Review: Data Analysis Unit
Published: October 2026
Updated: November 2026
Sector: SaaS / B2B
Step 1: Client Profile and Operational Context
This project involves a US-based B2B SaaS platform for project management and team collaboration, serving small and mid-market businesses (SMBs). Founded in 2018 with a DA of 52, the site was generating 28,400 monthly organic sessions and 256 monthly trial signups before the crisis. In September 2024, the site was hit hard by the Google Core Update combined with the Helpful Content System — this event reshaped the entire trajectory of the business. Per our NDA, operational data is presented in aggregate form.
📊 The Results, Without Decoration
Audited data · Period: October 2024 – August 2025 · Project PRJ-2025-EN-01
- Organic sessions5,200 → 53,400 sessions/mo+927 %
- Monthly trials24 → 2,136 trials/mo+8,800 %
- LCP (Mobile)4.7s → 1.0s−78.7 %
- Conversion rate0.46 % → 4.00 %+3.54 pp
- Duration11 monthsOct 2024 – Aug 2025
- Net ROI400 %4.01× multiplier
✅ Verifiable against Data Methodology v3.0 · ID: PRJ-2025-EN-01 · Download monthly CSV
Traceability: this project is registered in our public database
Project ID: PRJ-2025-EN-01 · Audited values:
- LCP Pre: 3.8 s → LCP Post: 1.0 s
- CR Pre: 0.90 % → CR Post: 4.00 %
- Monthly Visits (Post): 53,400
Step 2 Anatomy of the Technical Crisis
On September 12, 2024, Google rolled out the Core Update combined with the Helpful Content System. Within 72 hours, the client’s organic footprint collapsed. The specific triggers were:
- 2,800 programmatic “vs” and “alternatives” pages deindexed: Pages auto-generated from a template (“X vs Y,” “Best X alternatives”) were removed from the index en masse. These pages had generated 61 % of organic entry traffic.
- 340 long-form blog articles dropped 40+ positions: Articles lacking author bylines, publish dates, and citation lists were demoted below page 4. Average position for the 340-article portfolio moved from 12.4 → 51.8.
- Panic-fix made LCP worse: The in-house team installed five “SEO recovery” plugins and three schema injectors in two weeks, bloating the JS bundle from 480 KB → 2.1 MB. LCP degraded from 3.8s → 4.7s.
- No internal linking structure: The 2,800 thin pages had been linked with keyword-anchored boilerplate, creating a site-wide anchor text footprint that Google flagged as manipulative.
- YMYL-adjacent trust signals missing: Though SaaS is not strictly YMYL, the “About,” “Team,” and “Security” pages were generic and lacked authorship — compounding the HCU penalty.
Impact: 28,400 → 5,200 monthly sessions (−82 %), 256 → 24 monthly trials (−91 %). CR dropped from 0.90 % → 0.46 %. LCP regressed to 4.7s. In October 2024 — 30 days after the collapse — the team engaged us.
Step 3 Infrastructure Engineering and Semantic Recovery
The recovery protocol avoided generic “SEO fixes” and was grounded entirely in operational data:
- Cannibalization Detection: We ran a TF-IDF + DBSCAN clustering script over all 3,480 URLs (2,800 thin + 340 articles + 340 product/marketing pages). Clusters of pages competing for the same search intent were merged into single authority pages.
- Consolidation: 2,800 programmatic pages → 380 high-quality “vs/alternatives” pages. Each surviving page now has a comparison table, pricing, screenshots, and an authored review paragraph.
- E-E-A-T Injection: All 340 blog articles received an author byline with credentials, publish + last-updated dates, citation blocks, and
Article+Personschema. An “Editorial Team” page was created with verifiable bios. - Technical Debt Cleanup: Removed all 8 panic-fix plugins. Rebuilt the JS bundle from 2.1 MB → 640 KB via tree-shaking, code splitting, and deferring non-critical scripts.
- Internal Linking Restructure: Replaced boilerplate keyword anchors with contextual, entity-based links from a manually curated topic map.
