AI visibility is a 4-layer stack. Most operators are working on the wrong layer.
Most AI visibility work is happening at the wrong layer.
The pattern we keep seeing: an agency or business decides they need to "show up in ChatGPT." They sign up for an AI visibility tracking tool. They monitor their score weekly. Nothing improves. They conclude AI visibility is impossible.
It's not impossible. They're optimizing layer 4 while layer 1 is broken.
After running AI visibility audits across hundreds of brands since 2024, the problem space sorts into four distinct layers. Each layer addresses a different question. The order matters because layers depend on each other.
Layer 1: Entity coherence
Can AI systems identify your brand unambiguously? Is your business represented as a coherent entity across the data sources LLMs train on?
Most failure modes start here. A clinic with three slightly different name variants across the web doesn't show up in ChatGPT — not because it's invisible, but because AI systems can't decide which entity to surface.
Layer 2: Citation propagation
Once your entity is coherent, are you cited in the source material AI models train on and retrieve from? News, directories, industry publications, reference sites. Citation density at this layer is what makes you visible in retrieval-augmented AI search (Perplexity, ChatGPT Search, Google AI Overviews).
Layer 3: Editorial placement
Beyond raw citations, are you mentioned contextually in editorial content that AI systems treat as authoritative? Industry coverage, expert commentary, and journalist mentions. This is where LLMs form opinions about your brand, not just acknowledge its existence.
Layer 4: Behavioral signal
Once you appear in AI results, are users clicking through, dwelling, returning? AI search platforms increasingly use behavioral signals to reinforce or downgrade visibility — mirroring what Google did with traditional search a decade ago.
Why the order matters
You can run layer 4 campaigns at a brand with a broken layer 1. The campaigns will fail. The clicks won't compound because the underlying entity is too unstable for AI systems to consistently surface.
You can also obsess over layer 2 citations without fixing layer 1. The citations exist but don't help, because AI systems can't tie them to a coherent brand.
The right sequence is foundational: entity coherence first, citation density next, editorial placement, then behavioral signal. Skipping forward without the layers beneath is the most common money-losing mistake we see.
Honest caveat
This 4-layer stack is our working model from audits since 2024, not an industry standard. We don't have access to how each AI platform weights signals internally — the layers are inferred from observable outcomes across audit and campaign work, roughly 200 brand audits in total. The principle (foundation before tactics) is the durable part. Layer boundaries will keep refining as the AI search itself evolves.
The diagnostic question to start with: which layer is actually broken for your brand? Most operators assume layer 4 because that's where the tracking tools point. In practice, layer 1 or layer 2 is broken for ~70% of brands we audit.
If you're not seeing AI visibility traction, the audit is where I'd start.
Curious whether others working on AI visibility are using a similar layer model, or thinking about it differently.
Building Webido CTR since 2019
https://webidoctr.com/ai-visibility-audit (use IHAI20 for $20 off — $97 Starter / $197 Advanced)
The interesting part is that your framework assumes the right entity already exists.
One thing I'd be careful with is treating entity coherence as a technical problem only.
For a lot of companies, the entity is coherent, but the positioning isn't.
The brand gets mentioned consistently, but AI systems keep associating it with the wrong category, wrong use case, or wrong comparison set.
That creates a weird situation where visibility improves, but for the wrong queries.
Feels like there's a layer before citation propagation: making sure the market and AI systems agree on what the company actually is.
Curious whether you've seen that show up in the audits.
Good catch, aryan_sinh. You're right that "entity coherence" as I framed it leans toward identity coherence (is the brand identifiable?) and underplays categorical coherence (what category, use case, and competitive set does the brand get placed in?). Real distinction worth surfacing.
We see exactly this in audits. Common pattern: a vertical SaaS tool gets perfectly recognized as "X" but AI systems consistently slot it next to horizontal CRMs because the citation footprint trained the embedding that way. The brand is visible. The visibility is for the wrong queries.
The audit treats these as two separate diagnostic questions in practice, but the framework, as written, collapses them into one. Two reasons it probably deserves its own callout: (1) the intervention is different (categorical repositioning requires editorial work, not just citation/NAP fixes), and (2) it tends to be more expensive to fix because you're unwinding established AI associations rather than building from scratch.
Will surface this distinction in the next revision of the working model. Identity coherence is "can AI systems identify the brand?" Categorical coherence is "what does AI think the brand is FOR?" Both are foundational, but they fail differently.
That distinction is exactly the one I was getting at.
Identity coherence gets the brand recognized. Categorical coherence decides whether that recognition is commercially useful.
The dangerous version is when AI systems understand the company, but place it in the wrong buyer context.
If you’re open to it, send me your email. I’d like to share a tighter thought on how that distinction could show up in the audit flow without crowding the thread.
That "commercially useful" framing is sharp recognition without the right buyer context is exactly the failure mode most operators don't see until conversion data stops matching visibility data.
On the offline thread: I'd rather keep framework refinement in public if you're open to it. The IH thread becomes a useful record for other operators working through the same problem, and your distinction is good enough that it should live somewhere indexable. Happy to keep going here.