What Retrieval Optimization (RO) Looks Like at Each Layer

The retrieval optimization post made the case that SEO and AI visibility are two lenses on the same four-layer system. The response so far has been positive, with practitioners using the layer language to describe what they were already seeing. This is EXACTLY the confirmation and reason I’m building this framing. To me it’s inherently intuitive and representative of what we all see happening. I’m just organizing, building a framework, mapping the whole thing, and naming it.

What people are seeing in different parts and levels – brands ranking well in Google but invisible in AI. Content cited without the brand getting recommended. Sites Google refuses to index dominating ChatGPT, all are the pieces of the full framework. Putting it all together can produce a bit of an “ah ha” in the organization of it, but that then leads to the next part.

The follow-up question that keeps coming up: what does the work actually look like at each layer when you spec it for both lenses? This post is a high level walk-through of exactly that. This is the operational version of the Retrieval Optimization (RO) framing. Namely, what changes when you stop optimizing for one lens (SEO or AI Vis) and start optimizing for both. We’re going to go Layer by layer, with concrete examples of what the work looks like when you spec it for the full retrieval system instead of half of it.

A few ground rules before we walk through it. The four layers are L1 Entity Establishment, L2 Entity Depth, L3 Category Citation, and L4 Content Optimization. They’re sequential — every retrieval system asks them in order. And at each layer, the same source material gets read differently by Google and by AI platforms. The work isn’t different. The way you spec the work is different. This is the basis of what RO is and how it works.

L1 — Entity Establishment

The work at L1 is foundation. Directory listings, NAP consistency, Google Business Profile, Wikipedia, Wikidata, schema markup, social profile completeness. This is where infrastructure overlap is essentially total. Both Google and AI platforms check the same directories, the same review profiles, the same structured data sources.

What changes when you optimize for both lenses?

Mostly nothing about the work itself. The Yelp page is the Yelp page. The schema is the schema. But the DETAIL you populate is what matters. Google reads your local citations for ranking signals…..citation volume, NAP consistency, recency. ChatGPT reads the same listings for entity confirmation  with things like does this brand exist, what category is it in, what attributes does it have. If you fill out NAP and skip the attributes (price tier, hours, payment methods, accessibility, services offered), you’re feeding Google what it needs but starving the AI lens.

The practical version: when you audit a directory profile, every populated field feeds both systems. Every empty field is value left on the table for one or both. Generic schema (Article, Organization, BreadcrumbList) hurts AI citation rates compared to no schema at all — Growth Marshal tested 1,006 pages and found generic schema dropped citation from 59.8% (no schema baseline) to 41.6%. Attribute-rich Product/Review schema brought it to 61.7%. Google barely cares which type you use. AI cares enormously. It’s not just SEO.

So at L1, optimizing for both lenses means: populate every field with attribute-rich detail. Skip generic schema if you can’t make it specific. Treat every directory profile as entity training data, not as a citation count.

L2 — Entity Depth

L2 is where the most important divergence in the entire framework lives. Both systems benefit from the same press placements, the same earned media, the same authoritative coverage. But they pull completely different value from those placements.

Google’s algorithm reads a TechCrunch feature for: link equity, anchor text relevance, PageRank transfer, referring domain authority. ChatGPT reads the same TechCrunch feature for: entity depth in training data, brand-to-category associations, co-occurrence with other authoritative entities, attribute context.

Same article. Different cargo extracted from it.

So practically speaking when you brief a PR team for a placement, you have two specs to write, not one. The SEO spec covers anchor text, link target, follow vs nofollow, target domain authority. That’s familiar. The AI spec covers brand mention frequency in the body (not just the byline), explicit category labeling (“[Brand], a leading provider of [category]”), attribute context (specific products, capabilities, differentiators), and co-occurrence with other named entities in the same space.

A placement optimized only for the link captures the SEO value and leaves the AI value on the table. A placement optimized only for the brand mention captures the AI value and may miss the link equity. The best placements get both specs in the same brief.

The data backs this up structurally. @ahrefs studied 75,000 brands and found web mentions correlated 0.664 with AI visibility — three times stronger than backlinks (0.218). YouTube mentions correlated 0.737, the single strongest predictor in the study. The mention is doing more work than the link for AI visibility, but the link is doing the work for SEO. You need both, specced for both.

