Before we get into the framework, here’s the number that should reset how you think about this whole topic. Wellows.com’s analysis of 15,847 AI Overview results found that 47% of AI Overview citations come from pages ranked below position #5. Not the pages ranking highest. Not the most-linked domains. Pages that, by traditional SEO metrics, shouldn’t be winning at all.

That single data point tells you the game changed. Most teams haven’t updated their playbook.

The Three-Layer Evaluation Model Google Uses in 2026

Google doesn’t evaluate content through a single filter. There are three distinct layers, each with its own criteria, its own evaluation cycle, and its own gate. Failing any one of them produces different consequences. The mistake most content teams make is optimising for Layer 1 signals while ignoring the mechanics of Layer 3, which is where the AI-era visibility actually lives.

Layer 1 — Quality Rater Guidelines: The Human Standard That Trains the Machine

The QRG is 182 pages as of the September 11, 2025, update. It’s public, downloadable from guidelines.raterhub.com, and almost nobody in content teams has read it. Which is remarkable, because it’s a detailed specification of what your audience’s satisfaction looks like to the people training Google’s systems.

Raters score every page on two scales. Page Quality has five levels: Lowest, Low, Medium, High, and Highest. The gap between High and Highest isn’t minor — it’s the line between “good content” and content that demonstrates genuine first-hand experience plus original analysis that couldn’t be replicated by aggregating existing sources.

The second scale — Needs Met — has six levels. The one that matters most for AI Overview selection is this: content rated “Moderately Meets” will not be selected for AI Overviews, regardless of any other quality signals. “Highly Meets” is the floor. That’s an editorial standard, not a technical one. You can’t schema your way past it.

The September 11, 2025, QRG Update — What Actually Changed

Google called it a “minor update.” The PDF grew by one page. Two additions actually matter for content strategy.

First: YMYL expanded to include civic content. The September 2025 edition added a new sub-category — “YMYL Government, Civics & Society” — explicitly covering elections, voting information, trust in public institutions, and civic content. Any page touching these topics now carries the same high-scrutiny standards as medical, financial, and legal content. If you run a news or policy site and didn’t know this, your content may be getting evaluated against a much higher bar than you assumed.

Second: AI Overview responses can now be rated directly. For the first time in QRG history, the document added explicit criteria for evaluating AI Overview summaries — not just the web pages behind them. Raters can score AI-generated summaries against the same quality rubric as source content. The quality bar for AI Overview citation isn’t lower than the bar for ranking. It’s the same bar applied twice.

Layer 2 — Core Update Signals: Verified Outcomes from December 2025 and February 2026

December 2025 Core Update (Completed December 29)

SE Ranking’s SERP analysis found that 15% of pages previously ranking in the top 10 disappeared from the top 100 entirely after this update. ALM Corp’s analysis of 847 sites across 23 industries broke the outcome data into three content types. The numbers below are directional — ALM Corp is a named agency, not an independent academic audit — but the directional signal is clear enough to act on.

Source: ALM Corp analysis of 847 sites across 23 industries (January 2026); SE Ranking SERP analysis (December 2025). These are named sources with disclosed methodologies — not peer-reviewed. Use as strongly suggestive, not proven causal.

February 2026 Discover Core Update — The Historic First

Google’s February 5–27, 2026, update is genuinely novel. In 22 years of documented algorithm updates, Google had never issued a core update targeting Discover exclusively. Not Search. Not the full index. Discover only. If your Search Console traffic held steady while Discover traffic dropped in that window, this is why. The two surfaces are now evaluated and updated on separate tracks.

Google’s Search Central Blog stated three specific goals for the update: more locally relevant content from country-based sites; reduced sensational and clickbait content; and more in-depth, original, and timely content from sites with demonstrated topic expertise — evaluated topic by topic, not at the domain level.

