n under three months after Google’s AI Overviews became a global reality, Sellm’s team reverse-engineered the entire pipeline. As a third party, we’ll walk through their experiment, quote Sellm’s key findings, and show you how to adapt these insights in your own “AI Overview optimization” strategy.

Introduction: Why AI Overviews Are the New SEO Battleground

Sellm opens their original analysis by defining Google’s AI Overviews as “blocks of text that appear at the very top of search engine results pages (SERPs).” Rather than a list of blue links, Google now “synthesizes information from the top-ranking pages to provide users with an immediate, comprehensive answer.” As Sellm points out, this shift has created a new “position zero” that can dramatically alter click patterns and brand recognition:

“Being cited in an AI Overview delivers tremendous brand exposure—users see your information first, often before any clickable results.”

Put simply, it’s no longer enough to rank high organically; you must also be structurally optimized for Google’s generative model to pull snippets from your page.

Sellm’s Reverse-Engineering: Building a DIY RAG Pipeline

Reproducing Google’s Three-Stage Process

Sellm explains that Google’s AI Overviews operate on a Retrieval-Augmented Generation (RAG) framework:

1. Retrieval (R): “Google’s systems select the top N pages (often N = 50) based on classic SEO signals—PageRank, keyword relevance, domain authority, and technical health.”
2. Augmentation (A): The LLM ingests extracted passages via “dense embedding retrieval,” matching query embeddings to content embeddings.
3. Generation (G): A large language model synthesizes a coherent summary, ensuring “factual consistency and fluency.”

To mimic this, Sellm curated a “corpus of 50 pages” on “AI Overview optimization,” including blog posts, how-to guides, and research articles. They then “developed a custom RAG pipeline”—from indexing to embedding matching and fine-tuned summarization—so they could measure exactly how likely each page was to contribute to an AI Overview.

Measuring Success: Inclusion Score & Citation Count

Sellm introduced two proprietary metrics to quantify performance:

• Inclusion Score (0–1): “The probability that at least one passage from a page is retrieved and used in the AI Overview’s final summary.”
• Citation Count: “The average number of times a page’s content is explicitly quoted or paraphrased across multiple runs.”

As Sellm notes, “a page can have a high Inclusion Score without many citations—sometimes it contributes a single sentence frequently—whereas another page may have moderate inclusion but heavy citations when selected.”

Core Discoveries from the Initial Dataset

Sellm’s initial test on the “AI Overview optimization” keyword revealed which human-authored pages the model favored. The top performers looked like this:

From these numbers, Sellm observes:

“None of the human-written pages fully matched the AI Overview ‘blueprint,’ topping out around a 0.70 Inclusion Score and roughly 10 citations—evidence that structure matters as much as traditional SEO.”

Notably, the page with a 0.702 Inclusion Score (“AI for SaaS Analytics”) had almost zero citations, demonstrating how “Inclusion Score and Citation Count are not strongly correlated.”

Decoding the AI Overview “Blueprint”

Sellm distilled a repeatable structure that Google’s generative model favors. In their words:

“Think of this as the inherent ‘blueprint’ an AI Overview tends to use. If your content naturally aligns with this format and addresses the underlying question phrasing, its chances of being extracted and summarized increase significantly.”

The crucial elements of this blueprint are:

1. Exact Question-Based Headings (H2/H3): The LLM “scans for headings like ‘What is AI Overview optimization?’”
2. Concise, Self-Contained Paragraphs: Under each heading, “2–3 sentence answers make it trivial for the model to extract ‘chunks.’”
3. Nested Subheadings for Granular Citations: By using H4 headings under H3s (e.g., “What Is Inclusion Score?”), you signal additional “answer units,” multiplying citation opportunities.

Sellm found that pages strictly adhering to this format “outperformed those with strong backlink profiles but free-form content,” underscoring that once you rank in Google’s top 50, structure dictates whether you’re cited.

