AI Content + Google Penalties

01The 60-second TL;DR

02The truth Google never wanted public

Five things Google publicly denied for years. The May 2024 leak + DOJ trial confirmed all five.

The signals that actually run the show

03What actually gets penalized

Not "AI content." These 10 patterns. AI content tends to produce all 10 simultaneously, which is why it looks like an AI penalty. It isn't.

1. Programmatic / templated content at scale with thin per-page differentiation (the 50,000-hotel-pages-only-city-changes pattern; 98% deindex rate)
2. Zero human editorial review on AI output (raw GPT prose published as is)
3. High DA + low Brand Authority mismatch (Tom Capper's finding: HCU losers averaged BA 37, winners + neutral averaged 50-52)
4. Intrusive ads disrupting UX (auto-play video, fixed footer, multiple interstitials, mandatory video-before-content). Accumulates negative NavBoost over the 13-month window. December 2025 update specifically punished this.
5. Lack of identifiable authorship (anonymous bylines, "Staff Writer," AI-generated author personas with fake bios. Sports Illustrated case.)
6. Artificial freshening (updating bylineDate without updating content. The three-layer date system, bylineDate + syntacticDate + semanticDate, catches it.)
7. Stock images and scraped social images (Cyrus Shepard correlation: stock has "surprisingly strong negative impact")
8. Excessive affiliate-link density with no genuine product testing (CNN Underscored, Forbes Advisor, WSJ Buy Side commerce sections wiped November 2024 under Site Reputation Abuse)
9. Thin YMYL content without verifiable expert authorship (medical, financial, legal, increasingly relationship + parenting + nutrition)
10. Competitive differentiation failure (accurate, comprehensive, but adds nothing 10 other pages don't already cover. Zero information gain.)

04What survives (the HouseFresh story)

The biggest finding in the entire research: most HCU recovery sites recovered without changing their content.

HouseFresh got hit September 2023 HCU. They lost the bulk of their traffic. They built YouTube content and pursued high-profile collaborations. They did almost nothing to the existing articles. August 2024 core update: they exceeded their pre-HCU peak. Tom Capper's interpretation: their Brand Authority score caught up to their Domain Authority. The DA:BA mismatch resolved. The penalty lifted. Source: Glenn Gabe (G-Squared) tracked 400+ HCU sites. Lily Ray confirmed pattern at Amsive.

The patterns sites that survived (or recovered) share:

1. Genuine first-hand experience. NapLab (mattress affiliate, 6,200 → 132,000 monthly visitors) tested every product. Original photography. Methodology transparency. Quantitative scoring systems.
2. Identifiable credentialed authors with LinkedIn links + KG entity association.
3. Real brand presence. Organic search volume for the brand name. Reddit/Quora/LinkedIn citations. YouTube footprint.
4. Low ad density + clean UX. Positive NavBoost over the 13-month window.
5. Moderate publishing velocity. Not 50 articles/day from a content factory.
6. Original data or proprietary research (the only thing AI cannot generate from existing web content; the only path to positive information gain).
7. Direct navigation drivers (newsletter, app, community, free tools). Builds Chrome signal that exists outside Google's search funnel.
8. Off-site community engagement (real Reddit threads, forum discussions, mentioned in industry publications).

05The high-ranking page framework

Page-level rules. The exact "how many external links, what's the word count" answer.

Word count

• Informational pillar: 1,800 to 3,500 words.
• Commercial / "best of" listicle: 2,500 to 5,000 words.
• Product / category page: 600 to 1,200 words of unique copy plus product grid.
• Blog supporting article: 1,200 to 2,000 words.
• Hard rule: length follows topic depth, not the other way around. Top-3 SERP average is the floor, not the target. Padding is a penalty risk (Surfer found cited articles average 2,900+ words but the cause is depth, not length).

Headings (H1 through H4)

• One H1. Contains primary keyword. Maximum 60 characters.
• Five to ten H2s. Each maps to a real subtopic or "People Also Ask" query. Each H2 functions as an independently-retrievable passage for AI search.
• H3s nest under H2s for sub-questions.
• H4 only when genuinely needed.
• Headings should read like a table of contents that answers the search intent end to end.

Internal links

• 8 to 15 internal links per 2,000 words (sweet spot: roughly 1 per 200 words).
• Link to: the pillar page, sibling cluster pages, deeper supporting content.
• Anchor text: descriptive and varied. Never "click here." Never the same anchor twice on one page.
• Every new page should be linked FROM at least 3 existing pages within 7 days. Orphan pages die. Bi-directional linking increases AI citation probability by 2.7x (Yext 2025).

External links

• 2 to 5 external links per article to high-authority sources (.gov, .edu, original studies, Wikipedia for entity grounding, brand homepages).
• All target="_blank" rel="noopener". Don't nofollow legitimate citations. Outbound trust signals matter.
• Cite sources inline with the claim, not in a generic footer.
• Never link to direct competitors on commercial pages.

