"My traffic dropped 18% since Google launched AI Overviews. I don't know if AI is just answering the queries now, or if I'm getting cited and not noticing."
That's the question nobody in your last standup could answer cleanly. Here's what's actually happening: your pages are being read and not cited. Not because your content is wrong, not because your domain authority dropped, but because your pages aren't structured in a way LLMs can extract cleanly. That's the whole problem. And it's fixable.
This post gives you the seven structural rules we apply to every page we want Claude, ChatGPT, and Perplexity to quote. Not a framework. Not a mindset shift. Seven specific, named, implementable changes, with the one you can do before end of day flagged.
TL;DR
AI Overviews don't reward backlinks or keyword density. They reward structural clarity. SAIO (Search AI Optimization) is the practice of formatting pages so large language models can extract, trust, and cite them. The seven rules below — a 30-second TL;DR, a bulleted summary, numbered steps, internal links, per-section factual anchors, an author byline, and a comparison table — are the checklist we apply to every page we want LLMs to surface. Structured content gets cited 3-4x more than prose-only pages, according to LLMRadar scans across client domains. If your pages are still written for Google's crawler, they're invisible to Claude's citation graph. This post walks through each rule, explains why it works at the token level, and gives you a copy-paste implementation pattern for each one.
- LLMs front-weight the first 200 tokens of a page when building citations. If your answer isn't in the first 150 words, it won't be surfaced.
- Structured content (headers, bullets, tables, numbered steps) is cited 3-4x more often than prose-only pages, based on LLMRadar scan data across client domains.
- SAIO and SEO optimize for different graphs: SEO targets PageRank signals (backlinks, domain authority, keyword density); SAIO targets LLM citation graph signals (structural clarity, fact density, citable units).
- Comparison tables are the single highest-impact structural element. LLMs extract tabular data and often reproduce it verbatim in answers.
- An author byline with credentials isn't vanity metadata. It's a trust signal LLMs use when deciding whether a factual claim is citable.
What SAIO Is (and Why It's Not Just SEO with a New Name)
Nobody explains what SAIO even means, and the posts that try spend two paragraphs defining it abstractly before getting to anything you can do. So here's the short version: SEO and SAIO optimize for two completely different graphs.
SEO targets the PageRank graph. Backlinks, domain authority, keyword density in title tags and body copy. Google's crawler follows links, scores authority, and ranks pages accordingly. You know this. You've been doing it.
SAIO targets the LLM citation graph. When Claude or Perplexity pulls a source, it isn't checking your domain authority. It's asking one question: can I extract a clean, self-contained answer from this page? If the answer is buried in paragraph six, wrapped in transition sentences, with no headers breaking the content into discrete units, the LLM reads the page and moves on.
Here's where it gets painful: your current top-ranking pages may score zero in SAIO. Prose that ranks well for crawlers is often too dense and unstructured for LLM extraction. The writing style that Google rewards (comprehensive, flowing, 2,000-word essays) is exactly the writing style that makes Claude skip your page. Based on LLMRadar scans across client domains, structured content pages are cited by AI Overviews 3-4x more than prose-only pages. That gap is the problem this checklist solves.
| Signal Type | SEO (PageRank Graph) | SAIO (LLM Citation Graph) |
|---|---|---|
| Primary ranking factor | Backlinks and domain authority | Structural clarity and extractability |
| Content format rewarded | Comprehensive prose, keyword density | Headers, bullets, tables, numbered steps |
| Trust signal | Referring domain quality | Author credentials, factual anchors per section |
| Answer placement | Anywhere in the page | First 150 words or rarely surfaced |
| Optimization unit | Full page relevance score | Discrete citable units per section |
| Measurement tool | Google Search Console, Ahrefs | LLMRadar, brand mention tracking |
Rule 1: 30-Second TL;DR Paragraph in the First 150 Words
LLMs front-weight the first 200 tokens of a page when building citation candidates. If your answer is buried in paragraph four, after a two-paragraph introduction that sets context, it won't be extracted. That's not a ranking penalty. It's a structural one.
A compliant TL;DR does three things in one dense paragraph: names the problem, states the answer, and identifies the key mechanism. No preamble. No "in this post we'll explore." The information starts in sentence one.
The common failure pattern: articles that open with "In today's digital landscape..." That's 40 wasted tokens before any citable content appears. Claude has already moved to the next candidate page.
Implementation pattern: write the TL;DR last. After you know what the post actually says, compress the core answer into 80-100 words and paste it at the top. It should read like the answer to the question your headline promises, not like an introduction to the answer.
Rule 2: Bulleted Summary of 3-5 Key Takeaways Before Deep Content
Each bullet is a discrete, self-contained unit of meaning. LLMs can lift one bullet without needing surrounding context to make it coherent. That's why bullets beat prose for extraction. A paragraph requires the LLM to parse the whole thing to find the claim. A bullet hands the claim over directly.
Placement matters: the summary block belongs after the TL;DR and before the first H2, not at the end as a recap. Recaps are for human readers who scroll back up. LLMs read top to bottom. Put the summary where the extraction happens.
Write each bullet so it can stand alone as a quoted sentence in an AI answer. Subject, verb, specific claim, no pronouns requiring antecedents. "Structured content is cited 3-4x more than prose-only pages" is a citable unit. "This approach works better" is not.
Rule 3: Numbered Steps Where Applicable
Numbered lists tell the LLM this is a procedure, not an opinion. Procedures get cited as authoritative instructions. Opinions get paraphrased as perspective, or skipped. That's a meaningful difference when you're trying to show up as a source rather than a point of view.
