current catalogue as “updated every three hours” or sell freshness I cannot prove.

The next task is to clarify permitted use of the source and find independent sources whose automated-refresh rights are clear. If that portfolio does not exist, I will have to change the product promise rather than conceal the risk behind attractive numbers. The current restriction can be checked in the Imino Terms of Use; this is the source document, not my legal opinion.

Before scaling ads, I need three things: a reliable and permitted flow of fresh listings, an end-to-end journey from first card to payment, and genuinely useful assistance after payment. I will also test photos, translations, alert delivery, lease reviews and the real reasons people stop at pricing. Plans and prices should respond to observed behavior, not only to my intuition.

ArendaBE has already taught me three lessons. A large catalogue is not the same as a fresh one. A cheap bot start is not evidence of willingness to pay. And the most valuable product work often begins when I stop calling a prototype a business.

If you are building a SaaS product or Telegram bot, take the method rather than a promise of easy success: record the original hypothesis, show actual spend and funnel stages, name data-source constraints, and change course when evidence stops supporting the first plan.