Measuring Website Readiness for AI Search: The Carcass Method Framework
The most useful AI search readiness audit checks a website on the same four layers it is built on. The Carcass Method describes a website as Surface, Structure, and Data, with The Carcass as the substrate under all three. When the audit uses those same layers, every finding points to a specific part of the site and a specific fix, and the measurement framework and the theory behind the website are the same thing.
Start With the Four Layers
Peppermint Global describes a modern website as four layers, each read by a different audience. Surface is the visible website: brand, visual design, interface, and the signals that build human trust, and its reader is a person. Structure is the experience beneath the surface, covering interaction, navigation, wireframes, and the logic of user flow, and its reader is a person moving through the site. Data is the operational layer of business logic, systems, integrations, and core information, and its reader is other systems. The Carcass is the substrate under all three, read by AI models and agents. Peppermint Global describes The Carcass as the machine-legible layer that connects the whole system and makes its meaning intelligible to AI, a deliberately engineered architectural layer grounded in design, linguistics, semiotics, and information structure.
Because The Carcass renders the other three layers intelligible to AI, it is not a separate topic added at the end of an audit. It is where you check whether the meaning created in the Surface, Structure, and Data layers actually reaches the models that read the site.
| Layer | Reader | What it covers | Core audit question |
|---|---|---|---|
| Surface | People | Brand, visual design, interface, trust signals | Can a person quickly understand what this page is about and trust it? |
| Structure | People and UX | Interaction, navigation, wireframes, user flow | Can a visitor move from a claim to its proof without getting lost? |
| Data | Systems | Business logic, systems, integrations, core information | Is the core information accurate, connected, and the same everywhere? |
| The Carcass | Models and agents | The machine-legible substrate under all three | Does the meaning of the other layers survive when the surface is gone? |
The Test Underneath Every Layer: Does Meaning Survive De-Rendering?
One question holds the whole audit together. Strip a page of its visual surface: the layout, the colors, the imagery, everything a person uses to make sense of it. What is left is structure and data. Does the brand still resolve as a distinct entity? Do the product, its purpose, and its relationships still hold together?
That is the de-rendering test, and it sits at the center of the Carcass Method. Peppermint Global puts it simply: the test of good design is no longer only how it looks, but whether its meaning holds when the surface is gone. Each layer below is a specific way of asking that question.
Layer One: Surface
The Surface audit checks whether the visible website earns a person’s understanding and trust. On your most important pages, start with the first screen: does it say what the product does, who it is for, and what the visitor should do next? Then look at whether brand, visual design, and interface signal credibility, backed by real proof such as customers, reviews, or case studies. Key information should be written as text rather than locked inside images, video, or animation, and the visual hierarchy should match the importance of the content.
Evidence to collect: screenshots, short first-impression tests with people outside the company, and notes on any claim that only appears visually.
Layer Two: Structure
The Structure audit checks the experience beneath the surface: how people interact with the site and move through it. Navigation and user flows should lead from a claim to the features, integrations, use cases, and proof behind it. Each important topic needs one clear page, rather than several pages competing for the same question. Interactions such as tabs, accordions, and forms should keep essential information reachable, including with a keyboard and on mobile, and important pages should be linked from relevant pages with descriptive anchor text so that no page is left orphaned.
Evidence to collect: a page inventory, user flow maps, an internal link crawl, and a map of which page answers which buyer question.
Layer Three: Data
The Data audit checks the operational layer: the business logic, systems, integrations, and core information behind the experience. Contradictions here confuse visitors and make it harder for any system to describe the company correctly. Company, product, feature, and integration names should be identical across pages, documentation, and profiles. Pricing, plans, availability, and technical requirements should come from one source of truth and match everywhere they appear. Integrations and systems should be described the way they actually work, and any structured data should match the visible content of the page.
Evidence to collect: a fact sheet of core information, a list of conflicts found, and the pages where each conflict appears.
