What Two AIs Revealed About How AI Search Actually Decides What to Cite

The Experiment: Two AIs, One Question, Two Very Different Outcomes

Last week I ran a small experiment that turned into the clearest look I have had at how AI search actually works under the hood. I gave ChatGPT and Claude the identical prompt:

“I want a tool that analyzes my whole blog as a knowledge graph and tells me how to connect content so LLMs see it as coherent expertise.”

I never named VizzEx, the AI-visibility tool I co-founded with Carolyn Holzman. I wanted to see whether either model would get there on its own.

Why ChatGPT Found VizzEx in Eight Seconds—and Claude Never Did

ChatGPT took eight seconds. It ran a few queries behind the scenes and named VizzEx as the closest match, with a short list of alternatives underneath.

Claude took the better part of an hour. It asked good clarifying questions, then decided to build a tool instead of finding one: it crawled all 113 posts of a B2B marketing-automation company I founded called Genoo, and produced a beautiful interactive star-map of the blog. Then I asked it to recommend existing tools. It named five. It never named VizzEx. When I asked why, it gave me an honest answer: VizzEx was too new to be in its training data, too new to surface in the “best tools” roundups its searches leaned on, and its instinct had been to build rather than find.

(I told the full story, including Claude’s surprisingly candid confession, in my LinkedIn newsletter [WILL LINK TO NEWSLETTER SOON]. This piece is the part I could not fit there: why it happened.)

Four Layers of AI Visibility: Miss One and You Disappear

Here is the thing that stuck with me. This was not a story about one model being smarter than the other. It was a story about the machinery underneath every AI answer, and there are four distinct layers to it. Miss any one of them and you are invisible, no matter how good your work is.

Layer 1: Representation. Does the model already know you exist?

Start with the most uncomfortable layer, because it is the one almost nobody is optimizing for.

How AI Models Define ‘Domain Authority’ (It’s Not What You Think)

To a large language model, “authority” does not mean what it meant in traditional SEO. It has nothing to do with backlinks or a domain score. Jérôme Salomon, a senior technical SEO at Oncrawl, reverse-engineered the hidden JSON file behind ChatGPT’s search and confirmed the selection criteria directly with OpenAI’s support team. When ChatGPT decides which results to trust, it weighs four things: the title, the snippet, the publishing date, and “domain authority.” And domain authority, to the model, is not a Moz number or a backlink count. As Salomon puts it, there is no PageRank system in the model’s eyes. “Domain authority” means the model already learned to trust your brand from its training data. If your brand barely existed when the model was trained, you barely register now.

That single fact explains the whole experiment. Claude works from a frozen training cutoff plus the searches it chooses to run, and it had no stored signal for a brand-new tool, so it could not see VizzEx. ChatGPT’s live retrieval surfaced VizzEx anyway, because retrieval reaches past the training cutoff to whatever is currently indexed and mentioned.

From Backlink Rankings to Brand Mentions: The New Visibility Game

Salomon’s other finding matters just as much: the old backlink game is being replaced by a mentions game. The more your brand is mentioned across the web, the more likely you are to become part of the answer. As he frames it, it is no longer about ranking, it is about becoming the answer. This is exactly the shift we keep coming back to in why semantic relationship links are now the critical AI-visibility signal: visibility is built from how thoroughly and coherently you are represented, not how many links you have hoarded.

Takeaway 1: Representation Is Earned Over Time, Not Retrofitted

representation is earned over time, across the open web and in the training corpus. You cannot retrofit it overnight. But you can stop sabotaging it, which is what the next three layers are about.

Layer 2: Crawlability. Can the machine physically read you?

Say a model decides your page is worth a look. Now a bot has to fetch it and come away with usable content. This is the layer my co-founder Carolyn lives in, because it is invisible to marketers and brutal in its consequences.

Why AI Bots Don’t Run JavaScript—and What That Costs You

Salomon is blunt about the most common failure: the bots do not run JavaScript. Your content has to be present in the raw DOM. If your links, your key claims, or your structure only appear after client-side JavaScript executes, the fast fetch that decides citations never sees them. This is the precise reason I am careful about how internal links get added. It is why InLinks injecting links with JavaScript is a real problem for AI visibility and not a footnote: a runtime-injected link is invisible to a crawler that never runs the runtime.

Page Speed as a Citation Race: Why Slow Servers Lose to Faster Competitors

There is a second crawl killer that almost nobody talks about: speed. Salomon describes the live citation process as a race. When a model needs to cite, it fires off the fetch, and the first couple of servers to respond get quoted while the connection to the slower ones is cut. If your server is slow, you can be the best answer on the web and still lose the citation to someone faster.