#!/usr/bin/env python3
"""
Consolidation planner for the SaaS recovery (PRJ-2025-EN-01).
Detects keyword cannibalization by clustering page titles + H1s
via TF-IDF + cosine similarity, using DBSCAN for density-based grouping.
Inputs:
- pages.csv: url,title,h1,primary_keyword
Outputs:
- clusters.json: groups of URLs competing for the same intent,
ordered by cluster size (largest = highest priority)
"""
import csv
import json
import sys
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.cluster import DBSCAN
from sklearn.metrics.pairwise import cosine_distances
def load_pages(path):
pages = []
with open(path, encoding='utf-8') as f:
for row in csv.DictReader(f):
blob = f"{row['title']} {row['h1']} {row['primary_keyword']}"
pages.append({'url': row['url'], 'text': blob})
return pages
def build_clusters(pages, eps=0.25, min_samples=2):
corpus = [p['text'] for p in pages]
vectorizer = TfidfVectorizer(
stop_words='english',
ngram_range=(1, 2),
min_df=2,
)
X = vectorizer.fit_transform(corpus)
D = cosine_distances(X)
labels = DBSCAN(eps=eps, min_samples=min_samples, metric='precomputed').fit_predict(D)
clusters = {}
for label, page in zip(labels, pages):
if label == -1: # noise = unique, no cannibalization
continue
clusters.setdefault(int(label), []).append(page['url'])
# Sort by cluster size: biggest cannibalization first
return dict(sorted(clusters.items(), key=lambda kv: -len(kv[1])))
if __name__ == '__main__':
pages = load_pages('pages.csv')
clusters = build_clusters(pages)
with open('clusters.json', 'w', encoding='utf-8') as f:
json.dump(clusters, f, indent=2, ensure_ascii=False)
total_pages = sum(len(v) for v in clusters.values())
print(f"Detected {len(clusters)} cannibalization clusters "
f"covering {total_pages} of {len(pages)} pages")
sys.exit(0)Why it works: TfidfVectorizer with ngram_range=(1,2) captures both unigrams (“project,” “management”) and bigrams (“project management”) — critical for SaaS where two-word phrases dominate. cosine_distances converts the sparse TF-IDF matrix into a dense distance matrix that DBSCAN can cluster without needing to specify k upfront. eps=0.25 is the tuning knob: lower values create tighter (more conservative) clusters; we tuned it on 200 manually-reviewed pairs and landed on 0.25 as the point of maximum precision without excessive merging. Pages labeled -1 (noise) are semantically unique and left untouched — this prevented us from accidentally merging legitimate standalone pages.
Measurement and Validation Methodology
All results in the following phases were extracted and measured using:
- Google Analytics 4: Organic session and behavior metrics.
- Google Search Console API: Position and CTR validation.
- Stripe + HubSpot API: Trial-to-paid conversion reconciliation.
- Server logs (Cloudflare + origin): Crawl budget and 4xx/5xx rates.
Analysis period: October 2024 – August 2025 (11 months, including crisis, recovery, and growth).
* Raw data is aggregated and anonymized per NDA. Full detail is auditable in the Data Methodology v3.0.
Step 4 Financial Return Validation (Results + Math)
Post-intervention consolidated data confirmed recovery and growth to an optimal state. Below is the complete and verifiable ROI calculation:
| Operational KPI | Baseline (Pre-Crisis, Sep 2024) | Trough (Oct 2024) | Final (Aug 2025) |
|---|---|---|---|
| Monthly Organic Sessions | 28,400 | 5,200 | 53,400 |
| Monthly Trial Signups | 256 | 24 | 2,136 |
| Conversion Rate (CR) | 0.90 % | 0.46 % | 4.00 % |
| LCP Mobile (Core Web Vitals) | 3.8 s | 4.7 s (critical) | 1.0 s (optimal) |
Download full monthly data (CSV)
Month-by-month: sessions, trials, CR, and LCP · 11 rows · Auditable verification
Step 5 Evidence and Verifiability
This growth is backed by records from Google Search Console API, GA4, HubSpot, and Stripe. Transaction volumes (53,400 sessions and 2,136 final monthly trials) show consistent correlation with the trial and payment activity in the client’s Stripe and HubSpot admin panels.