Layer 3 — AI Overview Selection: The Citation Battleground at the Top of the SERP

AI Overviews now appear in over 60% of all searches. Pages cited earn 35% more organic clicks than non-cited pages with equivalent rankings (Search Engine Land, 2025). But here’s the catch worth understanding clearly: AI Overviews also reduce click-through rates by 34.5% on average for searches where they appear (Ahrefs, cited by eMarketer).

The math means you want to be the cited source. Being ranked but not cited is increasingly the worst position — you’re in a search landscape where clicks have declined overall, and you’re getting none of the citation benefit that partially compensates for it.

The Seven AI Overview Citation Factors — Ranked by Strength

Source: Wellows.com analysis of 15,847 AI Overview results across 63 industries. Industry study — not peer-reviewed. Treat all figures as directional. r = Pearson correlation coefficient between factor score and AI Overview selection rate.

The DA / Entity Density Paradox — A Cross-Source Synthesis

Here’s the thing no single source will tell you, because it requires combining two separate datasets.

E-E-A-T in 2026 — What Has Mechanically Changed

You’ve read the definitions. Experience, Expertise, Authoritativeness, Trustworthiness. Every SEO article from 2022 has them as four bullet points. Skip that. What matters now is what Google’s systems read as proxies for each — because the QRG tells raters exactly what structural signals to look for, and those signals tell you what to build.

Experience: The Hardest Pillar to Fake Algorithmically

The QRG’s Experience evaluation looks for specific structural signals, not claimed credentials. Three that raters are explicitly trained to identify.

Temporal specificity in outcomes. “We changed the CTA from ‘Learn More’ to ‘Get Your Free Quote’ in Q3 2025 and saw a 34% lift in form completions” reads as experience. “Changing your CTA can improve conversions” reads as aggregated knowledge. The specific quarter, the specific delta, the named actor — these are structural signals that a human rater recognises as first-hand. And increasingly, an algorithm trained on those ratings picks them up too.

Named scenarios only an insider would know. The specific failure mode of a tool that only someone who ran it in production would encounter. The edge case that contradicts the official documentation. The counterintuitive finding from a specific dataset. These aren’t things you can source from other articles — which is the point.

Exaggerated credentials are a Low rating trigger. The QRG states explicitly: “If you find the information about the website or the content creator to be exaggerated or mildly misleading, the Low rating should be used.” Claiming expertise you can’t substantiate isn’t neutral — it’s an active downgrade trigger. Author schema matters not because the schema itself signals quality, but because it gives raters a verifiable path to confirm the claims.

Trustworthiness: The Multiplier on Everything Else

Trust isn’t one signal — it’s the multiplier on all the others. A page with high Experience signals and low Trust signals scores lower than a page with moderate Experience and high Trust. The QRG’s trust evaluation runs through four channels: the page’s main content, the site’s off-page reputation, verifiable credentials disclosed on the page, and technical trustworthiness (security, editorial standards, correction policies).

The practical trust-building stack for 2026: named author with a Google-indexable bio page linked via Article schema; at least five external sources cited with named attribution, at least two being primary sources; a visible correction or update policy; HTTPS and Core Web Vitals at passing thresholds.

The part most teams skip is the author entity — specifically, the off-site citation pattern, mentions in authoritative sources, and entity co-occurrence in the Knowledge Graph. Schema is not trusted. It’s a pointer toward verifiable trust signals that exist elsewhere.

AI-Generated Content and Google’s 2026 Position — The Honest Picture

The December 2025 data settled this debate for anyone still having it. The problem was never AI tools. It was the absence of human editorial judgment. Google doesn’t penalize assistance. It penalizes absence of effort, originality, and added value — which the QRG defines precisely in three sections.

The Three QRG Failure Modes — Exact Section Language

Sections §4.6.3 through §4.6.6 of the September 2025 QRG identify specific abuse types. Three are directly relevant to AI content strategy.

§4.6.4 — Site Reputation Abuse. Third-party content is published on a high-authority host to exploit its established ranking signals. The risk isn’t just a spam penalty — it’s that the content misleads users into believing editorial standards were applied when they weren’t. Publishers hosting AI-generated partner content without review face this exposure specifically.