The Symmetry Gate: How Mismatched HTML and Rendered Pages Get Your Site Evicted

Carolyn’s forensic work pushes this further, into the access logs themselves. She reads server logs the way Salomon does, watching exactly which AI bots arrive, what they request, and what they get back. Her central finding has a name in our system: the Symmetry Gate. Modern indexers do a two-layer pass, first ingesting the raw HTML, then later rendering the JavaScript. If those two versions do not match, the page fails a parity check and gets evicted. As she lays out in her breakdown of why AI quietly drops most comparison pages, the failure is not editorial or moral. It is financial: reconciling a mismatched page costs the machine too much compute, so it simply stops trying.

We built a free tool that runs exactly this check. SymmetryGate.ai audits any URL for structural fidelity across ChatGPT, Claude, Gemini, and Perplexity, pinpoints the gaps where your raw structure and your rendered page drift apart, and calculates the compute tax a model pays to read you, so you can see whether your own pages are clean enough to be cited.

Chunk Autonomy: Why AI Strips Your Brand From Unlinked Claims

Her companion piece on the death of PageRank sculpting explains the mechanism behind it. AI does not read your page as one document. It chops it into independent chunks and evaluates each one in isolation. A chunk that makes a claim without a verifiable link to support it reads as an anonymous, unverified assertion, and the model strips your brand off it. Her line for this is the one I quote most: in the database-driven web, “nodes that are not connected do not exist.”

Takeaway 2: Fast, Raw-HTML Delivery Is the Foundation Everything Else Builds On

before any clever content strategy, make sure the machine can fetch you fast and read you in raw HTML. If it cannot, nothing else matters.

Layer 3: Coherence. Can the machine see your expertise, not just your words?

Now assume the bot reaches you, fast, in clean HTML. It can read every word. Here is where most good blogs still fail, and Claude’s audit of Genoo proved it on a real site.

Claude’s Audit of 113 Blog Posts Revealed a Field of Disconnected Islands

Claude crawled all 113 Genoo posts and its own analysis came back with the diagnosis: 21 orphan posts that nothing links to, “anchor” posts that nothing points at, and core concepts mentioned everywhere but owned nowhere. In its own words, a model “has to reconstruct your expertise from scratch every time.” Genoo has a genuinely good blog. To a machine, it still reads as scattered islands.

The beautiful knowledge graph diagram that Claude created to represent the analysis.

That is precisely what it means to see your content the way AI sees it: not as a collection of posts, but as a network of relationships the machine either can or cannot follow.

This connects straight back to chunk autonomy. If the model evaluates each chunk alone, then a pile of individually fine posts with no explicit relationships between them gives the model no way to assemble them into a body of expertise. Each chunk standing by itself, alone, rarely wins. That is the whole argument behind what semantic content analysis actually is: the difference between mentioning concepts and connecting them in a way a machine can follow.

VizzEx allows you to see the connectivity across islands in ways that are easily understandable.

Typed Semantic Relationships vs. Keyword Links: Why the Difference Matters to AI

And connection has to be the right kind. A keyword-matched link tells a system that two pages share a word. A typed semantic relationship tells the system how two ideas relate: that one is a prerequisite for another, that one is an integration pattern, that one supplies the evidence for another’s claim. That distinction is the difference between content that looks busy and content that demonstrates a methodology. It is also why a beautiful map of your content is not the same as a structured one, a point worth paying attention to given how AI evaluates the geometry of how your concepts sit in relation to each other.

Takeaway 3: Coherence Means Explicit, Typed Relationships—Not Just Good Writing

being readable is not the same as being coherent. Coherence means the relationships between your ideas are explicit, typed, and present in the text itself.

Layer 4: Execution. Can You Turn AI Diagnosis Into Machine-Readable Action?

This is the layer that turned Claude’s experiment into a lesson about tools, not just AI.

Why a Beautiful Knowledge Graph Is Not the Same as an Actionable Work Queue

Claude built a force-directed “constellation” of Genoo’s blog: 113 dots, color-coded, sized by inbound links, orphans ringed in red. It was genuinely beautiful. It was also, for a working marketer, close to useless. You have to read a legend to begin, the orphans are dots floating in black space, and when you are done admiring it, it still cannot tell you what to do. It is a picture you decode, not a list you work. I have tried the established knowledge-graph tools that produce the same kind of hairball, and I hit the same wall every time.

