The most important and critical aspects of Search Engine Optimization (SEO) in 2026 focus on creating genuinely helpful, trustworthy content while ensuring technical excellence and authority signals. SEO has evolved significantly with AI-powered search (like Google's Search Generative Experience), zero-click results, and a stronger emphasis on user intent and quality over keyword stuffing.

Google (and other engines) uses hundreds of signals, but a few core pillars and factors drive most results.

1. E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness)

This is one of the most critical overarching concepts today. While not a single "ranking factor," Google heavily weighs signals that demonstrate these qualities, especially for YMYL (Your Money or Your Life) topics like health, finance, or advice.

  • Experience: Show firsthand knowledge (e.g., author bylines, case studies, real examples).
  • Expertise: Demonstrate deep understanding through comprehensive, accurate content.
  • Authoritativeness: Build reputation via citations, mentions, and recognition in your field.
  • Trustworthiness: Use clear sourcing, transparent authorship, contact info, secure site (HTTPS), and positive user signals.

Why critical? Low E-E-A-T content gets filtered out, especially in AI summaries and competitive niches. Human, original content consistently outperforms generic AI-generated material.

2. High-Quality, People-First Content & User Intent

Content remains king, but it must solve real problems and match what users are actually searching for.

  • Create comprehensive, original, helpful content that goes beyond surface level (topical depth and entity relationships).
  • Prioritize search intent (informational, navigational, transactional, commercial investigation).
  • Focus on freshness: Regularly update content.
  • Structure for readability and AI citation: Clear headings, concise answers (e.g., 40-60 word summaries), lists, tables, and schema markup.

Tip: Shift from "keyword-focused" to "audience-first." Answer questions thoroughly and provide unique insights or data.

3. Technical SEO (The Foundation)

Search engines must be able to crawl, understand, and deliver your site efficiently. Neglecting this kills everything else.

Key elements: - Core Web Vitals and page speed (fast loading, stable layout, interactivity). - Mobile-friendliness and responsive design. - Proper site architecture (logical URL structure, internal linking, XML sitemaps, robots.txt). - Structured data (Schema markup) for rich snippets. - Indexability (fix crawl errors, duplicates, orphaned pages). - HTTPS security and accessibility.

4. On-Page SEO

Optimize individual pages to signal relevance clearly:

  • Title tags and meta descriptions (compelling and keyword-relevant).
  • Header tags (H1-H6) for structure.
  • Natural keyword usage (primary + semantic/LSI terms).
  • Optimized images (alt text, compression).
  • Strong internal linking to distribute authority.

5. Off-Page SEO & Brand Authority

This builds external trust and authority.

  • High-quality backlinks from relevant, authoritative sites remain powerful (though quality > quantity).
  • Brand mentions, social signals, and overall online reputation.
  • Branded search volume and direct navigation signals.

In 2026, brand salience (being known and associated with your category) is increasingly important.

Emerging/Additional Critical Aspects

  • AI & Answer Engine Optimization (AEO): Optimize for direct answers in AI overviews and voice search. Focus on being citable and visible across platforms (Google, YouTube, etc.).
  • User Experience (UX): Low bounce rates, high engagement, intuitive navigation. Page experience signals matter.
  • Topical Authority: Become the go-to resource in your niche through content clusters.
  • Multi-Platform Visibility: Don't rely only on Google—optimize for YouTube, social, and other discovery surfaces.

Prioritization Advice

  1. Fix technical issues first (crawlability and speed).
  2. Ensure strong E-E-A-T and content quality.
  3. Build topical depth and on-page optimization.
  4. Earn authority through links and branding.
  5. Measure success by visibility, impressions, and business outcomes—not just traffic (due to zero-click searches).

SEO is a long-term, holistic strategy. Tactics change, but fundamentals like helpful content + technical health + authority endure. Focus on users first, and search engines will reward you. Stay updated via Google's official guidelines and tools like Search Console.

Implementing E-E-A-T (Experience, Expertise, Authoritativeness, and Trustworthiness) is a practical, ongoing process focused on demonstrating that your content and site are reliable, created by knowledgeable people, and genuinely helpful. Google doesn't have a direct "E-E-A-T score," but its systems and human raters evaluate signals related to these qualities, especially for YMYL topics (health, finance, legal, safety, etc.).