Images

• 1 hero image above the fold (original beats stock 10x).
• 1 image per 300 to 500 words thereafter.
• Original photos, screenshots, custom diagrams beat stock.
• Every image: descriptive alt, descriptive filename, <figcaption> when adding context, lazy-load below the fold, WebP/AVIF, under 200 KB.
• Add at least 1 original chart or comparison table. Gets cited in AI Overviews.

Schema markup (mandatory)

• Article or BlogPosting with author, datePublished, dateModified, publisher.
• Person schema on author with sameAs linking to LinkedIn / X (entity verification for Google Knowledge Graph).
• BreadcrumbList.
• FAQPage if you have a real FAQ section. Highest AI-citation probability of any schema type.
• HowTo if step-based.
• Organization site-wide.
• Review / AggregateRating only if real.

E-E-A-T signals (heaviest weight in 2026)

• Author bio block under H1: photo, credentials, 2-3 sentence bio, links to LinkedIn + X + author archive.
• "Reviewed by [credentialed expert]" line on YMYL topics.
• "Last updated [date]" visible. Refresh content quarterly with real updates (not just date stamps).
• First-person experience injection: at least one paragraph of "I tested / I built / I measured" with a real artifact (screenshot, photo, data).
• Original data or research: even one stat you generated yourself moves the needle.

On-page mechanics

• Title tag: primary keyword in first 60 characters. Brand at end. Format: Primary Keyword | Subtitle: Brand.
• Meta description: 150-160 characters. Primary + secondary keyword + CTA.
• URL slug: 3-5 words, hyphenated, primary keyword, no dates, no stop words. /best-voice-ai-agents/ not /2026/05/01/the-10-best-voice-ai-agents/.
• Primary keyword: in H1, first 100 words, one H2, one image alt, URL, meta. Density ~0.5 to 1.5%. Never stuff.
• Semantic / entity coverage: use Surfer / Frase / Clearscope to ensure the page covers the full entity set the top 10 SERP covers. Biggest lever for ranking.
• TOC for any page over 1,500 words. Anchor links boost AI Overview citation.

Engagement (NavBoost loves these)

• Hook in first 50 words that satisfies the query directly.
• Bold key sentences for skimmers.
• Bullet lists every 300-500 words.
• One embedded video or interactive widget (drives dwell time).
• "Jump to" TOC and back-to-top button.
• Comments enabled (engagement signal).
• Related posts grid at bottom (drives session depth).

Technical layer

• Core Web Vitals all green: LCP under 2.5s, INP under 200ms, CLS under 0.1.
• Mobile-first: every element 16px font minimum, tap targets 48px minimum.
• HTTPS, valid SSL, HTTP/2.
• canonical tag on every page.
• hreflang if multi-language.
• Clean robots.txt, sitemap.xml submitted to Google Search Console.
• No render-blocking JS for above-the-fold content.

What to NEVER ship

• Native <select> dropdowns on customer-facing pages.
• Stock photo as the only image.
• "In conclusion" / "In summary" / "It's not just X, it's Y" cadence (GPT slop).
• Author = "Admin" or no author at all.
• Dates with no actual update to the body when you bump them.
• Affiliate links above the fold without disclosure.
• Programmatic templates with thin / repeated / spun content.
• More than 3 ads above the fold.

06The 8-layer hybrid pipeline

The production SOP. Hybrid content ranks 34% higher on average than unedited AI content (2025 SEO analysis). Pure AI hits Google top 10 in 28% of cases but only 3% reach the top 3.

1. AI draft. GPT-5.4 / Claude Opus 4.7 / Gemini 3.1 Pro. Persona prompt with credentials. Target keyword cluster. Brief with desired headings. Competing URLs in context. Fact-check list of claims to verify. Output: 800-2,000 word raw draft. Do NOT publish.
2. SME review. Subject-matter expert reads for factual accuracy, not prose quality. Upwork/Contra freelancer at $25-75/article, or in-house SME for core topics. Non-negotiable for YMYL.
3. First-hand experience injection. 2-4 spots per article: "When we tested X, we found Y specific result with numbers." Original screenshots. Original photography. Replace generic advice with specific metrics. 3-5 injected experience signals per article.
4. Fact-check layer. Verify every statistic against primary source. Flag any claim older than 18 months as stale. Inline citations with hyperlinks. Use Perplexity for verification, Google Scholar for academic, .gov/.edu for regulatory.
5. Original asset addition. At least one original asset per article: chart with real data, infographic, screenshot, video embed. The AI-content-cannot-replicate moat.
6. Humanizer pass (optional). Only if you have compliance contracts requiring AI-undetectable output. For pure SEO publishing: skip. Spend the budget on more SME review time.
7. E-E-A-T markup. Structured author bio. "Reviewed by" attribution. "Last Updated" date. Article + Person + Organization + FAQPage schema.
8. Technical polish. Compress + alt-tag images. TOC for 1,500+ word articles. 3-5 internal links. 2-3 external citations. Page loads under 2.5s LCP.

