Skip to main content
StackraStackra
Case Study
6 min readJuly 12, 2026Verified against codebase

We Tracked 1.1 Million US Business Sites Across Three Years of the Web Graph. Local Businesses Are Sliding Toward the Edge.

Three years of Common Crawl webgraphs (7.9 billion link edges), every domain in our 1.2M-site US corpus, four separate metrics. The pattern is consistent: link authority is draining out of local business categories toward digital-native ones, and for every site that climbed meaningfully, ten declined. What business category you're in turns out to predict your trajectory, right down to the subcategory.

Industry
Internal data research, US business web corpus
Stack
BigQuery, Common Crawl webgraph (2024-2026), CrUX, HTTP Archive
Outcome
A per-domain, three-year authority and popularity panel for 1,142,892 US business sites, and a clear finding: business category predicts authority trajectory even after controlling for starting position, and the categories losing link authority are not the ones losing traffic.

At a glance

Everyone tells small businesses to 'build authority,' but nobody measures which kinds of businesses are actually gaining or losing it. We had three years of complete web graphs and a classified US corpus, so we measured it.

1,142,892
Sites tracked
every registrable domain in the US corpus; 982,437 present in all three annual web graphs
7.9 billion
Link edges analyzed
Common Crawl domain webgraphs: 1.88B (2024), 2.23B (2025), 3.90B (2026)
10 to 1
Decliners vs climbers
74,694 sites dropped 20+ authority percentile points from 2024 to 2026; only 7,657 climbed that much
-10.6 points
Steepest subcategory decline
used car dealers, optometrists, and apartment communities; over half of each dropped 10+ points
-0.4 to -0.6
Most stable subcategories
roofing and HVAC contractors; only ~10% dropped 10+ points
SaaS, +0.2
Only positive median anywhere
PageRank percentile; every other category declined on every metric

Three pain points, by site type

Tap the one that sounds like your site to read the full story.

1

Raw referring-domain counts lie across crawl years

Affects: Anyone comparing backlink counts across time using crawl-based data (Common Crawl, or any tool built on it).

If this is you

If your referring-domain count dropped 40% year over year, that number alone tells you almost nothing.

What it looks like

Common Crawl's graph construction swings 2-4x between releases. A typical test domain went from 67 referring domains to 37, which looks like collapse, but its percentile within each year's graph barely moved (98.4 to 96.0).

What worked

Convert every count to a percentile within its own year's graph before comparing years. All findings below use within-year percentiles.

Result

Cross-year trends that survive the crawl's own volatility, instead of measuring it.

2

Categories that start lower have more room to fall

Affects: Any cohort trend analysis where groups start at different levels.

If this is you

The obvious cut is confounded: SaaS sites start near the 99th percentile and literally cannot fall far.

What it looks like

The raw category gradient (real estate -6.1 vs SaaS -0.7) could have been pure regression to the mean.

What worked

Re-ran the whole comparison restricted to domains that all started in the same authority band (85th to 95th percentile), so every category had identical room to move.

Result

The gradient survived nearly untouched: real estate -5.8 with 36% of sites dropping 10+ points, SaaS -0.8 with 11%. Same starting line, 3.2x the odds of a serious decline.

What we measured

Four metrics per domain per year, 2024 through 2026, all from the same annual Common Crawl webgraph releases plus Google's CrUX popularity data: referring domains (converted to within-year percentile), PageRank position, harmonic centrality position (how central the site sits in the web's link structure, not just how many links it has), and CrUX popularity rank. Positions are normalized by each year's graph size, since the graph itself went from 173M to 121M ranked domains over the window.

The category gradient, controlled

Every category below started in the same 85-95th percentile authority band in 2024. The differences in where they ended up are category effects, not starting-position effects:

Median authority percentile change, 2024 to 2026 (same starting band)
Real estateWellnessRetailAutoRestaurantsB2BNonprofitSaaS-8pts-6pts-4pts-2pts0pts

All categories started in the 85-95th percentile band. Real estate declined 7x faster than SaaS from the same starting line.

Authority percentile change 2024 to 2026, same starting band
CategoryMedian changeShare dropping 10+ points
Real estate-5.836.1%
Wellness and beauty-5.432.3%
Retail-5.332.0%
Automotive-4.628.4%
Restaurants-4.023.5%
B2B services-2.718.3%
Nonprofits-1.712.7%
SaaS-0.811.2%

n = 1,581 to 54,205 per category. Local and physical business categories decline fastest; institutional and digital-native categories hold.