Here’s a clear, actionable framework to implement it effectively in 2026.

1. Demonstrate Experience (First-Hand Knowledge)

Show that you or your contributors have real-world, practical involvement with the topic.

  • Publish case studies, before-and-after results, screenshots, logs, or original photos from real projects.
  • Include specific details only someone with hands-on experience would know (e.g., challenges faced, lessons learned, exact outcomes).
  • Use first-person accounts or “What We Tested” / “Our Process” sections.
  • For product reviews or tutorials: Share your actual usage, timelines, and real results.

Tip: Document workflows as you work and turn them into content.

2. Demonstrate Expertise (Knowledge & Skill)

Prove deep understanding through quality and credentials.

  • Have content written or reviewed by subject matter experts (SMEs).
  • Create in-depth, comprehensive guides that cover “what,” “why,” and “how.”
  • Add detailed author bylines and bios on every article: Include credentials, years of experience, certifications, affiliations, and links to LinkedIn or professional profiles.
  • Develop author hub pages or dedicated profile pages for key contributors.
  • Use original research, surveys, data analysis, or peer-reviewed sources.

Action: Create or strengthen an “About the Author” section and an “Our Team” or “Experts” page.

3. Build Authoritativeness (Recognition & Influence)

Become a recognized leader in your field.

  • Earn high-quality backlinks and mentions from reputable industry sites, publications, or organizations.
  • Pursue speaking engagements, awards, interviews, or collaborations with known experts.
  • Build topical authority through content clusters: Cover a subject comprehensively with interlinked pages.
  • Encourage and showcase positive reviews, testimonials, and social proof (with real names/photos where possible).

Off-page tip: Focus on digital PR, guest posting on authoritative sites, and consistent brand mentions.

4. Establish Trustworthiness (Reliability & Transparency)

This is often the most important element.

  • Be transparent: Clear “About Us” page, contact information, physical address (if applicable), and privacy policy.
  • Cite reputable sources with outbound links to primary studies, government sites, or official documentation. Fact-check everything and disclose methods.
  • Keep content accurate and up-to-date — regularly refresh older articles with dates and “Last Updated” notes.
  • Technical signals: Use HTTPS, fast loading times, secure forms, and good page experience.
  • Respond professionally to reviews and comments; manage reputation across platforms.

Bonus: Add schema markup for authors, reviews, and FAQs to help search engines understand your credibility signals.

Step-by-Step Implementation Roadmap

  1. Audit Your Site (Week 1–2)

Review existing content for weak E-E-A-T signals. Use Google’s Search Quality Rater Guidelines as a reference. Check for missing author info, outdated content, or lack of sources.

  1. Strengthen On-Page Elements (Ongoing)
  • Add author bios + credentials to all new and priority pages.
  • Enhance About Us and contact pages.
  • Implement structured data (Schema.org).
  1. Create/Refresh Content

Prioritize high-impact pages (especially YMYL). Rewrite thin content with real experience, data, and sources. Aim for people-first, helpful content.

  1. Build External Signals

Focus on earning links, reviews, and mentions. Consistency across the web (same bio, branding) helps.

  1. Monitor & Iterate

Track performance in Google Search Console (impressions, clicks, queries). Watch for updates in helpful content systems. Measure user engagement (time on page, bounce rate, conversions).

Quick E-E-A-T Checklist

  • Real author bylines + detailed bios with credentials.
  • Original research, case studies, or firsthand examples.
  • Proper sourcing and fact-checking.
  • Transparent business info and contact details.
  • High-quality backlinks and positive reputation signals.
  • Regularly updated, accurate, well-structured content.
  • Strong technical foundation (speed, security, mobile-friendliness).

E-E-A-T is most powerful when it’s authentic — not gamed. Focus on genuinely helping your audience with expertise you actually possess. Over time, this builds sustainable authority and better performance in both traditional search and AI overviews.

Start with your most important pages and high-traffic content, then scale. If you’re in a competitive or YMYL niche, prioritize this heavily. For the latest official guidance, refer to Google’s “Creating Helpful, Reliable, People-First Content” documentation.

Google's Search Quality Rater Guidelines are a public document (currently ~182 pages, latest major version September 11, 2025) that instructs thousands of human Search Quality Raters on how to evaluate search results.

