What Does Luxembourg’s Web Talk About?

Using BERTopic on CommonCrawl archives, I applied unsupervised topic modeling to Luxembourg websites. The results reveal what .lu domains talk about — from property listings and investment funds to sushi menus and scout camps.
Published

February 23, 2026

In my previous post, I found that 75% of Luxembourg websites offer French, while Luxembourgish appears on just 9%. But language is only part of the story — what are these websites actually about?

Using the same CommonCrawl sample, I aggregated all crawled text per website-year, removed repeated boilerplate (navigation bars, footers, cookie banners), and let BERTopic discover themes on its own — no predefined categories, just multilingual embeddings, clustering, and keyword extraction across all years at once.

A Map of Luxembourg’s Web

What happens when you let an algorithm read every .lu website and group them by content? The treemap below shows the topics discovered in 2024, organized into 15 categories. Click any category to explore its individual topics, then click the header to zoom back out.

All topics discovered in 2024, grouped by category

Show code 
# Compute category totals
cat_totals = {}
for t in topics_2024:
    s = t['sector']
    cat_totals[s] = cat_totals.get(s, 0) + t['count']

sorted_cats = sorted(cat_totals.keys(), key=lambda s: -cat_totals[s])
total_classified = sum(cat_totals.values())

# Build treemap arrays
ids = ['Luxembourg .lu']
labels = ['Luxembourg .lu']
parents = ['']
values = [total_classified]
colors = ['#faf9f7']
customdata = [f'{total_classified:,} websites classified']

for cat in sorted_cats:
    ids.append(cat)
    labels.append(cat)
    parents.append('Luxembourg .lu')
    values.append(cat_totals[cat])
    colors.append(cat_colors.get(cat, '#999'))
    n = sum(1 for t in topics_2024 if t['sector'] == cat)
    customdata.append(f'{cat_totals[cat]} websites across {n} topics')

for t in topics_2024:
    ids.append(f"{t['sector']}/{t['id']}")
    labels.append(t['label'])
    parents.append(t['sector'])
    values.append(t['count'])
    colors.append(cat_colors.get(t['sector'], '#999'))
    customdata.append(t['words'])

fig_treemap = go.Figure(go.Treemap(
    ids=ids,
    labels=labels,
    parents=parents,
    values=values,
    branchvalues='total',
    marker=dict(colors=colors, line=dict(width=1.5, color='white')),
    customdata=customdata,
    hovertemplate='<b>%{label}</b><br>%{value} websites<br>%{customdata}<extra></extra>',
    textinfo='label+value',
    textfont=dict(size=13),
    maxdepth=2,
    pathbar=dict(textfont=dict(size=13))
))

fig_treemap.update_layout(
    height=550,
    margin=dict(t=30, r=10, b=10, l=10),
    paper_bgcolor='white'
)

fig_treemap.show(config={'displayModeBar': False})

Note: Based on 4,629 classified websites in 2024. The remaining 3,153 websites (40.5%) were too unique to fit any cluster. Categories are manual groupings of BERTopic’s automatically discovered topics.

TipKey Findings

Real estate dominates Luxembourg’s web with 643 websites — more than finance and healthcare combined. In a country where housing prices rose by nearly 90% between 2015 and their peak in 2022 (source), it makes sense that appartement, m², and chambres are the most common vocabulary on .lu domains.

Investment funds are the second-largest single topic (277 websites), with keywords almost entirely in English: fund, investment, asset management, tax, equity — reflecting Luxembourg’s role as Europe’s largest fund administration centre.

Construction is the most fragmented category: 11 distinct topics from roofing (toiture, charpente) to tiling (carrelages, salle de bain), reflecting a highly specialized trade sector where each craft maintains its own web presence.

The topic keywords also reveal the multilingual character of Luxembourg: appartement, chauffage (French), Fenster, Fassade (German), Musek, Scouten, Haff (Luxembourgish), and fund, cloud (English) — all coexisting in the same .lu domain space, as explored in my previous analysis.

How Stable Is This Landscape?

The treemap shows a snapshot of 2024, but has the composition always looked like this? Because the model runs globally across all years, each topic has a consistent identity — so I can track how the share of each category evolved from 2016 to 2024.

