A page can be crawlable but never retrieved. It can be retrieved but not used as evidence. A brand can be cited without being recommended, or recommended without its own domain being cited. Even a recommendation does not prove that a buyer clicked or converted.
These are different states. Treating them as one “AI visibility score” hides the actual constraint and encourages the wrong work. This framework separates the system into 12 layers so teams can diagnose where evidence stops and choose a test that matches the problem.
Four connected systems
The 12 layers
1. Content discovery
A provider needs a path to the content. Internal links, sitemaps, public URLs, and crawler access help discovery. OpenAI recommends allowing OAI-SearchBot for ChatGPT search eligibility; Google requires pages to be indexable and eligible for snippets before they can appear as supporting links in AI features.
2. Crawling and parsing
Access is not understanding. Important claims should exist as readable text, page structure should be coherent, and structured data must match visible content. A blocked CDN or ambiguous rendered page can break this layer without saying anything about brand demand.
3. Entity resolution
The system needs to distinguish the company, product, domain, category, and similarly named entities. Consistent naming and connected first-party pages reduce ambiguity. Entity clarity prevents avoidable identity errors; it does not guarantee selection.
4. Query relevance
Buyer prompts express situations, constraints, and trade-offs—not only keywords. A useful map covers category questions, comparisons, use cases, objections, implementation, and proof. Google says its AI features may issue multiple related searches across subtopics and sources, making narrow keyword matching incomplete.
5. Source usefulness and trust
A source must contribute something the answer needs: a definition, fact, comparison, limitation, method, or first-hand result. Google’s people-first guidance emphasizes original information, clear sourcing, substantial value, and accurate authorship. More pages without more evidence do not solve this layer.
6. Cross-source corroboration
Commercial claims often need support beyond a company’s own page. Independent documentation, customer evidence, public research, and consistent product facts can reduce uncertainty. Providers do not publish a universal weighting formula, so treat corroboration as a testable hypothesis—not a guaranteed ranking factor.
7. Evidence extraction
Even a relevant page may be hard to use if the answer is buried or detached from its conditions. Put the claim, scope, date, method, and limitation close together. Tables and concise sections help readers locate evidence, but formatting cannot rescue weak substance.
8. Citation selection
A citation is evidence attached to an answer. It does not automatically mean the cited company is endorsed. ChatGPT warns that search citations can be incomplete, outdated, or wrong. Measure citation presence and support quality separately.
9. Brand recommendation
A recommendation answers a choice question: which product fits the stated job and constraints? This requires fit, not merely visibility. Record whether the brand was named, described positively, recommended for a situation, or selected as the primary choice.
10. Comparison and ranking
The same product can be appropriate for one scenario and weak for another. A defensible report preserves the prompt and rationale rather than presenting a context-free rank. The Aivius AI Video Enhancement Benchmark shows matched prompts producing different primary picks across providers.
11. Buyer action and attribution
Exposure is not revenue. Track available referrals, landing behavior, qualified actions, assisted conversions, and sales feedback while acknowledging unobservable journeys. OpenAI says ChatGPT referral links include utm_source=chatgpt.com; that is useful evidence, not a complete attribution system.
12. Continuous experiments
Version the prompt set, change one meaningful variable where possible, repeat matched measurements, record provider failures, and compare evidence over time. A before-and-after change is informative; it is not automatically proof of causation.
Use the framework as a constraint map
| Observed problem | Inspect | First test |
|---|---|---|
| Owned pages never appear as sources | Layers 1–3 and 7 | Verify access, rendered text, canonicalization, identity, and extractable evidence. |
| Third parties are cited but the brand is absent | Layers 4–6 and 9 | Compare buyer-intent coverage and independent evidence for observed competitors. |
| Brand is mentioned but not recommended | Layers 8–10 | Test explicit use cases, constraints, proof, and differentiation. |
| Recommendation rises but pipeline does not | Layers 11–12 | Audit referral capture, landing alignment, sales attribution, and competing causes. |
Three evidence classes keep the model honest
- Observed: actual answer, citation URL, provider status, timestamp, and matched text.
- Inferred: a plausible explanation supported by patterns, but not disclosed as an internal platform rule.
- Controlled: variables a team can change—access, content, evidence, distribution, landing experience, and measurement design.
Only observed evidence should become a reported fact. Inferences belong in the diagnosis with confidence and alternatives. Controlled variables become the prioritized experiment backlog.
Sources and further reading
- Google Search Central: AI features and your website.
- Google Search Central: helpful, reliable, people-first content.
- OpenAI: Publishers and Developers FAQ.
- OpenAI: Searching the web with ChatGPT.
- Aivius: Mention vs Recommendation vs Citation.
Find which layer is blocking growth
Start with a directional snapshot, then turn observed gaps into a human-reviewed audit.
Get a free growth snapshot →