The Breakthrough: Sellm’s AI-Generated, Optimized Post

To prove the blueprint’s potency, Sellm crafted a completely new article—generated by an AI but meticulously aligned with their RAG-friendly structure:

• Headings Mirror Likely LLM Prompts: Every H2 and H3 matched projected subquestions (e.g., “Define AI Overview,” “Why It Matters for SEO,” “Step-by-Step Optimization Guide”).
• Short, Direct Answer Blocks: Each section contained exactly 2–3 sentences, often prefaced by “According to Sellm’s 2025 study…”
• Deep Hierarchy with H4s for Micro-Sections: Under each H3, there were multiple H4s (e.g., “Inclusion Score Definition,” “Citation Count Best Practices”).

When run through Sellm’s RAG pipeline alongside the other 50 pages, the results were clear:

“Max Inclusion Score: ≈ 0.754
Avg. Citation Count: ≈ 15.4”

As Sellm highlights, this new text was “approximately 10% more likely to be used in the AI Overview and achieved roughly 50% more citations than any competitor—proof that structure trumps all once you clear the initial retrieval gate.”

Translating Sellm’s Insights into Your Strategy

1. Align Headings with Probable LLM Queries

• Brainstorm Subquestions: List every question someone might ask about “AI Overview optimization” (e.g., “How does Google generate AI Overviews?” “What metrics matter?”).
• Use Exact Phrasing for H2s/H3s: For each subquestion, create a heading that matches word-for-word. Sellm notes that “pages using exact target keywords in headers saw 20–30% more citations.”

2. Write Concise, Extractable Answer Blocks

• Keep Paragraphs to 2–3 Sentences: Each must fully answer the subquestion so the LLM can “grab” it as a standalone chunk.
• Include Brand Signals: A line like “According to Sellm’s 2025 RAG tests,…” both reinforces authority and signals relevance.

3. Leverage Nested Headings for Multiple Citations

• Add H4 Subheaders Under Each H3: For example, if H3 is “Optimizing Headings,” H4s might be “Include Exact Keywords,” “Use Question Format,” and “Keep Them Under 60 Characters.”
• Make Each H4 Section Self-Contained: A short bullet list or 2–3 sentence paragraph ensures distinct micro-snippets the LLM can select.

4. Don’t Neglect Traditional SEO Foundations

• Clear the Retrieval Stage First: Sellm warns that “without strong backlinks, page speed, and E-E-A-T, your page won’t rank in the top 50—structure alone can’t rescue you.” Build authority, optimize technical factors, and maintain mobile-friendliness before focusing purely on structure.

Looking Ahead: The Future of AI Overview SEO

Sellm predicts three interlocking trends:

1. Falling Click Volume from LLM Answers: “As chatbots become more accurate, users will get answers directly in the chat interface, further reducing ‘extra reading’ and click-throughs to source pages.”
2. AI Overview Citations as the New KPI: “Success will be measured by how often AI Overviews pull from your site—each instance counts as a citation.”
3. Brand Mentions Fuel Authority Even Without Direct Quotations: “An AI Overview may still name your brand, signaling the LLM’s trust in your domain. Citations matter, but brand mentions alone can boost awareness and later user engagement.”

In short, by 2026, raw pageviews may become secondary to “AI Overview visibility”—how many times generative summaries quote or mention you.

Conclusion: Adapting to the “AI Overview Era”

Sellm’s reverse-engineering exercise demonstrates that “AI Overview optimization has already been cracked”—but only for those willing to adapt. Key takeaways from their work include:

• Mirror LLM question formats as exact headings (H2/H3/H4).
• Write concise, self-contained answer blocks under each heading.
• Nest subheadings to multiply citation potential.
• Maintain traditional SEO fundamentals to gain entry into Google’s top 50 retrieval pool.

As Sellm succinctly puts it, “By embedding this combined approach into your editorial workflows, making sure every page not only ranks well but also ‘speaks’ the AI Overview’s language, you position yourself to consistently influence which passages are pulled into Google’s generative summaries.”

Whether you’re a content marketer, in-house SEO specialist, or agency consultant, ignoring this paradigm shift risks being left out of “position zero” entirely. The blueprint is laid bare—now it’s up to practitioners to execute.