- Most GEO advice focuses on AI outputs marketers cannot directly control. The more realistic work sits in two areas: making content retrievable and building stronger brand associations across trusted sources.
- AI search visibility is also harder to measure than referral traffic suggests, because brands can be recommended inside AI answers without receiving a single trackable click.
THE 5 LAYERS BEHIND EVERY AI RESPONSE 🤖 AND HOW TO DO REAL AI SEO NOT FAKE GEO…
Every ChatGPT answer, every Google AI Overview, every Perplexity response your customers see is currently generated through these exact five layers.
And SEOs can ONLY really influence TWO of… pic.twitter.com/yrcaHUqFzH
— Charles Floate 📈 (@Charles_SEO) July 7, 2026
A July 7 thread from Charles Floate put a sharper frame around one of the biggest problems in generative engine optimization: a lot of the work being sold as GEO targets parts of AI systems that marketers cannot actually reach.
Floate’s argument is not that AI search optimization is fake. It is that the useful part is narrower than the market often makes it sound.
His framework breaks an AI answer into five layers: latent knowledge, system prompts, retrieval, evidence integration and final generation. SEOs, he argues, can meaningfully influence only two of them: the long-term knowledge layer, where brands become associated with topics and categories over time, and the retrieval layer, where AI systems pull current information from the web.
That distinction matters. It separates real AI search work from vague claims about “optimizing for LLMs” as if marketers can directly control how ChatGPT, Gemini, Perplexity or Google AI Overviews decide to phrase an answer.
The practical takeaway is much less magical. Brands need to be findable, understandable, corroborated and consistently mentioned in the places AI systems rely on.
The Five Layers Behind an AI Answer
Floate’s model starts with latent knowledge. This is what a model has absorbed during training: the associations between brands, topics, people, categories, sources and claims.
A brand that appears repeatedly beside a category across news sites, Reddit discussions, Wikipedia-style references, PDFs, industry pages and reviews has a better chance of becoming part of that broader association network. That does not change overnight. It is built through repeated, consistent presence across sources the model may learn from.
The second layer is the system prompt. This is the hidden instruction layer that shapes how the model behaves, formats answers, handles sensitive topics and decides what kind of response is allowed. Marketers do not have access to it.
The third layer is active retrieval. This is where the work starts to look very familiar to SEOs. AI systems need accessible content to retrieve, whether through search indexes, browsing systems, partner data or other grounding sources. If a page cannot be crawled, indexed, understood or matched to the query, it is unlikely to become part of the answer.
The fourth layer is evidence integration. This is where retrieved information gets weighed against other signals. A claim that appears only on a brand’s own site is weaker than a claim supported by reviews, media coverage, comparison pages and third-party references.
The fifth layer is final generation: the answer written token by token. This is the part users see, and often the part marketers obsess over. It is also the part they control least.
That is the core of Floate’s critique. The visible answer is not the optimization surface. The inputs that make a brand retrievable and credible are.
Why This Pulls GEO Back Toward SEO
One of the more useful parts of Floate’s argument is that retrieval is not a new discipline wearing a new label.
If an AI system needs to fetch information before answering, then classic SEO foundations still matter: crawlability, indexability, page structure, internal linking, topical relevance, fresh content and clear entity relationships.
Cyrus Shepard’s meta-analysis of 54 AI citation studies, published in May 2026, pointed in the same direction. URL accessibility and search rank were among the strongest citation factors across ChatGPT, Gemini and Perplexity.
That does not mean AI search works exactly like Google rankings. It does mean the retrieval layer still depends heavily on whether content can be found and understood in the first place.
Where AI search changes the work is in extraction.
AI systems do not read pages like human visitors moving through a designed experience. They pull, compress and compare pieces of information. Dense prose, vague positioning and buried facts make that harder. Short sections, clear headings, named entities, specific claims and direct comparisons give retrieval systems cleaner material to work with.
That is not “writing for robots.” It is making useful information easier to extract without losing meaning.
The Attribution Gap Is Already Showing Up
The measurement problem is just as important as the optimization problem.
Aleyda Solis published a framework in April 2026 that separates AI search measurement into three layers: Presence, Readiness and Business Impact.
Presence asks whether a brand appears in AI answers and how it is described. Readiness checks whether the technical and content conditions for that visibility are in place. Business Impact asks whether any of that visibility leads to commercial value.
That last layer is where standard analytics start to break.
A user may ask ChatGPT which PR platform fits their agency, read the answer, form a preference and then search the brand name on Google or go directly to the website. The AI answer influenced the decision, but the referral click never happened. GA4 may credit organic search, direct traffic or branded search instead.
The AI system did the recommendation. Analytics saw only the final visit.
That makes AI visibility easy to underestimate. Referral traffic from ChatGPT, Perplexity or Gemini is useful, but it does not capture every recommendation, comparison or shortlist that happens inside an AI interface.