Raters assess two main things for sample queries:

  • Page Quality (PQ) — How well a page achieves its purpose, focusing on quality, reliability, and helpfulness.
  • Needs Met (NM) — How well the result satisfies the user's intent (Fully Meets, Highly Meets, Moderately Meets, Slightly Meets, Fails to Meet).

Important disclaimer: These ratings do not directly rank pages. They help Google test and improve its algorithms by providing human feedback on what constitutes good results. No single rating affects any specific page.

Core Concepts in the Guidelines

#### 1. Page Quality Rating Raters evaluate based on: - Purpose of the page — What is it trying to do (inform, sell, entertain, etc.)? - Main Content (MC) — The primary reason users visit the page. High-quality MC requires significant effort, originality, talent/skill, and value. - Supplementary Content (SC) and Ads — These should not distract from or overwhelm the MC. - Website reputation and content creator reputation — Checked via reviews, external sources, etc.

#### 2. E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) This is a key framework for assessing Page Quality, especially for YMYL (Your Money or Your Life) topics like health, finance, legal, safety, or topics impacting well-being.

  • Experience: First-hand or practical knowledge (e.g., personal use, case studies).
  • Expertise: Demonstrated knowledge and skill.
  • Authoritativeness: Recognized as a reliable source in the field.
  • Trustworthiness: Accuracy, honesty, safety, transparency, and reliability. Trust is the most important aspect of E-E-A-T.

Raters look for clear authorship, sourcing, "About Us"/contact info, and evidence of the above.

#### 3. YMYL Topics Pages on topics that could significantly impact a person's health, financial stability, safety, or society require much higher standards of E-E-A-T. Low E-E-A-T here can lead to Lowest Quality ratings.

#### 4. Quality Levels (Lowest → Highest) - Lowest Quality: Harmful, misleading, untrustworthy, spammy, deceptive, low/no effort content, scaled content abuse, expired domain abuse, etc. - Low Quality: Lacking E-E-A-T, filler content, distracting ads, low effort/originality. - Medium Quality: Average, satisfies basic needs but not exceptional. - High / Highest Quality: Excellent MC, strong positive reputation, very high E-E-A-T, demonstrates clear value and expertise.

#### 5. Needs Met Rating Focuses on how well the result matches user intent: - Fully Meets (FullyM): Perfectly satisfies the query. - Highly Meets (HM), Moderately Meets (MM), etc., down to Fails to Meet (FailsM).

Raters consider query intent (Know, Do, Website, Visit-in-person, etc.), locale, freshness, and multiple possible meanings.

Key Takeaways for Website Owners & SEO (2026)

The guidelines reinforce people-first content: - Prioritize genuine helpfulness over search manipulation. - Demonstrate E-E-A-T clearly (bylines with credentials, sources, transparency, original insights/experience). - Avoid low-effort, AI-generated filler, scaled content abuse, or deceptive practices. - High Page Quality + strong Needs Met = better performance in Google's systems. - AI content is acceptable if it adds real value and follows the same quality standards.

Google's "Creating Helpful, Reliable, People-First Content" page summarizes these ideas for creators and explicitly recommends reviewing the rater guidelines for self-assessment.

Where to Read the Full Guidelines

Reading the guidelines (or at least key sections on E-E-A-T, Page Quality, and examples) is one of the best ways to align your content strategy with what Google values. Focus on creating content that real users find exceptionally helpful and trustworthy — the algorithms are trained to reward that. Google's Helpful Content System (now integrated into core ranking systems) prioritizes original, helpful content created primarily for people rather than for manipulating search rankings.

History and Current Status (as of 2026)

  • Launched as the Helpful Content Update in August 2022.
  • In March 2024, Google integrated it into its core ranking systems. It is no longer a standalone periodic update but uses a variety of ongoing signals.
  • The goal remains the same: Reward people-first content and reduce visibility of low-quality, unhelpful, or "search engine-first" content (e.g., mass-produced, AI-generated filler, or content made just to rank).

It operates at both page level and site-wide signals. A site with a lot of unhelpful content can see broad ranking impacts.