07Optimizing for AI search (GEO / AEO / LLMO)

Where the puck is going. Mike King (iPullRank) calls this Relevance Engineering. The core architectural shift: Google's AI Mode decomposes one user query into 6-12+ synthetic sub-queries (per Google patent US20240289407A1). Content that only answers the head term gets cited once. Content that covers the full sub-query fan-out gets cited 6-10 times per response.

Citation patterns by platform

Universal citation patterns (Surfer 36M study)

• YouTube ~23.3%, Wikipedia ~18.4%, Google.com ~16.8% dominate every vertical.
• Video is the single most-cited content format across every vertical. An article + companion YouTube video doubles citation surface area.
• UGC sites get nearly 2x more citations than brand-owned content.
• For YMYL: verified authorship + primary research citations + visible credentials are decisive.

Content structure that gets extracted

H2: [Question format: "How does X work?"]
[Direct 40-60 word answer. Self-contained. No context required.]

[Supporting explanation: 150-300 words with evidence.]

[Bullet points or numbered list for scannable structure.]

[Inline citation: "According to [Source], [specific stat]."]

Information gain (Google patent US11200288B2)

The patent describes assigning each page an "information gain score" measuring new information above and beyond what the user has already encountered in the current session. The 8th article a user reads about "content marketing" scores near zero. An article with a unique data point, unusual framing, or novel entity relationship scores high. The patent was originally written for "automated assistants and chatbots." The scoring logic is embedded in how AI Overviews selects sources.

The AI content trap: AI-generated content that synthesizes existing web content scores zero information gain by definition. It is a recombination of already-indexed facts. The only path to positive information gain is original data, first-person experience, expert interviews, or genuinely novel framing.

08AI humanizer benchmark

The honest answer to "should we run our content through Undetectable.ai before publishing?"

09The operator camps

The community is in its most fractured state ever. Eight operators, four camps, no consensus.

10The 30-day recovery playbook

For a site hit by HCU, scaled content abuse, or general quality demotion. Recovery timeline: 3-6 months for algorithmic penalties (must wait for next core update). 67 days average for manual actions with reconsideration request.

Days 1-3: Audit and triage

• Export Google Search Console: all pages, clicks, impressions, average position, last 12 months.
• Export GA4: sessions, engagement rate, conversion rate per page.
• Build a spreadsheet. Add a Decision column: Delete / Improve / Keep.
• Segment thresholds: Delete = under 10 impressions AND 0 clicks in 12 months. Improve = impressions exist but CTR under 2% OR position 11-25. Keep = consistent traffic + 2+ minute average session.
• Identify your 10 highest-traffic pages pre-penalty. These are your recovery anchors.

Days 4-7: Content delete pass

• Execute all Delete decisions: 301 redirect to most relevant category page, or noindex if not ready to remove.
• Goal: reduce crawl budget waste, signal to Google that the site has stopped producing thin content.
• Do NOT delete pages with backlinks. 301 redirect those to preserve link equity.

Days 8-14: Author + E-E-A-T overhaul

• Replace "Staff Writer" and byline-less articles with real named authors.
• Write full author bio pages (200+ words, credentials, LinkedIn link, professional photo).
• Add "Last Updated" dates to all improved articles.
• Add "Reviewed By [Credentialed Expert]" to YMYL content.
• Implement Person schema for all authors with sameAs linking to LinkedIn and X.
• Implement Organization schema for the site.

Days 15-21: Top 10 content upgrade

For each of your 10 recovery anchor pages:

• Add original data/research or first-person experience section.
• Add original screenshot or photograph.
• Add FAQPage section (5-7 questions) at the bottom + matching FAQPage schema.
• Add Article + Author + FAQPage schema.
• Update any statistics older than 18 months.
• Add 3-5 internal links to related cluster content.
• Improve title tag: add year, add specific number/stat, make it answer-shaped.
• Improve meta description: add first-hand experience signal ("We tested 12 tools...").

Days 22-28: Technical + brand layer

• Audit Core Web Vitals. Fix pages with LCP above 4 seconds.
• Build or upgrade topical clusters around core content.
• Set up email newsletter (even if 0 subscribers; the form + cadence is an entity signal).
• Publish 1 piece of genuinely original research (a 20-50 respondent survey on Pollfish at $50-200 is enough).
• Submit updated sitemap to Google Search Console.

Days 29-30: Reconsideration + monitoring

• If manual action: submit a reconsideration request via GSC. Document every change. Include links to improved content. Explain what was deleted and why.
• If algorithmic: no reconsideration. Wait for next core update.
• Set up weekly GSC monitoring. Impressions trending up = recovery in progress.

Pruning case studies (verified)

• CNET (2024): Deleted hundreds of thousands of pages. +29% search traffic.
• Anonymous brand (Dec 2025): Removed 600,000+ pages. +30% clicks + impressions, sustained 4 months later.
• Home Science Tools (Inflow): Pruned/improved blog. +64% strategic content revenue.
• 192-post deletion case: 20% of content removed. Traffic increased.

11Tool stack

12Bottom line for our properties

Distilled from ~25,000 words of research into the actions that move the needle.