The gap between a graph you admire and a queue you work is the entire point of execution. It is the difference between a tool that diagnoses and a tool that helps you finish. A diagnosis tells you that you have 21 orphans. Execution tells you which post to link each orphan from, writes the sentence that makes the connection, explains the relationship type so the link is meaningful to a machine, and tracks whether you actually did it.

VizzEx provides easy to understand Connectivity tiers, with an action plan that can easily be implemented. The guide that makes it easy.

And execution has to respect Layers 2 and 3 or it is worthless. A link written for you only counts if it lands in real HTML the crawler can read, not JavaScript it will skip. The relationships only count if they are typed and explicit. Carolyn’s work goes one step further here, encoding those relationships into the page’s structured data so the machine is handed the network directly instead of being asked to infer one. That is the move from flat schema to relational schema she describes in the chunk-autonomy breakdown: give the crawler a pre-assembled map of how your facts connect, and you stop making it pay the compute tax of figuring it out.

Takeaway 4: Analysis That Doesn’t End in Completed Markup Is Just Decoration

Analysis that does not end in completed, machine-readable work is decoration. The job is not to see the problem. It is to finish the fix.

Getting cited is easy. Staying cited is the loop.

Here is the part that ties all four layers together, and it is the part almost nobody understands.

Remember Layer 1: if you are too new for the training data, you are invisible without live retrieval. But live retrieval is not a consolation prize. It is how you get cited without being in the training data at all. When a model hits a question it cannot answer from memory, it runs a live search, grabs a source, and hands it what amounts to a probationary citation. You are in the answer. For now.

That citation is on loan, though, not fused into the model. Weeks later, the model goes back and examines the source more closely, and on that second look, Layer 2 is the gate. If your page cannot be read cleanly and fast, you are evicted from the answer. That eviction is not random. It follows a measurable curve: citations decay with a 4.5-week half-life, the math of which is laid out in the breakdown of inference metabolism.

Pass all four layers, though, and the loop runs the other way. You survive the re-check. You keep getting cited. And the sources that keep getting cited are the ones that get pulled into the next version of the model’s training data, which is how you finally earn Layer 1 for real. That is the whole game in one line: getting cited once is luck, staying cited is architecture, and architecture is how temporary visibility becomes permanent.

Genoo vs. VizzEx: A Before-and-After Look at AI-Visible Content Architecture

Put the four layers together and you get a simple test you can run on any site: is it represented, crawlable, coherent, and executable?

I have two blogs that sit on opposite ends of that test, which makes the contrast easy to see. Genoo and VizzEx are two different companies. I founded Genoo; I co-founded VizzEx with Carolyn. What connects them is me, and the fact that I watched the islands problem play out on one blog and built the cure into the other.

How Deliberate Link Architecture Makes Every Chunk Citable by AI

Genoo’s blog is the “before”: good writing, scattered structure, a field of disconnected dots to a machine. Understanding why that scattering happens, and how to read it systematically, is exactly what separates scattered islands to a machine from a coherent, citable knowledge structure. The VizzEx blog, The Signal, is the “after,” and the post you are reading is itself an example of it. Notice that the links above do not say “click here.” Each one names a relationship and points to the single post that owns that idea: the forensic crawl mechanics live in one place, the semantic-relationship argument in another, the comparison-page autopsy in another. The same vocabulary runs across posts written by two different authors. That is not an accident of good writing. It is architecture, applied deliberately, so that when a model chunks this page it finds every claim connected to a verifiable home.

Your Four-Layer AI Visibility Checklist: Representation, Crawlability, Coherence, Execution

AI will not reward expertise it cannot see. And “see” turns out to mean four specific things, in order:

1. Represented so the model already knows your brand exists.
2. Crawlable so the bot can fetch you fast and read you in raw HTML.
3. Coherent so your ideas are connected with explicit, typed relationships. Understanding what explicit, typed relationships means at the content level is the foundation for making this layer work.
4. Executable so the work actually gets done, in the markup the machine reads.

The Marketers Who Win AI Search Will Be Legible at All Four Layers

Claude failed Layer 1 on a brand-new tool and, when it tried to help, stopped at a diagnosis it could not turn into action. ChatGPT cleared Layer 1 through live retrieval and pointed me at a tool that handles the rest. Neither outcome was about intelligence. Both were about machinery.

That machinery is now the game. The marketers who win the next few years will not be the ones who publish the most. They will be the ones whose expertise is legible to the machine at every one of these four layers.

If you want to see where your own blog stands, that is exactly what we built VizzEx to not only fix, but guide you effectively..