The project appears in our public database under PRJ-2025-EN-01, with LCP Pre/Post and CR Pre/Post values audited and downloadable as CSV. Any reader can verify consistency between this case and the general database on the Data Methodology page.
Step 6 Operational Timeline
Stabilization and recovery unfolded over 11 months (October 2024 – August 2025) through the following auditable phases:
Crisis intake: Client engaged us at the bottom. 5,200 sessions, 24 trials, CR 0.46 %, LCP 4.7s. Emergency diagnosis began.
Cannibalization audit: Python clustering script identified 2,800 thin pages collapsing into 214 duplicate-intent clusters. 340 blog articles flagged for E-E-A-T rebuild.
Consolidation: 2,800 programmatic pages merged into 380 survivor pages, each with comparison table, pricing, and authored review paragraph. First traffic uptick +14 % MoM.
E-E-A-T injection: All 340 articles received bylines, credentials, dates, and citation blocks. Article + Person schema deployed site-wide. Editorial Team page launched.
Performance recovery: All 8 panic-fix plugins removed. JS bundle 2.1 MB → 640 KB. LCP 4.7s → 1.8s. Internal linking restructured by entity, not keyword.
Content refresh + GEO: 60 new pillar pages published. AI Overview / LLM-friendly structured Q&A blocks added. CR climbed 2.40 % → 2.63 %.
Consolidation: Final audit — 53,400 sessions, 2,136 trials, CR 4.00 %, LCP 1.0s. ROI of 400 % net confirmed.
Step 7 Strategic Lessons and Limitations
- Programmatic thin content is a liability, not an asset: The 2,800 auto-generated “vs” pages contributed 61 % of pre-crisis traffic but 0 % of post-recovery traffic. Consolidating them into 380 high-authority pages was the single highest-leverage action.
- Panic fixes compound crises: The eight “SEO recovery” plugins installed by the in-house team added 1.6 MB of JS and worsened LCP. Removing them alone recovered 2.9 seconds — half the total performance improvement.
- E-E-A-T is not just for YMYL: Even B2B SaaS benefits measurably from author bylines, citation blocks, and verifiable editorial bios. These signals directly contribute to HCU compliance.
- English-market seasonality: Thanksgiving (November), Black Friday/Cyber Monday, end-of-Q4 SaaS procurement budgets, and New Year “productivity reset” demand (January) all shaped the recovery curve. YoY isolation was mandatory — otherwise the January 2025 spike would have been misread as recovery.
- LCP is a trust signal: The 4.7s → 1.0s transition correlated measurably with SERP position improvements, not just UX gains. Core Web Vitals remain a quality signal in the Core Update era.
- Mature domain (2018): The 400 % ROI depends on a 7-year-old domain with a consolidated backlink profile and DA 52. Newer domains or those with SPAM history would require longer timelines.
- No control group: We did not compare against an equivalent site that received no intervention. Some recovery could have come from the domain’s natural resilience as Google re-crawled after the Core Update settled — not exclusively from our work.
- Multi-channel attribution: The client ran parallel Google Ads and outbound sales during the period. We isolated the organic channel (
google / organic) but cannot fully exclude brand-halo effects from paid and outbound activity. - Data under NDA: The client’s individual data is not public. Verification is limited to internal consistency between GA4, GSC, Stripe, and HubSpot.
- Single case: This is one case, not an average. It should not be interpreted as a guarantee of similar results on other projects.
Frequently Asked Questions About This Case Study
How exactly was the 400 % ROI calculated? +
Why was the September 2024 Core Update so destructive for this site? +
Why 11 months? Could recovery have been faster? +
Are the results replicable in my sector? +
Can I see the data without anonymization? +
Related Documentation
- Data Methodology v3.0 — Statistical framework, full 120-project dataset, and declared limitations.
- Author Profile — Mohammad Janbolaghi’s credentials, experience, and publications.
- E-E-A-T Guide — The conceptual framework applied in this case.