Layer Four: The Carcass
The Carcass audit checks whether the meaning built in the other three layers reaches AI models intact. Peppermint Global’s service page breaks The Carcass into four components, and each one becomes an audit question. Semantic architecture asks whether the structure is clear enough for a model to disambiguate without guessing. Entity and referential design asks whether AI resolves what the company is, as a stable entity, rather than only what it claims. Retrieval-ready content asks whether content is written at the passage level, with consistent terms and concepts, so it can be extracted and cited. Machine-legible interaction asks the de-rendering question: does meaning stay intact in the HTML and structure rather than living only in the visuals?
Access is the precondition for all four. Check that the crawlers you care about can reach and index the page; OpenAI, for example, says sites that opt out of its OAI-SearchBot crawler will not be shown in ChatGPT search answers. Check that important text is in the initial HTML; in a Vercel and MERJ analysis, the major crawlers from OpenAI, Anthropic, and Perplexity did not execute client-side JavaScript. Schema markup is part of this layer, but Peppermint Global calls it table stakes: entity clarity, authority, and information gain matter more for citation.
A clear headline that only renders after a script runs is a Surface success and a Carcass failure.
Evidence to collect: URL Inspection results, initial HTML compared with the rendered page, a de-rendered read of key pages, entity and terminology checks, and structured data validation.
Where Outcomes Fit
Outcomes are not a fifth layer. They are the signals that show whether the four layers are working, and they should be reported separately from the audit itself. The most useful signals are impressions from AI Overviews and AI Mode in Search Console’s Generative AI performance report, which Google says rolled out worldwide as of August 31, 2026, for properties with enough impressions; ChatGPT referral traffic, which OpenAI tags with utm_source=chatgpt.com; and a stable set of buyer questions tested in AI tools over time, treated as a sample because answers vary between runs.
A Four-Layer Audit Baseline
Adapt this table to your site. If you score findings, define the criteria so another reviewer would reach the same result from the same evidence.
| Layer | Evidence to collect | Example of a finding |
|---|---|---|
| Surface | First-impression tests, page reviews | Homepage hero describes the technology but not the user outcome |
| Structure | Page inventory, internal link crawl | Integration pages are not linked from the product page |
| Data | Fact sheet, cross-system comparison | Pricing page and FAQ list different plan limits |
| The Carcass | De-rendered read, initial vs rendered HTML | Main product explanation only appears after JavaScript runs |
How to Prioritize Findings
Work through findings in the order of their effect. Fix Carcass blockers first, such as pages crawlers cannot reach or read, because they hide everything above them. Next, resolve Data conflicts, since inconsistent facts undermine every page that repeats them. Then repair Structure gaps so that claims connect to proof, and improve Surface clarity on the pages that matter most. Re-test after each release, and keep tracking outcomes separately.
How to Turn Findings Into a Work Plan
Each finding should name the layer, the URL or template, the problem, the evidence, the expected impact, and a next step. Assign an owner and track status. Grouping the work plan by layer also shows which team owns each fix: brand and design for Surface, UX for Structure, product and operations for Data, and engineering with content for The Carcass.
Conclusion
A readiness audit is most useful when it measures a website on the same layers it is built on. Auditing Surface, Structure, Data, and The Carcass gives every finding a clear place and owner, and the de-rendering test shows where meaning is lost between what people see and what models read.
If your team needs a structured four-layer review of its website, explore Peppermint Global’s AEO-optimized website service.
FAQ
What Are the Four Layers of the Carcass Method?
Surface, Structure, Data, and The Carcass. The Carcass is the machine-legible substrate under the other three that makes the whole system intelligible to AI.
Why Not Audit Technical Access as a Separate Layer?
Because technical access is part of The Carcass. Crawlability, rendering, and markup determine whether the meaning in the other layers reaches machines, so they belong in the same framework rather than beside it.
Does the Carcass Method Produce an Official Readiness Score?
This article describes how to organize a readiness review around the four layers. It is not a Google metric, and any internal score should define its criteria clearly.
Can a Strong Four-Layer Audit Guarantee AI Citations?
No. The audit removes barriers and shows where meaning is lost, but AI systems decide what to cite, and answers vary between runs.
