Fynbos Wellness: 5,638 Real Products,
Modelled End-to-End

1 · A catalogue this size actually behaves like a business

442 brands across 173 product types is a real assortment problem, not a toy dataset — the top 5 product types alone (Herbal Medicine, Face Care, Combination Products, Minerals, Bath & Shower) account for 1761 of the 5,638 products — real long-tail assortment economics, where a handful of categories carry most of the depth and hundreds of smaller categories fill out the rest.

Top brands by product count

2 · The catalogue, with real photos

A sample of 12 real products — the most expensive, plus a random in-stock sample — with their actual images from the retailer's own CDN.

Top 10 most expensive products

3 · Shop by Solution — the real category structure

Fynbos Wellness's own "Shop by Solution" navigation (Bone & Joint Health, Weight Management, Hair/Skin/Nails, Vegan, Detox, Happy Gut, Love Your Heart, Kids' Health, Woman's Wellness, Men's Health, Sexual Wellness and more) is encoded directly in this data as product tags — every number below is that real taxonomy, not a reconstruction.

4 · Worst-performing / at-risk products

This dataset has no sales, order, or conversion data — it's a catalogue snapshot, not a transactions log. "Worst performing" can't honestly mean "lowest sales" here. What it can mean, from signals this data actually has: products that are out of stock (a dead SKU generating zero revenue right now, regardless of demand) and products flagged by the existing IsolationForest as priced well outside the norm for their product type (a real pricing-error or poor-market-fit signal). The table below is a live pandas query joining fact_product against the anomaly model's output — not hand-picked examples.

Top 15 of 281 flagged products, dual-flagged items first, sorted by price descending within each group. Full list: ml/outputs/anomalies.csv joined against fact_product.Availability.

5 · Data quality

6 · Machine learning — four real models, compared properly

Four algorithms (Linear Regression, Ridge, Random Forest, Gradient Boosting) were compared to predict price from product type, brand tier and stock status, via 5-fold cross-validation on all 5,638 products — not one model picked and presented alone.

The simplest model won. Plain Linear Regression (R² = 0.539, std only 0.069 across folds) beat both tree-based models, including Gradient Boosting. At this feature set — mostly one-hot categorical columns, no continuous predictors to carve nonlinear splits from — the extra flexibility of Random Forest and Gradient Boosting bought overfitting, not accuracy. 4 models compared via 5-fold CV on 5,638 rows - a genuinely stable sample size for cross-validation (unlike the sibling property project's 70 rows), so this leaderboard is trustworthy without the fold-safety workarounds that project needed.

What actually drives price

Brand is a real, intuitive price signal in this catalogue: premium and specialist brands carry a structural price premium independent of product type, consistent with how the wellness retail category actually prices.

7 · SEO and AI-search visibility

AI-generated shopping answers (Google AI Overviews, ChatGPT, Perplexity, Bing Copilot) now depend heavily on structured data to know what to cite. Research from 2026 puts 65-71% of AI-cited pages already using structured data, and Google AI Overviews now appear on roughly 14% of shopping queries — a 5.6x increase in four months — with Product schema (accurate price + availability) being what AI shopping answers key off directly. FAQPage schema separately shows a 67% AI-citation rate for queries it directly answers.

Real Product schema, generated from actual data

Real long-tail keyword opportunities (derived from actual co-occurring data)

Long-tail search terms below are built by combining fields that genuinely co-occur in this catalogue — not invented examples:

FAQ block — real questions this data can actually answer

FAQPage schema is worth adding specifically where the answer is a real, computed fact from this data — not a generic marketing claim:

Structured-data adoption and AI-citation figures above are 2026 third-party research findings on AI search behaviour generally, cited for context — not measurements made on this project's own (offline, non-indexed) HTML file.

8 · Competitor and market context

Since this project's reader-facing narrative uses a fictional company name ("Fynbos Wellness"), this section is real market research presented as competitive context — it does not claim "Fynbos Wellness" is, or is modelled directly on, any specific one of the real companies named below.

Real named competitors in this space

Dis-Chem is the dominant pharmacy-led retailer in this category by scale. The real Wellness Warehouse chain ranks around #5 by traffic among online wellness competitors. Faithful to Nature operates online-first. On the brand side, Herbalife, USN and Amway SA are established players selling through multiple channels rather than a single storefront. This is real competitive-landscape research, not a claim about which of these this project's catalogue came from.

What this project actually is (and isn't)

This is a 5,638-product analytical sample with real brand, category and solution-tagging depth — not a live retailer, and not a claim to compete with Dis-Chem or Wellness Warehouse on scale. Its honest value is the kind of cross-brand, cross-category price/anomaly/solution-tagging analysis most single retailers don't expose publicly — a demonstration of the analytical layer, not a market-share claim.

Market-size, CAGR, channel-share and regulatory-compliance figures above are 2026 third-party market research findings, cited for context.

9 · Real business challenges — what a wellness retailer could actually do

11 · How this storefront would track a real funnel

view product → add to cart → purchase, using one of the actual products above. Every button fires a real dataLayer.push(), logged live.

12 · How this was actually built — problems solved, cost held to cents

This project's real story: the first technical approach was blocked, and the fix was to find the right public endpoint, not to give up or fake the data.

Cost discipline as a working practice

Every S3 bucket, Athena workgroup and Glue database created for this 5,638-row build was deleted the same session it was built, then re-verified empty afterward rather than assumed gone from a delete command's exit code. Nothing was left running "just in case."