Babyland's pages, built from the products up.
Babyland's categories would not combine into the pages parents search for, so facet pages built top-down came up empty. Our new take on Content Engineering started from the products, and the rewritten category pages now win more clicks per view.
See the results- Client
- Babyland
- Industry
- Baby and children's retail
- Markets
- Sweden
- Period
- Published November 2025, measured through May 2026Nov 2025 to May 2026
- Our part
- Content Engineering: data model, facet pages, category content
- Award
Silver. Most innovative campaign (SEO), small.Official 2026 winners
Challenge
Every page needs products behind it.
Babyland sells baby and children's products online in Sweden, from prams and car seats to LEGO and party supplies: more than twelve thousand of them.
Content Engineering builds a page for every specific need in a range: a modular cot, a fleece footmuff, an organic sleeping bag. Each page has to be matched against what is in stock, or it opens onto an empty shelf with a headline. We call that match product density.
The usual route runs top-down: combine the filters a site already has, check which pairs people search for, then look for products to fill each page. At Babyland the filters were some five hundred legacy categories, brands mixed with product types and occasions, marvel next to marvel 7 and marvel 9. Combined, they gave mostly noise to sift through.
What we found
Start with the products, not the categories.
The structure the pages needed was already in the products. It had just never been written down the same way twice.
So we turned the recipe round. Instead of combining categories and hoping products would fit, we let the catalogue describe itself and built the data model on top of it. A page could only be proposed if products stood behind it, so product density was built in from the first pass.
The work ran in passes, each one small and patient. A pass worked through the products in batches and kept going until three batches in a row changed nothing, an idea borrowed from multi-pass video encoding. Slower than tagging everything at once, but facets only combine when a tag means the same thing on every product.
Two tools split the work. Embeddings, a way of measuring how close two pieces of text are in meaning, proposed: which group a product probably belonged to, which existing page a new one resembled. A language model then decided. Similarity on its own made confident mistakes; the model caught them.
We tried letting the models read the product photos as well. It cost several times as much without a matching gain, so the run that shipped read product text alone. The same kind of test settled the rest: each step got the model that did that job best, not one model for everything.
What Babyland got is a way of working as much as a set of pages: a data model grounded in the range, demand checked before anything is written, and a guard against pages competing with each other. As models improve, each step can be tested again and swapped.
In
Product text
Names, descriptions and specs for every product in the range.
Product photos
Tried, then dropped: several times the cost, no matching gain.
01 Groups
Let the products form the groups
Propose groups
Models read the products in small batches, adding, merging and removing groups until three batches in a row change nothing.
Place every product
Embeddings propose the likeliest group. A language model decides.
02 Attributes
Give each group its own attributes
Build each group's set
Material, pattern, age band, style: a set that fits the group, not the whole shop. It repeats until nothing changes.
Tag every product
Against its own group's set, so a tag means the same thing on every product.
03 Demand
Ask the searcher, then check the shelf
Phrase the combinations
Attribute combinations become Swedish search phrases.
Keep what people search for
Phrases without search demand go before a word is written.
Check the existing pages
Embeddings find the closest page Babyland already has. A language model keeps, drops or merges.
- “leksaker Geomag” dropped: a Geomag page already exists.
- “Marvel samlarfigurer” merged into the bigger Marvel figurer.
- “Babblarna spel” kept: it has demand of its own.
Out Pages
Write only what passed
Write and link
Title, headings and text for each page, with internal links from the same embeddings.
Every page has products behind it.
The work
The result
Chosen more often when shown.
Measurement notesSix months after publication, the rewritten category pages turned a larger share of their impressions into clicks than in the six months before.
The first pages live were Babyland's existing categories, just over two hundred of them, rewritten from the new model: titles, headings, descriptions and body text, with internal links chosen by the same embeddings. They went live in one upload in November 2025.
Clicks rose too, but we don't claim them: the months after include Christmas and the months before do not. Click-through rate and average position depend less on the season, which is why we lead with them.
The new facet pages were prepared with their product sets and queued for a phased rollout, so none of these figures come from them.
- Click-through rate on the rewritten pages, 1.62% to 1.78%
- +9.5%
- Average position on the same pages, from 10.21
- 8.91
Measurement notes and definitions
- Sources
- Google Search Console for clicks, impressions, click-through rate and average position.
- Scope
- Organic Google search to the 228 existing category pages on babyland.se rewritten in November 2025. All queries to those pages.
- Window
- 17 November 2025 to 26 May 2026 (191 days) against 1 May to 8 November 2025 (192 days). The publication week, 9 to 16 November, is in neither.
- Excluded
- The new facet pages, not yet published. Paid search is not in Search Console data.
| Period | 1 May to 8 Nov 2025 | 17 Nov 2025 to 26 May 2026 |
|---|---|---|
| Clicks | 31,462 | 37,694 |
| Impressions | 1,940,800 | 2,123,469 |
| Click-through rate | 1.62% | 1.78% |
| Average position | 10.21 | 8.91 |
1 May to 8 Nov 2025
- Clicks
- 31,462
- Impressions
- 1,940,800
- Click-through rate
- 1.62%
- Average position
- 10.21
17 Nov 2025 to 26 May 2026
- Clicks
- 37,694
- Impressions
- 2,123,469
- Click-through rate
- 1.78%
- Average position
- 8.91
Google Search Console, babyland.se, the 228 rewritten category pages.
Not a year-on-year comparison, so the rise in clicks and impressions is not claimed. The weekly series in the entry covers 227 of the 228 pages.
What this case says
Build the data model from what you sell.
The hard part was never generating pages. It was deciding which ones deserve to exist.
Content Engineering is only as good as the data model under it. When a site's own categories cannot be combined, tidying them keeps their faults; the products already hold the structure. Build from them, and product density stops being a check at the end and becomes the place you start.
Your case
Are your categories hiding your range?
Bring one problem from your own category; we start with the evidence you already have.
or write to info@pixel.se