Estimated reading time: 10 minutes
There’s something beautifully recursive about asking an AI how it decides what to believe.
That’s exactly what we did. At Harmony Nexus Vitae, our work centers on a deceptively simple question: how do you make a brand visible to AI search engines? Not to Google’s blue links — to the AI systems that are increasingly replacing them. ChatGPT, Gemini, Perplexity, Claude, Copilot. The engines that don’t show you a list of results — they become the result.
To optimize for something, you need to understand how it thinks. So we went straight to the source. We sat down with Perplexity and asked it, plainly: How do AI systems decide what information is trustworthy? What sources do you prioritize? And why?
The answers didn’t just confirm our working theories. They revealed a structured trust hierarchy that has direct, practical implications for anyone working in GEO.
This is The GEO Lab. We run experiments so you don’t have to guess.
The Experiment
The methodology was simple by design. Rather than studying AI behavior from the outside — running test queries and reverse-engineering the logic — we asked the AI directly about its decision-making architecture.
We posed a series of escalating questions to Perplexity:
- How do AI systems measure the reliability of information?
- Are there sources considered to have a certain “objectivity”?
- What specific knowledge bases and structured data sources are used?
- How does human evaluation factor into the process?
What emerged was remarkably candid and far more structured than most marketers realize.
Finding #1: The Four Layers of AI Trust
Perplexity described a multi-layered system for evaluating reliability. Not a single filter — four interconnected mechanisms working simultaneously:
Training data patterns. The models have been trained on massive datasets where they learned to distinguish quality responses from poor ones. This is baked into the model’s weights — it’s not a rule, it’s an instinct.
Grounding against external sources. When an AI has access to the web or structured databases, it cross-references its generated output against those sources. If the model says something that contradicts what reliable sources say, the grounding layer catches it.
Automated metrics. Accuracy rates, hallucination rates, internal consistency — these are continuously measured and optimized. The AI is literally being scored on how often it gets things right.
Human evaluation loops. And here’s what most people miss: real humans are continuously reviewing AI outputs, scoring them for accuracy, clarity, safety, and usefulness. These human judgments are fed back into the system to retrain and refine the model.
For GEO practitioners, this last point is critical. It means that the trust signals we optimize for aren’t just algorithmic — they’re also shaped by human judgment about what constitutes a credible, useful source. Quality content isn’t just a ranking signal. It’s a training signal.
Finding #2: The Trust Pyramid
When we pushed Perplexity on which specific sources it considers reliable, it didn’t give us a vague answer. It described what amounts to a clear trust pyramid. Here it is, from highest authority to lowest:
Tier A — Scientific research and technical documentation. Peer-reviewed journals (Nature, Science, The Lancet, IEEE), preprint repositories (arXiv, bioRxiv), and official technical documentation (W3C standards, RFCs, language documentation). This is the gold standard. If a peer-reviewed paper says it, the AI treats it as near-fact.
Tier B — Public institutions and international organizations. Governments, regulatory agencies, central banks, and bodies like the UN, WHO, OECD, World Bank, European Commission. Official statistics, legislation, health guidelines. Hard data from authoritative bodies.
Tier C — High-reputation media. BBC, The New York Times, The Guardian, Le Monde, El País, and specialized outlets like Financial Times, MIT Technology Review, Wired. Important for current events and context, but the AI knows these have editorial lines and will cross-reference when sources disagree.
Tier D — Educational publishers and reference works. University textbooks, Britannica, Oxford dictionaries, RAE. Consolidated theory and established definitions.
Tier E — Collaborative knowledge bases. Wikipedia, Wikidata, and similar open encyclopedias. Extremely useful for breadth and general context, but treated as “softer” than Tiers A-D because of the inherent possibility of errors or vandalism.
Tier F — Blogs, forums, and social media. Personal blogs, Reddit threads, X/Twitter posts. Valuable for practical, real-world use cases and community knowledge, but explicitly given less weight when contradicting sources from Tiers A through E.
Read that last line again: blogs and social content are explicitly deprioritized when they contradict higher-tier sources. This isn’t speculation — the AI told us this directly.
Finding #3: The Structured Data Backbone
Beyond the trust pyramid, Perplexity confirmed the critical role of structured knowledge bases in AI decision-making:
Wikipedia is embedded in the training data of most major models. It’s not just a source they reference — it’s part of their foundational understanding of the world. When we say that 3% of GPT’s training data comes from Wikipedia, this conversation confirmed why that matters so much.
