PageRank SEO: Why Modern Search Optimization Is an Information Architecture Problem

September 3, 2026 · Daniel Kliewer

Modern SEO is an engineering discipline where search reputation, information architecture, code quality, machine-readable representation, and user experience converge. It should be understood not as optimizing individual pages, but as optimizing the information graph a website presents to search engines and users.

This essay makes the case, walks through the implementation, and releases the open-source pagerank-seo SDK that turns the idea into something a Python script, a CI pipeline, or an autonomous agent can run.

From keyword stuffing to graph reasoning

SEO used to be a textual problem. In the late 1990s, the problem was literal text: keyword density, meta keywords, hidden text. The original PageRank paper already disagreed. The paper opens by noting that "the importance of a web page is an inherently subjective matter… But there is still much that can be said objectively about the relative importance of web pages." The objective thing Page and Brin said was: a page's importance propagates through the links pointing at it, not the words inside it.

Twenty-six years later, almost every documented Google concern about a website is, at some level, a graph concern:

• Crawlable structure. Crawlers extract links from known pages to discover new ones. A site with a broken internal-link graph cannot be fully crawled.
• Canonicalization. Google clusters duplicate pages and chooses a canonical representative. A site that fans out into hundreds of near-duplicate URLs forces the engine to do more work, with worse results, for everyone.
• Helpful content. The unit of analysis is not a page; it's a site — and a site's quality is, in part, a graph property.
• Sitemaps. Sitemaps tell Google about pages and the relationships between them. The word is literally "site-map" — a map of the structure.

The unit of optimization that matters is the site-as-graph, not the page-as-text.

Modeling a website as a graph

A graph G = (V, E) has nodes and edges. For a website:

• Each page is a node.
• Each internal hyperlink is a directed edge, carrying metadata (anchor text, rel attribute, DOM position).

From this graph we compute PageRank (public-domain algorithm, damping factor 0.85, dangling-node correction), weighted PageRank (edges weighted by link position and rel), in/out degree, weakly-connected components, orphan pages (in-degree zero, excluding seed), Gini coefficient of PageRank, and top-1 PageRank share.

I am explicitly not claiming that the PageRank number produced here equals Google's score. It is an analytical tool for reasoning about the site. All the original PageRank patents expired in September 2019.

The six layers

SEO splits into six interacting layers, each depending on the ones below it:

1. Technical representation — charset, title, meta description, canonical, viewport, JSON-LD, robots directives.
2. Information architecture — URL hierarchy, breadcrumbs, sitemap, orphan pages, deep pages.
3. Link graph — PageRank, weighted PageRank, in/out degree, weakly-connected components, Gini, top-1 share.
4. Semantic relevance — title/content overlap, heading structure, duplicate titles, thin content.
5. User experience — <nav>, <main>, <header>, <footer>, lang, alt.
6. Reputation architecture — HTTPS, /about, /contact, /privacy, Organization schema.

I am careful to distinguish reputation signals (publicly observable properties) from ranking signals (whatever is actually inside Google's head). I do not claim to know the latter.

The Search Quality upgrade

The methodology now incorporates concepts derived from the publicly documented Search Quality Rater Guidelines (General Guidelines v10.1.1, September 2025). This is a significant upgrade.

The guidelines make explicit that SEO engineering must care about two distinct questions:

Page Quality asks: "What is this page and how well does it fulfill its own purpose?"

Needs Met asks: "Given this query and user intent, how useful is this page?"

A high-quality page can fail a particular query. A keyword-stuffed page can fail every query. The SDK models both.

Page Purpose

The guidelines make identifying the true purpose of a page an early step. The SDK classifies pages into: informational, transactional, navigational, entertainment, community, software/tool, commerce, personal expression, reference, service.

Content Classification (MC / SC / Ads)

The guidelines distinguish:

• Main Content (MC): directly helps the page achieve its purpose
• Supplementary Content (SC): contributes to UX without being primary
• Advertisements/Monetization: treated separately

Advertising is NOT inherently a reason for a low Page Quality rating. The SDK models it as a structural observation, not a penalty.