Same earned media campaign. Same target list. Different brief instructions for each placement. That’s L2 retrieval optimization. Are we starting to see that it’s not just SEO here that covers?

L3 — Category Citation

L3 is where AI visibility lives or dies for most brands. The question every retrieval system asks at L3 is the same: “is this entity relevant to this category?” But the sources each system checks to answer that question diverge significantly.

Google determines category relevance primarily through organic SERP position — your ranking for category queries. If you rank #1 for “best CRM software,” Google has answered the question. ChatGPT determines it through training-layer pre-selection (which brands does the model already know belong in this category) plus listicle scanning (what does the editorial coverage of “best CRM software” articles look like). Perplexity heavily mirrors Google’s results — 91% domain overlap with Google’s top 10. Gemini often answers category queries from training data without retrieval at all.

Same question. Different sources checked. Different weights.

The practical version of L3 optimization is the work that addresses both pools at once. SERP ranking work – keyword targeting, content depth, topical authority, internal linking, link velocity – covers the Google side and Perplexity by extension. Editorial listicle placement, review platform activation, and brand-mention density across category-relevant editorial coverage handles the ChatGPT and Gemini side. I don’t make the rules, I just follow the patterns.

The brands that show up strong on Google but invisible on ChatGPT typically have great organic ranking for their category and zero presence on the editorial listicles ChatGPT scans. The brands that show up on ChatGPT but weak on Google typically have strong PR and earned coverage but poor on-page SEO. Both are doing half of L3, but just from opposite measuring sticks.

The fix at L3 is auditing both pools — your SERP position for category queries AND your appearance in the top editorial listicles ranking for those same queries. If you’re missing from either, you’re invisible to whichever system relies on that pool. The work isn’t doubled. It’s specced to land in both.

L4 — Content Optimization

L4 is the strongest convergence in the framework. The signals that get content cited by AI — answer-first structure, statistics over keywords, declarative language, evidence-backed claims, semantic HTML, clean section boundaries — are also what makes content perform well in Google. Both systems are trying to do the same thing at L4: extract the best answer from the page. They reward the same structural qualities because the task is identical.

The divergence is narrow. ChatGPT prefers tighter freshness cycles than Google’s AI Mode (60-day cadence vs evergreen depth). Section length sweet spots differ slightly (120-180 words for ChatGPT, 100-150 for AI Mode). Google still rewards comprehensive coverage and topical clusters more aggressively than ChatGPT does.

But the foundational content spec here for L4 is still just one spec. Answer-first paragraphs because @Kevin_Indig‘s analysis of 1.2M ChatGPT citations showed 44.2% come from the first 30% of the page. Statistics and data because the peer-reviewed Princeton GEO study (Aggarwal et al., KDD 2024) measured adding statistics as the single most effective optimization technique tested, with up to 41% AI visibility lift.

L4 retrieval optimization isn’t writing two versions of the page. It’s writing one version that follows the structural rules both systems reward, with awareness of the small platform-specific tuning at the edges. This is likely the best “GEO Is just SEO” level of the four, as it has that overlap and contains the most SEO-like tactics so it’s easy to say SEO covers it. But again, there are very specific AI-retrieval mechanisms happening here, so really this is a layer with 2 mechanisms.

What this actually means

At L1, optimizing for both lenses costs nothing extra. It’s the same directory work with more detail in every field.

At L2, optimizing for both lenses means writing two specs in the same brief. Same placements. Different instructions for what each system needs to extract from them.

At L3, optimizing for both lenses means auditing both pools — your organic ranking for category queries AND the editorial listicles ChatGPT pulls from for the same queries. The work to fix each is different, but the diagnosis is one diagnosis.

At L4, optimizing for both lenses is essentially one spec. The structural rules that win Google citations are the structural rules that win AI citations. Edge tuning for freshness and section length is the only divergence.

That’s retrieval optimization made operational. Not a new discipline you have to invent from scratch, but a new way of seeing the work you’re already doing — and specifying it so neither lens leaves value on the table.

The brands doing this well aren’t running two parallel programs. They’re running one program with both lenses applied at every layer. Every directory, every press placement, every editorial mention, every page on the site — specced for what Google extracts AND for what AI extracts from the same source.

The work isn’t twice as much. It’s the same work, told twice in the brief, executed once in the world.

More coming. Going to break down the sh*t out of this.