Subcategories separate even harder

The same controlled comparison at the subcategory level (minimum 1,000 sites each) spreads wider than the categories do, and it splits within categories: sushi restaurants declined -6.5 while breweries declined -1.8.

Share of sites dropping 10+ authority percentile points, by subcategory
ApartmentsUsed carsOptometryClothingHair salonsRoofingHVAC0%15%30%45%60%

Same starting band for every subcategory. Over half of apartment-community, used-car-dealer, and optometry sites dropped 10+ points; roofers and HVAC held.

Subcategory extremes, same 85-95 starting band
SubcategoryMedian changeShare dropping 10+ points
Used car dealers-10.651.5%
Optometry practices-10.651.3%
Apartment communities-10.652.3%
Hair salons-8.141.8%
Clothing retail-7.942.9%
HVAC contractors-0.69.2%
Roofing contractors-0.410.6%

Over half of the used-car-dealer, optometry, and apartment-community sites in the band dropped 10+ percentile points in two years. Roofers and HVAC barely moved.

Centrality falls harder than link counts

Harmonic centrality measures where a site sits in the web's link structure rather than how many links it has. On this metric the local decline roughly doubles: real estate's median harmonic percentile fell -13.7 and wellness/beauty fell -10.0, versus -1.2 for SaaS. Local business sites are not just collecting fewer links; they are drifting toward the periphery of the graph, further from the well-connected core that crawlers and AI systems traverse first. PageRank tells the same story as the link counts (SaaS at +0.2 was the only positive median we found on any metric), which is expected since PageRank is link-volume-weighted. Harmonic is the metric that adds new information.

Median harmonic centrality percentile change, 2024 to 2026 (same starting band)
Real estateWellnessAutoRetailHealthcareB2BSaaS-16pts-12pts-8pts-4pts0pts

Centrality (position in the web's link structure) declines roughly twice as fast as raw link counts for local categories.

Traffic follows a different map than links

Here is the part that stops this from being a simple 'local businesses are losing' story. We ran the same category comparison on CrUX popularity movement (all sites starting from the same popularity tier), and the ordering changed: home services sites hold their link authority better than almost anyone but were among the worst on popularity decline (44.7% dropped a tier), while automotive sites bleed link authority yet came closest to net traffic gains of any category. And content/media sites hold both their links and their centrality while losing popularity faster than any other category at the top tier (52.1% of their well-ranked sites dropped). Sites that keep their citations while losing their clicks is exactly the pattern you'd expect as AI search answers more queries without a visit. Links and traffic are different assets moving on different schedules, which matches what our earlier backlinks research found at the individual-site level.

Share of sites losing a popularity tier, 2024 to 2026 (same starting tier)
PortfolioHome svcsWellnessContentRestaurantsSaaSRetailAuto0%15%30%45%60%

A different ordering than the link-authority chart: home services holds links but loses traffic; automotive loses links but holds traffic; content/media holds links and centrality while bleeding popularity.

How to read your own numbers

The single most important reading rule: the median site in almost every category drifted down, because the linked-domain pool keeps growing with fresh sites while incumbents age. A raw decline is the baseline condition of the web, not a verdict on your site. The question that matters is whether you are declining faster or slower than your category's median. A restaurant that lost 3 percentile points matched its cohort exactly; an HVAC contractor that lost 3 points fell five times faster than its peers. Same number, opposite meanings.

Method and caveats

Panel: every registrable domain in Stackra's 1.2M-site US corpus, joined to the Common Crawl domain-level webgraphs for 2024, 2025, and 2026 (the April-June release each year) and to CrUX popularity data. Absence from a crawl year is recorded as absent, never as zero links. Percentiles are computed within each year's graph among domains with at least one inbound link. Known bias we can't fully remove: Common Crawl covers small local sites more thinly than big ones, which can inflate local-category declines; the starting-band control reduces this but does not eliminate it. Category labels come from the corpus's Gemini-verified 25-type taxonomy, documented in the corpus case study.

LB
Luke Beck
Founder, Stackra
Last verified July 12, 2026
Every number in this write-up was produced by live BigQuery queries against stackra_corpus.us_corpus_authority_panel and stackra_us_corpus on 2026-07-12, the same day of publication, including the start-percentile-controlled re-runs. The panel build SQL is versioned in the Stackra repo (scripts/us-corpus-authority-panel.sql). Common Crawl release-to-release volatility and the local-coverage bias are documented limitations, not footnotes discovered later.

Want to see what bots see on your site?