Share of classified websites per category, 2016–2024

Show code 
years = sector_evo['years']

# All named sectors sorted by latest-year share (largest first), then "Other"
named = [s for s in sector_evo['sectors'] if s['name'] != 'Other']
named_sorted = sorted(named, key=lambda s: -s['shares'][-1])
other = [s for s in sector_evo['sectors'] if s['name'] == 'Other']

fig = go.Figure()

# Add named sectors (largest on bottom for stacked chart)
for sector in reversed(named_sorted):
    fig.add_trace(go.Bar(
        x=years,
        y=sector['shares'],
        name=sector['name'],
        marker_color=cat_colors.get(sector['name'], '#999'),
        hovertemplate=f"<b>{sector['name']}</b><br>%{{y:.1f}}% of classified websites<br>%{{x}}<extra></extra>"
    ))

# Add "Other" (unmapped topics) at the bottom
if other:
    fig.add_trace(go.Bar(
        x=years,
        y=other[0]['shares'],
        name='Other',
        marker_color=cat_colors.get('Other', '#e0e0e0'),
        hovertemplate='<b>Other (unmapped topics)</b><br>%{y:.1f}%<br>%{x}<extra></extra>'
    ))

fig.update_layout(
    height=500,
    margin=dict(t=20, r=20, b=100, l=60),
    plot_bgcolor='white',
    paper_bgcolor='white',
    barmode='stack',
    legend=dict(orientation='h', y=-0.22, x=0.5, xanchor='center',
                traceorder='normal'),
    hovermode='x unified',
    bargap=0.25
)

fig.update_xaxes(
    title='Year', dtick=1,
    gridcolor='#eee', zerolinecolor='#eee'
)
fig.update_yaxes(
    title_text='Share of classified websites (%)',
    gridcolor='#eee', zerolinecolor='#eee',
    range=[0, 100]
)

fig.show(config={'displayModeBar': False})

Note: Each bar shows the share of classified websites belonging to each category. All 15 categories from the treemap are shown; “Other” covers the remaining smaller topics without a category assignment.

TipKey Findings

The composition is remarkably stable. Real estate, finance, and construction have held their positions since 2016, reflecting an economy whose core pillars haven’t changed.

The one visible shift is data protection: barely present before 2018, it grew into a significant category after GDPR required businesses to publish privacy policies and cookie notices. This isn’t a new sector — it’s existing websites adding standardized legal text that the algorithm detects as a recurring pattern.

Methodology

A single global BERTopic model processes all 81,585 website-years at once, producing consistent topic IDs that can be tracked across years without heuristic matching.

1. Data Preparation
From the same CommonCrawl sample used in my language analysis (81,585 website-years, 2013–2024), I aggregated all crawled pages per website-year. Within each website, repeated paragraphs — navigation menus, footers, cookie banners — were deduplicated by comparing normalized text blocks across pages. This removed an average of 25% of text volume while preserving all unique content.
2. Sentence Embeddings
Each website’s deduplicated text was encoded using BAAI/bge-m3, a multilingual sentence transformer with a context window of 8,192 tokens. The 1,024-dimensional embeddings were computed on GPU (one SLURM array job per year).
3. Global BERTopic
All embeddings across all years are combined into a single BERTopic run: UMAP reduces the 1,024 dimensions to 5, HDBSCAN clusters them (minimum cluster size = 50), and c-TF-IDF extracts keywords from the full text (up to 50,000 characters), using n-grams (1,2) with stopwords for English, French, German, Portuguese, Dutch, and Luxembourgish. BERTopic’s built-in topics_over_time() method then tracks how each topic’s prevalence changes year by year.
4. Category Aggregation
The 191 automatically discovered topics were manually grouped into 15 categories (Real Estate, Finance & Law, Construction, etc.) based on their keywords and representative documents. This mapping covers the 94 largest topics. The remaining smaller topics are grouped as “Other”.

Citation

For attribution, please cite this work as:

Garbers, J. (2026, February). What Does Luxembourg's Web Talk About?
Retrieved from https://github.com/julio-garbers/blog/tree/main/bert_topic_websites_lux

BibTeX:

@misc{garbers2026topics,
  author = {Garbers, Julio},
  title = {What Does Luxembourg's Web Talk About?},
  url = {https://github.com/julio-garbers/blog/tree/main/bert_topic_websites_lux},
  year = {2026}
}

All scripts available on GitHub