When to number versus when to bullet: if the sequence matters (do step 2 before step 3), number it. If order is arbitrary, bullet it. Don't number things that aren't actually sequential just to signal authority. LLMs have seen enough content to recognize when step order is cosmetic.
Each step should include three elements: a verb in the imperative, the object, and the expected outcome. "Add a TL;DR paragraph in the first 150 words so LLMs can extract your core answer before the page body begins" is a citable step. "Optimize your opening" is not.
Structured content (headers plus bullets plus tables plus numbered steps combined) is cited 3-4x more than prose-only pages according to LLMRadar scan data. No single element does it alone. The combination is the signal.
Rules 4 and 5: Internal Links and Per-Section Factual Anchors
Rule 4 is about the citation graph at the cluster level, not just the page level. LLMs weight pages that link to related authoritative content because it signals the domain has depth. One well-optimized post with no internal links looks like an isolated document. Three posts linking to each other, plus a product page, looks like a domain that knows what it's talking about.
The link target formula: 2-3 sibling posts in the same topic cluster, plus one product or solution page. Enough to signal depth without diluting the page's own citable density. And use descriptive anchor text that names the destination topic. "Click here" contributes nothing to the citation signal. "How LLMRadar measures brand citation frequency" does.
Rule 5 is simpler: every H2 section needs at least one grounded factual claim. A specific number, a named source, or a named methodology. LLMs use factual density as a proxy for trustworthiness. Pages that assert things with supporting evidence rank higher in the citation candidate pool than pages that state opinions without attribution.
The minimum viable citation isn't a peer-reviewed study. It's a specific number tied to a named source or methodology. "3-4x more citations, based on LLMRadar scans" clears the bar. "Much more effective" doesn't. The failure mode here is opinions stated as facts without attribution. Those are exactly what LLMs either paraphrase without credit or skip entirely. Each H2 that contains a grounded factual claim is a discrete citation opportunity. Five H2s with grounded claims is five chances to be cited in a single AI Overview response.
Rules 6 and 7: Author Byline with Credentials and Comparison Tables
Rule 6: models trained on web data have learned that bylined content with specific credentials is more reliably accurate than anonymous content. That's not a philosophical point. It's a pattern baked into training data. A page with "Christine Johnson, Head of Content, 6 years in B2B SaaS SEO" is a different trust signal than a page with no attribution at all.
A compliant byline contains four things: name, role or title, a one-line statement of relevant expertise, and optionally a link to an author page or LinkedIn profile. This is also where SAIO and Google's E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) framework converge. Author credentialing helps both. It's one of the few structural changes that improves your standing in two graphs simultaneously.
Rule 7 is the one to prioritize if you're only going to do one thing before end of day. Comparison tables are the single highest-impact SAIO structural element. LLMs extract tabular data and present it directly in answers, often verbatim. A well-structured table is the closest thing to a guaranteed citation.
Table construction rules for LLM extraction: include a clear header row with named attributes, keep cell content to one fact per cell, and make the table self-explanatory without surrounding prose. If someone lifted the table out of your post and dropped it into a Slack message, it should still make sense. That's the test.
Place the comparison table at the natural decision point in the post, where a reader would ask "how does this compare to the alternative?" That's also where LLM extraction is most likely to occur. For most posts, that's somewhere in the first half, not at the end after three pages of context. The SAIO vs. SEO table earlier in this post follows that rule exactly.
The Full SAIO Checklist: Run It Before You Publish
Here are the seven rules as a pre-publish gate. Each one either passes or fails. No partial credit, because LLM extraction is binary: the page gets cited or it doesn't.
- Does the page open with a TL;DR paragraph that names the problem, the answer, and the key mechanism, all within the first 150 words?
- Is there a bulleted summary of 3-5 key takeaways placed after the TL;DR and before the first H2?
- Are sequential procedures written as numbered steps, each with an imperative verb, an object, and an expected outcome?
- Does the page include 2-3 internal links to sibling posts in the same topic cluster, plus one link to a product or solution page, with descriptive anchor text?
- Does every H2 section contain at least one grounded factual claim (a specific number, a named source, or a named methodology)?
- Is there an author byline with name, role or title, and a one-line statement of relevant expertise?
- Is there a comparison table placed at the natural decision point, with a header row, one fact per cell, and no dependency on surrounding prose to make sense?
For existing pages, prioritize by organic traffic volume first, then by topic. High-traffic pages covering how-to, comparison, and definition queries return the highest SAIO ROI because those are the query types where AI Overviews appear most frequently. LLMRadar data suggests 2-6 weeks from page update to citation appearance in AI Overviews, so changes you make this week should show up in your brand mention tracking by end of the month.
What to track after implementation: brand mention frequency in LLM outputs, not just traditional ranking positions. Your Search Console impressions and click-through rates tell you what Google is doing with your pages. They don't tell you whether Claude is citing you. Those are two different measurements now, and you need both.
The seven rules aren't aesthetic preferences. They're structural affordances that LLMs are trained to recognize as signals of extractability and trustworthiness. A page without a TL;DR in the first 150 words, without numbered steps, without a comparison table, and without per-section factual anchors is a page that Claude will read and not cite. Apply the checklist. Audit your existing pages in priority order. Then measure. If you want to know exactly where your domain stands in the LLM citation graph before you start, the LLMRadar Audit gives you that scan for $197.
Run the LLMRadar Brand Audit →
Next post: How the LLMRadar $197 audit works under the hood. Methodology, what the scan actually measures, and how to read your report.