E-E-A-T

Experience, Expertise, Authoritativeness, Trust — analyzed as evidence, not keywords:

• Experience: first-hand evidence, demonstrations, original observations
• Expertise: credentials, depth of knowledge, technical accuracy
• Authoritativeness: recognized authorship, citations, independent references
• Trust: transparency, accuracy, responsible ownership, disclosures

Originality and Scaled Content

The guidelines identify scaled content abuse as a lowest-quality concept: content produced at scale with little effort, originality, or added value. The SDK detects patterns like hundreds of structurally identical pages, minimal variation, and programmatic keyword substitution.

The Reputation Graph

The guidelines treat reputation as requiring investigation beyond what a website says about itself — independent reviews, references, news articles. The SDK observes on-site signals and instructs the agent to investigate independent sources separately.

The SDK: pagerank-seo

pip install pagerank-seo
pagerank-seo audit https://example.com \
    --max-pages 50 \
    --max-depth 3 \
    --output markdown \
    --out-dir reports/

Or in Python:

from pagerank_seo import AuditConfig
from pagerank_seo.auditor import SiteAuditor

auditor = SiteAuditor(AuditConfig(
    start_url="https://example.com",
    max_pages=50,
    max_depth=3,
))

report, quality = auditor.audit_with_quality()

print(report.composite_score)          # 50.9 (real run against example.com)
print(quality.overall_quality_score)   # Page Quality heuristic
print(report.recommendations[:3])      # CRITICAL → LOW prioritized
print(quality.scaled_content.detected) # scaled-content pattern?

The output is a structured AuditReport plus a SearchQualityReport containing per-page quality reports, scaled-content detection, and site reputation abuse risk.

The honest constraints

It is not Google's algorithm. The PageRank computed is the public-domain algorithm from 1999. Google's modern ranking system contains hundreds of signals.

It is not a ranking predictor. The composite score is a project-internal engineering health metric. A "100" score does not guarantee first-page rankings.

It does not measure page speed at runtime. The SDK observes HTML signals only; it does not run a headless browser.

It does not crawl sites you don't own. The crawler respects robots.txt by default and exposes conservative knobs.

The Search Quality Rater Guidelines are not a ranking formula. Human ratings do not directly move individual pages up or down in search results. The project uses these concepts as an analytical framework, not as a reproduction of Google's systems.

The agent workflow: an executable loop

DISCOVER SITE → UNDERSTAND SITE PURPOSE → CRAWL → BUILD INFORMATION GRAPH
→ CLASSIFY PAGE CONTENT → ANALYZE PAGE PURPOSE → ANALYZE TECHNICAL STRUCTURE
→ ANALYZE INTERNAL LINK GRAPH → ANALYZE CONTENT QUALITY → ANALYZE ORIGINALITY
→ ANALYZE REPUTATION → ANALYZE E-E-A-T → IDENTIFY SPAM / ABUSE PATTERNS
→ MODEL QUERY INTENT → ANALYZE QUERY-PAGE FIT → GENERATE RECOMMENDATIONS
→ IMPLEMENT APPROVED CHANGES → RE-CRAWL → COMPARE BEFORE / AFTER

I shipped this loop as a Hermes skill. The critical instruction:

Do not optimize blindly. Measure the information architecture first.

Every recommendation carries a verification_method. No change without a reason, no reason without evidence, no evidence without a re-audit to confirm.

Try it

pip install pagerank-seo
pagerank-seo audit https://your-site.example/ --output markdown

Repository: github.com/kliewerdaniel/pagerank-seo · Hermes skill: skills/seo-audit/SKILL.md · Methodology: docs/methodology.md · Score formula: docs/scoring.md · Sources: docs/research/sources.md

If you find a real bug, open an issue. If you build something with it, I'd like to hear about it.

Daniel Kliewer is a software engineer who has worked around search-quality evaluation and built the open-source sovereign-knowledge-compiler, kleincannon-video-pipeline, and other local-first tools. He writes at danielkliewer.com.