What Google Considers "Helpful Content"

Google's systems evaluate content based on questions like these (self-assessment recommended from their official guide):

#### Content Quality Questions - Does it provide original information, reporting, research, or analysis? - Is it a substantial, complete, and comprehensive description of the topic? - Does it offer insightful analysis or information beyond the obvious? - Does it add substantial additional value instead of just summarizing or rewriting others? - Would you bookmark, share, or recommend this page? - Is it well-produced, free of sloppy errors, and comparable to high-quality printed material?

#### People-First vs. Search Engine-First People-first content (rewarded): - Created for an existing or intended audience that would find it useful directly. - Demonstrates first-hand expertise and depth of knowledge. - Leaves readers feeling they've learned enough and had a satisfying experience. - Has a clear primary purpose or focus for the site.

Search engine-first content (demoted): - Primarily made to attract search traffic. - Mass-produced on many topics without real expertise. - Heavy use of automation (including AI) without adding unique value. - Keyword-stuffed, thin, or trending-chasing content. - Leaves readers needing to search elsewhere for better info.

Connection to E-E-A-T

The system heavily incorporates E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness), with Trust being the most important. It applies stricter standards to YMYL topics (health, finance, safety, etc.).

AI-Generated or AI-Assisted Content

AI content is not banned if it is helpful, original, and follows the same quality standards. Google recommends transparency (e.g., disclosures) when automation plays a significant role, especially if readers would expect to know "how" it was created.

Practical Advice for Success

  1. Audit your content — Use Google's self-assessment questions. Remove or improve thin/unhelpful pages.
  2. Focus on depth and originality — Add real insights, data, case studies, or firsthand experience.
  3. Demonstrate E-E-A-T — Use clear bylines, author bios with credentials, sourcing, and "About Us" transparency.
  4. Improve page experience — Fast loading, mobile-friendly, intuitive design.
  5. Monitor performance — Use Google Search Console to track impressions/traffic. Recoveries can take time (often aligning with core updates).
  6. Avoid shortcuts — No bulk publishing, date manipulation, or low-effort automation solely for rankings.

The Helpful Content principles are now a foundational part of Google's core algorithm alongside other systems (e.g., spam detection, freshness, and reviews systems). Sites that consistently create high-value content for real users tend to perform best long-term.

Official Resources: - Creating Helpful, Reliable, People-First Content — The best starting point. - Guide to Google Search Ranking Systems - Search Quality Rater Guidelines (for deeper E-E-A-T understanding).

Focus on genuinely helping your audience, and the algorithms are designed to reward that.

Google's SpamBrain is an AI-powered (machine learning) spam detection and prevention system designed to identify and neutralize spam, manipulative tactics, low-quality content, and deceptive practices in search results.

Key Facts About SpamBrain

  • Launched: 2018 (Google publicly acknowledged it in 2022).
  • Core Technology: Self-learning AI/machine learning that continuously adapts to new and evolving spam techniques. It analyzes patterns at massive scale during crawling and indexing.
  • Capability: Detects spam so effectively that many spammy pages are prevented from being indexed at all (they never appear in results).
  • Impact: Google reported that SpamBrain helped make 99% of visits from Search spam-free in certain periods, and it detects significantly more spam over time (e.g., 5x more in 2022 vs. 2021).

What SpamBrain Detects

SpamBrain targets a wide range of violations of Google's spam policies:

  • Link spam: Manipulative or artificial links (buying/selling links, link farms, sites created primarily to pass link value). It can neutralize the impact of unnatural links rather than just penalizing sites.
  • Scaled content abuse: Mass-produced, low-effort, or AI-generated content without meaningful value or human oversight.
  • Other tactics: Keyword stuffing, cloaking (showing different content to users vs. Google), hidden text, parasite SEO, site reputation abuse, and various deceptive practices.

It works alongside other systems and is frequently improved. Spam updates (periodic rollouts) often reflect enhancements to SpamBrain.

How SpamBrain Fits Into Google's Ecosystem (2026)

  • It is always running and improved continuously — not just during named updates.
  • Spam updates (e.g., March 2026 spam update, which rolled out extremely quickly) are often refinements to SpamBrain for better enforcement of existing policies.
  • It complements other systems like the Helpful Content System, core ranking algorithms, and E-E-A-T evaluations. Low-quality or spammy content is demoted or removed, while high-quality, people-first content is rewarded.