Wikidata serves as a structured verification layer. Where Wikipedia provides narrative context, Wikidata provides machine-readable facts: relationships, properties, identifiers. “Paris — is capital of — France” as a graph relationship, not a sentence. This is how AI systems verify concrete data points.
DBpedia extracts structured information from Wikipedia and represents it as RDF graphs — another layer of machine-readable knowledge that AI systems use for semantic verification.
And then there are domain-specific databases: GeoNames and OpenStreetMap for geography, UMLS and SNOMED for health, CrossRef and arXiv for academic citations, plus Google’s own Knowledge Graph (built on the now-discontinued Freebase).
The implication for GEO is enormous: if your entity exists in Wikidata with proper structured properties and references, you have a presence in the verification layer that AI systems use to confirm facts. If you don’t exist there, you’re invisible to that entire mechanism.
Finding #4: Art Gets Special Treatment
We specifically asked about how AI handles art and cultural content — a domain we work in deeply. The response revealed an interesting nuance.
For art, the AI distinguishes between three sub-layers:
Hard facts come from museums and official collections (Prado, Louvre, MoMA, Tate), cultural heritage institutions (UNESCO, ICOM), and catalogue raisonnés maintained by artist foundations. These are trusted for dates, techniques, dimensions, provenance, and current collection information.
Contextual interpretation comes from art history manuals, academic journals, and curatorial texts from exhibitions. This is where styles, influences, and symbolic readings live.
Contemporary discourse comes from specialized critics and publications (Artforum, Frieze) and essays by critics or artists themselves. The AI acknowledges this layer is highly subjective and, when there’s disagreement, should present multiple interpretations rather than a single truth.
For anyone working on AI visibility for artists or cultural institutions, this is a roadmap. Your factual data needs to live in museum databases and Wikidata. Your interpretive authority needs to come from academic and curatorial sources. And your critical presence needs to appear in recognized art publications.
What This Means for Your GEO Strategy
This experiment wasn’t academic. Every finding translates into actionable strategy:
Invest in structured data presence. Wikidata entries, Schema.org markup, and consistent entity information across the web aren’t optional — they’re the verification backbone that AI systems rely on. If the AI can’t confirm your existence through structured data, your content has to work much harder to be trusted.
Cite upward in the trust pyramid. When your content references peer-reviewed research, official statistics, or recognized institutional sources, you’re not just adding credibility for human readers. You’re aligning with the same sources the AI uses as ground truth. The Princeton GEO research paper found that citing credible sources increased AI visibility by up to 77%.
Don’t fight the hierarchy — use it. Your blog post won’t outrank Nature in the AI’s trust model. But your blog post that cites Nature inherits credibility by association. Build your content on the shoulders of higher-tier sources.
Multi-source consistency matters deeply. The AI doesn’t trust any single source absolutely. It cross-references. This means your entity information needs to be consistent across your website, Wikidata, Wikipedia (if applicable), Google Business Profile, social media profiles, and industry databases. Inconsistency is a credibility penalty.
Human quality standards drive AI behavior. The human evaluation loop means that content which feels authoritative, well-structured, and genuinely useful isn’t just subjectively better — it’s literally what the AI has been trained to prefer. Write for the human evaluators, and the AI follows.
The Meta-Lesson
There’s something worth noting about the method itself. We didn’t run thousands of test queries. We didn’t use expensive monitoring tools. We asked the AI a direct question, and it gave us a remarkably transparent answer.
This is itself a GEO insight: AI systems are increasingly willing to explain their own reasoning. The era of pure black-box optimization is giving way to something more conversational, more transparent, more human. The best GEO strategies won’t just optimize for AI — they’ll optimize with AI, using these systems as research partners rather than opaque gatekeepers.
That’s what The GEO Lab is about. We ask the questions that matter, document what we find, and share it openly.
Because in a world where AI decides what to trust, understanding that decision is the most valuable intelligence there is.
This is the first installment of The GEO Lab, where we run real experiments on AI search visibility and publish the results. Subscribe to get notified when the next experiment drops.
Have a question you’d like us to test? Get in touch at hello@harmonynexusvitae.com