SA Accommodation Intelligence Platform

Executive Summary

South Africa's short-term accommodation market is booming — and the data proves it. This report analyses 1,011 unique LekkeSlaap listings spanning 18 SA regions and Namibia, overlaid with 30,000 simulated GA4 web sessions and three machine learning models to uncover the pricing dynamics, web analytics patterns, and market intelligence that matter most to accommodation operators, digital marketers, and data engineers.

What We Built

• A Python ETL pipeline that scraped, cleaned, and normalised 1,013 dirty records into a production-grade dataset — fixing price formatting, promotional name noise, duplicate location strings, and outlier prices before any analysis ran.
• A BigQuery star schema (dataset: accommodation_intelligence, region: africa-south1) with six tables, partition pruning, and cluster keys — all loaded via bq load (free ETL within BigQuery's 1 TB/month free tier).
• Synthetic GA4 web events (30,000 sessions, 28,628 funnel events) modelling a real booking platform — traffic sources, devices, SA provinces, POPIA consent signals, and a 5-step booking funnel.
• Three machine learning models: a Gradient Boosting price regressor (R²=0.1726), a Random Forest tier classifier (Accuracy=0.397), and a demand score regressor — all trained on listing features and deployed to BigQuery for ranking.

Chapter 1: The Dirty Data Problem

Before any chart was drawn or model was trained, the raw scraped data told a sobering story: it was filthy. Real-world web scraping almost always is. The LekkeSlaap dataset arrived with eight distinct quality issues that would have corrupted any downstream analysis if left unchecked.

What Was Wrong

How We Fixed It

All cleaning ran in 01_etl_clean.py before a single byte reached BigQuery:

1. Name cleaning: regex stripped the pattern (?:Flash Deal\s+)?(?:\d+%\s*off!?\s*)+
2. Price normalisation: replaced R prefix + spaces, cast to float, capped at R20,000
3. Location deduplication: matched ^(.+?),\s*
\s+(.+)$ and collapsed the repeat
4. Type mapping: dict-based replace() for the two non-standard values
5. Deduplication: drop_duplicates(subset=["property_id"])

Chapter 2: The Market — Who Lists on LekkeSlaap?

With clean data in hand, the market picture snaps into focus. LekkeSlaap is overwhelmingly a self-catering platform — think cottages, apartments, and holiday homes where guests cook for themselves. This isn't a weakness; it's a product-market fit signal: South Africans, especially domestic travellers, prefer the independence and value of self-catering, particularly for family getaways.

Listing Type Breakdown

• Self Catering: 67% of all listings — the platform's core identity
• Guest House: 16% — typically B&B-adjacent, owner-operated
• Lodge: 6% — higher-end, often near nature reserves
• Game Lodge: 3% — premium niche, commanding top prices
• B&B: 3%, Hotel: 2% — underrepresented (LekkeSlaap is not a hotel channel)

Price Tier Distribution

After cleaning, prices span from R99 to R20,000 per night (post-outlier cap), with a median of R1,380 and mean of R2,002. The higher mean vs. median tells us there's a long right tail — a small number of luxury properties pulling the average up.

Regional Hotspots

LekkeSlaap's listing distribution closely tracks South Africa's domestic tourism routes: Cape Town, Garden Route, Gauteng (Joburg + Pretoria), and KwaZulu-Natal dominate. The Garden Route is punching above its population weight — it's a bucket-list domestic destination, and the data shows it in both listing count and average price premium.

Chapter 3: GA4 & Web Analytics — The Digital Layer

Raw listing data only tells half the story. The other half lives in the analytics layer: how users find, browse, and ultimately book accommodation online. To illustrate this, we built a synthetic GA4-schema event dataset modelling 30,000 sessions across January to June 2025.

Traffic Source Mix

Organic Google search (32%) is the dominant acquisition channel — exactly what you'd expect for a destination with strong brand search intent ("LekkeSlaap Cape Town"). Paid Google CPC (18%) supplements organic coverage, particularly for competitive destination keywords. Social media (Facebook + Instagram, 19% combined) drives discovery for users who weren't already searching — critical for inspirational content about new destinations.

Device Reality Check

Mobile accounts for 61% of sessions — but mobile users show a lower average engagement time and higher bounce rate than desktop users. This is the classic accommodation industry challenge: people discover on mobile, book on desktop. Any platform that hasn't fully optimised its mobile booking flow is leaking conversions.

The 5-Step Booking Funnel

Province-Level Booking Patterns

Western Cape users confirm the most bookings — unsurprising given Cape Town's tourism dominance. Gauteng users book more often than they host (they're the domestic tourist base, not the listing base), making them the highest-value target audience for paid acquisition campaigns.

POPIA Consent Mode

South Africa's POPIA (Protection of Personal Information Act) requires explicit consent for analytics tracking — analogous to GDPR in Europe. Our simulated data shows a 78.6% consent rate, meaning 21.4% of events are measured via GA4's modelled (cookieless) measurement. This must be handled in GTM via the consent_update event and gtag('consent','update',...) call, mirroring the Bash/TFG POPIA implementation.