Implications for SEO and Website Owners

  • Focus on quality: Create original, helpful content with genuine E-E-A-T. Avoid automation-only or scaled production of thin content.
  • Avoid manipulation: No paid links for ranking, keyword stuffing, cloaking, or schemes designed purely to game rankings.
  • Recovery: If hit by SpamBrain (traffic drop during a spam update), fix the issues thoroughly and wait for the next refresh or core update. Recoveries can take time.
  • Best practice: Follow Google's official spam policies strictly. Legitimate sites with strong, user-focused content are generally safe.

Official Resources

SpamBrain represents Google's shift toward scalable, intelligent spam fighting using AI. The safest long-term strategy remains creating genuinely helpful, trustworthy content that serves users — exactly what the Helpful Content and E-E-A-T systems also reward.

Google's SpamBrain is a proprietary, AI-driven machine learning platform (not a single model) launched in 2018 and continuously refined. Google has not released detailed technical papers, model architectures, or code, so public knowledge relies on official blog posts, patents on related spam detection, expert analyses, and observed behaviors.

Core Machine Learning Approach

SpamBrain is a self-learning / adaptive system that improves over time by analyzing vast datasets of web pages, links, content, and user signals. Key characteristics:

  • Supervised & Semi-Supervised Learning: Trained on large labeled datasets of known spam vs. legitimate examples (e.g., organic links vs. manufactured ones). It learns patterns and generalizes to detect new variants.
  • Pattern Recognition at Scale: Excels at identifying complex, multi-dimensional patterns across billions of pages that rule-based systems miss. This includes subtle correlations in link networks, content quality, timing, and site behavior.
  • Deep Learning Techniques: Later iterations incorporate advanced deep learning (e.g., neural networks) for nuanced detection, including subtle spam patterns and low-value AI-generated content. Natural Language Processing (NLP) helps analyze content authenticity, context, and semantics.
  • Probabilistic Scoring: Assigns confidence scores rather than binary yes/no decisions, allowing nuanced actions like neutralizing specific signals instead of full penalties.
  • Continuous / Online Learning: Updates dynamically as new spam emerges, making it robust against evolving tactics.

Key Detection Areas & Likely Techniques

  1. Link Spam Detection (Primary Strength):
  • Analyzes link networks and graph structures (e.g., cross-linking patterns, shared ownership signals via technical footprints like IPs/DNS, timing correlations).
  • Distinguishes natural editorial links from artificial ones using features like anchor text distribution, contextual relevance, velocity of acquisition, and site purpose.
  • Neutralization over Penalization: Often devalues manipulative links without demoting the entire site.
  1. Content & Scaled Abuse:
  • Detects thin, templated, spun, or mass-produced content (including low-value AI-generated material).
  • Uses NLP for semantic analysis, topic coherence, and quality signals (e.g., originality, depth, engagement proxies).
  1. Other Signals:
  • Site reputation, user behavior patterns, cloaking, keyword stuffing, and broader manipulative behaviors.
  • Integrates with other systems (e.g., Helpful Content, core ranking) for holistic evaluation.

How It Likely Works (Inferred Pipeline)

  • Data Ingestion: Crawls and processes signals during indexing (prevents many spam pages from entering the index entirely).
  • Feature Engineering: Extracts hundreds/thousands of features (link graphs, content embeddings, behavioral metrics, technical fingerprints).
  • Model Ensemble: Multiple specialized ML models running in parallel as a "platform," with outputs combined for final decisions.
  • Action: Block from index, neutralize signals (e.g., ignore links), or demote visibility.

This is similar to other Google ML systems (e.g., those using embeddings, graph neural networks, or transformers for pattern detection), but customized for spam.

Limitations in Public Knowledge

Google keeps specifics confidential to prevent gaming. Older patents exist on link-based spam detection and content analysis, but none are confirmed as direct SpamBrain implementations.

Implications for SEO

  • Focus on Authenticity: Build genuine value, natural links, and high-quality content. Avoid patterns (e.g., unnatural velocity, footprints, thin sites).
  • Long-Term Resilience: SpamBrain rewards sustainable practices and punishes shortcuts, especially at scale.

For the most authoritative info, refer to Google's Web Spam reports and spam policies. SpamBrain evolves rapidly—staying aligned with E-E-A-T and Helpful Content principles remains the best defense.