For brands, this means AI search reporting needs more than session counts. It should also track whether the brand appears, how it is framed, which competitors appear beside it and whether the recommendation is accurate.
Evidence Integration Is Where Many Brands Get Filtered Out
The evidence layer is the part of Floate’s framework that deserves more attention.
Being retrieved is not the same as being used. A page can rank, be crawled and enter the candidate set, but still fail to appear in the final answer if the claim is not supported elsewhere.
That matters for content strategy. A brand can say it is the best option for a category on its own site. But if that claim does not appear across reviews, third-party articles, comparison pages, customer discussions or industry references, the model has little reason to treat it as reliable.
Solis’s Readiness layer points to the same issue. Brands need clear claims, consistent positioning and external corroboration. Without that, visibility becomes fragile.
Ahrefs’ study of 75,000 brands found that branded web mentions had a much stronger correlation with AI Overview brand visibility than backlinks. That does not mean backlinks no longer matter. It does suggest that off-site brand presence now carries a different kind of importance in AI search.
This is where SEO, PR, content and brand marketing start to overlap. AI systems are not only looking at whether a page exists. They are also looking for signals that make the information easier to trust.
The GEO Market Needs More Precision
Floate’s thread also lands because the GEO tooling market has become crowded very quickly.
Some tools are useful. Visibility monitoring can show where a brand appears. Technical tools can help with crawl access and structured content. Content platforms can make pages easier to update and extract from.
But none of that gives marketers direct control over system prompts or final model generation.
That is where some GEO claims become slippery. Promising visibility is not the same as measuring it. Improving content structure is not the same as controlling an answer. Tracking citations is not the same as earning them.
Search Engine Land’s analysis of 8,000 AI citations found that official company sites and blogs accounted for fewer than 4% of citations, while media, review sites and mainstream publications appeared much more often.
That finding is uncomfortable for brands that still treat their own website as the center of every visibility strategy. In AI search, third-party validation can matter more because it helps the system compare options from a source that is not the brand itself.
The realistic stack is not one GEO tool. It is strong SEO, structured content, third-party mentions, review visibility, comparison coverage, clear entity signals and measurement that looks beyond referral traffic.
What the Work Actually Looks Like
Floate’s framework leaves marketers with two practical tracks.
The first is long-term entity building. Brands need to become consistently associated with the categories, problems and use cases they want to own. That means earning mentions across trusted publications, industry resources, review platforms, communities, video platforms and comparison pages.
This does not work as a short campaign. If the goal is to influence model familiarity and brand associations, consistency matters more than bursts.
The second track is retrieval optimization. This is the faster-moving work: make high-value pages crawlable, keep them current, structure them clearly and build content around the prompts buyers actually ask.
Comparative searches deserve special attention. Queries like “best tool for,” “alternative to,” “which platform is better” and “what should I choose for” are closer to purchase decisions than broad informational searches. They are also the kinds of prompts where AI systems often summarize options instead of sending users to a list of blue links.
Ahrefs’ Q1 2026 AI Search Benchmark Report, based on 730,000 AI responses and 146 million search result pages, found that YouTube mentions had one of the strongest correlations with AI brand visibility. That reinforces the larger point: AI search visibility is not only a website problem. It is a presence problem across the sources AI systems can observe.
For SaaS brands, that means product pages, documentation, comparison pages, YouTube explainers, reviews and third-party articles all matter. For ecommerce, product data, reviews, merchant feeds and category authority matter. For service businesses, recommendations, local proof, case studies and clear positioning become more important.
The right KPI depends on the business model. Ecommerce teams should watch linked citations and comparison visibility. Lead-gen businesses should care about recommendation rate. SaaS companies need to track representation accuracy, because being described incorrectly inside an AI answer can damage conversion before a user ever reaches the site.
What Marketers Should Take From This
The useful version of GEO is not about hacking AI answers.
It is about becoming easier to retrieve, easier to understand and easier to trust.
That means the closed layers are the wrong place to spend energy. Marketers cannot buy access to the system prompt. They cannot reliably control final answer wording. They cannot force a model to cite them because a page has been rewritten with AI-search language.
They can improve the inputs.
They can build pages that answer buyer questions clearly. They can earn mentions from sources AI systems already trust. They can make claims consistent across owned and third-party surfaces. They can monitor how AI systems describe the brand. They can fix inaccurate positioning when it appears. They can treat AI visibility as a search, content, PR and brand problem at the same time.
That is less exciting than the GEO sales pitch, but it is more useful.
AI search is not replacing SEO with a completely new discipline. It is adding retrieval, extraction and evidence layers on top of the web that already exists. The brands that adapt fastest will likely be the ones that stop chasing control over the final answer and start improving the signals that shape whether they enter the answer at all.