§4.6.5 — Scaled Content Abuse. “Many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value for website visitors.” Scale alone is not the issue. Scale without originality or user value is.

§4.6.6 — Low-Effort MC. Google’s exact language: “The Lowest rating applies if all or almost all of the MC on the page is copied, paraphrased, embedded, auto or AI generated, or reposted from other sources with little to no effort, little to no originality, and little to no added value for visitors to the website. Such pages should be rated Lowest, even if the page assigns credit for the content to another source.”

---

Pre-Publication Quality Gate — 12 Checkpoints Verified Against 2026 Standards

Every checkpoint below has a specific threshold. Not “ensure content is high quality” — specific enough that an editor who just produced the wrong thing can point to the exact mistake.

Content-Level Checks

• 01
  Passage architecture Each topical passage is 127–167 words (target 140–155) and answerable standalone — a reader who sees only that passage gets a complete answer to one specific question. This is your AI Overview citation unit.
• 02
  Entity density: count to 15 The page contains 15+ named, Google Knowledge Graph-recognized entities (people, organizations, concepts, places), each in meaningful context. Count them. Not keywords — entities. If you’re below 15, you have a citation gap that no amount of schema will fix.
• 03
  Needs Met target: declare it before writing State the Needs Met target before the brief is written: “Fully Meets” or “Highly Meets.” If you can’t write one sentence naming the specific query this article fully answers, the brief isn’t ready.
• 04
  Originality gate (QRG §4.6.6) At least one named, verifiable example with a specific outcome, date, and named actor. This is what satisfies the “effort and originality” requirement. Aggregated advice without named examples fails it.
• 05
  AI Overview structure: 5 citation-ready passages At least 5 passages structured as: question → direct answer → supporting evidence. Apply FAQ schema to each. These are your citation candidates for AI Overviews.
• 06
  Source quality: ≥5 external, ≥2 primary At least 5 external sources cited with named attribution. At least 2 are primary sources (government documents, official research, platform announcements) — not other blogs summarizing those sources.

Technical-Level Checks

• 07
  Core Web Vitals — current thresholds (not 2023 thresholds) LCP <3.0s. INP <200ms — INP replaced FID as a Core Web Vitals metric in March 2024. If your audit stack still checks FID, it’s two years out of date. CLS <0.1. Verify in field data via CrUX in Search Console, not only Lighthouse lab scores.
• 08
  Schema coverage: four types, validated before publishing Article schema (name, author, datePublished, dateModified, publisher). FAQ schema on all H3/H4 question-format sections. BreadcrumbList. SpeakableSpecification on definitional paragraphs. Validate via Google’s Rich Results Test before publishing.
• 09
  Author entity: indexable bio + external sameAs links Named author has a Google-indexable bio page. Article schema includes an author entity with credentialOf or sameAs links to external profiles. Schema is a pointer — not the trust signal itself.
• 10
  Multi-modal: original diagram + adversarial data table At minimum: one original diagram and one data table with an adversarial column (a column that states why the evidence might not apply to the reader’s specific context). Alt text on all images — accessibility and crawl signal both.
• 11
  Vintage compliance: argue continued validity for data older than 3 years Any data older than 3 years needs its continued validity explicitly argued — state both the original publication year and the most recent update year, then explain why the underlying mechanism hasn’t changed enough to invalidate the figure.
• 12
  §4.6.6 self-test: apply the Lowest rating definition to your own work Read the Lowest rating definition from QRG §4.6.6. Apply it to your article. If any section could be characterised as “little effort, little originality, little added value” — rewrite before publishing, not after.

---

What This Means for Your Role

---

Sources and Confidence Levels

Every figure in this article has a named source and a disclosed confidence level. “Directional” means the source is credible but not independently audited — use the figures as strongly suggestive, not proven causal. “Verified” means primary source confirmed at a canonical URL.