Chapter 4: GTM Implementation — Tagging a Booking Platform

Building a booking platform's analytics layer in Google Tag Manager requires careful event architecture. Unlike ecommerce (where GA4's default purchase event maps cleanly to a transaction), accommodation booking has a longer, more conversational funnel. Here's how to model it:

dataLayer Architecture

Key Events and Parameters

• listing_view: fires when a property page loads · params: property_id, listing_type, price_zar, price_tier, region
• search_nearby: fires on map/filter interaction · params: search_term, filters, result_count
• contact_host: fires on the "Contact" CTA · params: property_id, price_zar, has_promo
• booking_initiated: equivalent to GA4 begin_checkout · params: checkin_date, checkout_date, nights, price_zar
• booking_confirmed: equivalent to GA4 purchase · params: transaction_id, price_zar, payment_method, province
• consent_update: POPIA compliance · updates analytics_storage, ads_storage flags in consent mode

Custom Dimensions to Register in GA4

Register these as event-scoped custom dimensions in GA4 Admin → Custom Definitions:

Chapter 5: Machine Learning — Pricing Intelligence

Three models were trained on the cleaned LekkeSlaap dataset to extract actionable intelligence. All used the same feature set: Listing Type, Region, City, Log(Review Count).

Model 1: Price Regression (Gradient Boosting)

Can we predict a property's nightly price from its characteristics alone? Yes — with a R² of 0.1726 and a mean absolute error of R1066.66. The model reveals that listing type and region are the two dominant price drivers. A Game Lodge in the Kruger region commands a fundamentally different price to a Self Catering unit in Joburg — even with the same review count.

Model 2: Tier Classifier (Random Forest)

A four-class classifier (Budget / Mid-Range / Premium / Luxury) trained with balanced class weights to handle the market's natural skew toward Budget and Mid-Range. Achieved accuracy of 0.397 on hold-out test data. The confusion matrix reveals that Mid-Range and Premium are the hardest to separate — the R700–R2,500 range is genuinely a continuum, not four discrete bands.

Model 3: Demand Scorer (Gradient Boosting)

A composite score from 0–100 combining normalised review count with tier-based weights. Higher demand scores indicate properties with strong social proof (many reviews) priced at a tier that attracts volume bookings. Achieves R² of 0.9606. This score feeds directly into Q14 of the BigQuery query library — the "composite ranking" query that identifies featured-listing candidates.

Chapter 6: BigQuery Architecture — Free ETL in the Cloud

One of the most common misconceptions about Google Cloud is that ETL is expensive. In this platform, all data loading and transformation is free:

Free ETL Tools Available in BigQuery

1. bq load — The CLI command loads CSV/JSON/Parquet from local disk or GCS buckets. Loading is free; only the resulting storage is billed (≈R0.017/GB/month in africa-south1).
2. GA4 → BigQuery Daily Export — Enable in GA4 Admin → BigQuery links. Automatically exports events_YYYYMMDD tables every 24 hours, free, no pipeline needed.
3. Scheduled Queries — SQL-based transformations on a cron schedule. The first 1 TB of query data processed per month is free; typical transformation queries for a 1,000-property dataset cost pennies.
4. Dataform — Google's native dbt-equivalent. Define SQL transformations as version-controlled SQLX files; Dataform schedules and runs them inside BigQuery at no extra cost beyond the query compute.
5. BigQuery Omni — Query data in AWS S3 or Azure Blob Storage without moving it. Useful if the LekkeSlaap scrape output lives in a different cloud.

Dataset Structure

Chapter 7: Conclusions & Recommendations

For Accommodation Operators

• If your listing has under 50 reviews and your price is below the regional median, you are almost certainly underpriced — the ML model will flag you in ml_predictions.
• Flash deals work — but our data shows properties running promotions already tend to have lower base prices, suggesting they're competing on price rather than quality. Invest in review generation before discounting.
• Mobile is discovery; desktop is conversion. Make your photos exceptional on mobile (first impression) and your booking form frictionless on desktop (last mile).

For Platform Product Teams

• The biggest funnel drop (−40%) is listing view to search nearby. A "similar properties" feature with ML-ranked alternatives could recapture this audience before they exit to a competitor.
• POPIA consent rate of 78.6% is acceptable but improvable. A cleaner consent banner with a clear value exchange ("Accept to see personalised recommendations") typically lifts consent by 8–15 percentage points.
• Game Lodges and boutique Hotels are underrepresented (combined <5%) but command the highest prices and demand scores. A supply-acquisition campaign targeting these property types would improve revenue per listing without diluting the platform's self-catering identity.

For Data Engineers

• Partition your BigQuery tables on event_date from day one. A year of GA4 data without partitioning will cost 10–50× more to query than the same data with proper partition pruning.
• Register property_id as a GA4 custom dimension. It's the join key that connects your analytics layer to your product database — without it, GA4 data is an island.
• Schedule a weekly Dataform job to refresh ml_predictions. Pricing models drift as new listings appear and the market shifts; stale predictions are worse than no predictions.