Graph Neural Networks (GNNs) are a powerful class of machine learning models designed specifically for graph-structured data, where entities (nodes) are connected by relationships (edges). They excel at spam detection because spam often involves coordinated networks, unnatural link patterns, clusters of fake accounts/reviews, or anomalous behaviors that are hard to spot in isolation.

Why GNNs Are Effective for Spam Detection

Traditional ML models treat data points independently, but spam frequently relies on relational patterns: - Link spam: Networks of sites linking to each other artificially (link farms, PBNs). - Review spam / fake accounts: Groups of coordinated entities behaving similarly. - Botnets / phishing: Connected infrastructure with shared footprints. - Content abuse: Clusters of low-quality, mass-produced pages.

GNNs propagate information across these connections via message-passing, allowing nodes to learn representations based on their own features and their neighbors (and neighbors' neighbors). This captures multi-hop relationships and structural anomalies effectively.

How GNNs Work in Spam Contexts (High-Level)

  1. Graph Construction:
  • Nodes: Web pages, domains, accounts, reviews, IPs, etc.
  • Edges: Links, shared hosting, co-occurrences, temporal correlations, user interactions.
  1. Node/Edge Features:
  • Content embeddings (from NLP models), metadata (age, traffic), behavioral signals.
  1. Message Passing & Aggregation:
  • Nodes aggregate information from neighbors (e.g., via Graph Convolutional Networks — GCN, Graph Attention Networks — GAT).
  • Multiple layers allow learning higher-order patterns (e.g., "this site links to many low-trust sites that also link to each other").
  1. Prediction:
  • Node classification (spam/not spam), edge prediction (suspicious link?), or graph-level anomaly detection.

Common variants include Graph Attention Networks (GAT) for focusing on important neighbors and heterogeneous GNNs for different node/edge types.

Applications in Spam & Fraud Detection

  • Web/Link Spam: Early research applied GNNs to datasets like WEBSPAM-UK2006 for classifying spammy pages using link graphs.
  • Review Spam: Models like GCN-based Anti-Spam (GAS) detect fake reviews on platforms by modeling reviewer-product graphs.
  • Social Bots / Email Spam: Attention-based GNNs for bot detection in networks; Semantic GNNs for email classification.
  • Fraud/Anomaly Detection: Widely used in finance (transaction graphs), e-commerce, and cybersecurity (malicious domain detection via GNNs on infrastructure graphs).
  • LLM-Generated Spam: Hybrid models combining language model embeddings with GNNs to detect coordinated fake content.

Google's Role: Graph Foundation Models (GFMs)

Google has heavily invested in graph ML (e.g., TensorFlow GNN library). In 2025, they introduced Graph Foundation Models (GFMs) — an evolution beyond traditional GNNs.

  • Key Advantage: Traditional GNNs are often tied to a specific trained graph and don't generalize well to new graphs. GFMs are more like "foundation models" for graphs — pre-trained to work across diverse, unseen relational data (e.g., turning multiple database tables into a unified heterogeneous graph).
  • Spam Application: Tested at scale on spam detection in Google Ads, achieving 3x to 40x gains in precision over strong tabular baselines by leveraging connections across dozens of large graphs.
  • This is highly relevant to SpamBrain, which handles massive-scale link networks, content clusters, and relational signals. While Google doesn't publicly detail SpamBrain's exact internals, graph-based techniques (including advanced GNN/GFM approaches) align perfectly with its ability to detect and neutralize coordinated manipulative patterns.

Implications for SpamBrain & SEO

SpamBrain's machine learning likely incorporates graph-based methods (or similar relational learning) to: - Identify unnatural link velocity, topology, and shared footprints. - Neutralize manipulative signals without full-site penalties. - Scale to billions of pages while adapting to new tactics.

For SEOs: Avoid patterns that create detectable graph anomalies (e.g., reciprocal linking schemes, private blog networks with similar footprints, scaled low-value sites). Focus on natural, earned relationships and genuine content — these are harder for graph models to flag as outliers.

GNNs/GFMs represent a shift toward structural intelligence in spam fighting, making isolated tactics less effective and rewarding authentic, user-focused ecosystems. Google's public research (e.g., TF-GNN and GFMs) gives insight into the direction of their systems.

For deeper dives, check Google's Research blog on Graph Foundation Models and academic resources on GNNs for fraud/spam.