Why We Simulate AI Retrieval Without Calling a Single AI API
SiteNexis models retrieval quality using deterministic algorithms on measurable content signals — not live queries to AI systems. Here is why this design choice makes the results more trustworthy, not less.
The most intuitive way to know how AI systems retrieve your content would be to ask them. Submit your pages to ChatGPT, Perplexity, and Claude, observe which content gets cited, and use the results to diagnose retrieval failures. This sounds scientific. In practice, it is the worst possible approach to building a reliable AI visibility analysis system — and SiteNexis deliberately does not do it.
Why Live AI Queries Are the Wrong Method
Live AI queries fail as a measurement method for three reasons. First, they are not reproducible: the same query to the same AI system produces different results on different runs because of temperature, context, and constantly updated retrieval indexes. Second, they are not auditable: when a live query produces a citation, you cannot determine which measurable content signal caused it. You see the output, not the mechanism. Third, they are not privacy-safe at scale: submitting client website content to third-party AI APIs for analysis creates data sharing obligations that would be unacceptable for any business domain.
What Algorithmic Simulation Provides Instead
SiteNexis measures content signals that AI retrieval systems are known to process: chunk boundary quality, entity clarity scores, factual density ratios, boilerplate ratios, schema alignment, context window positioning, and trust signal density. The Retrieval Quality Score is computed from these measurements using a deterministic formula. The same content always produces the same score. The score is fully decomposable into its components. Every deduction maps to a named issue with a specific recommendation. This is what a measurement system should provide — not the stochastic output of a black-box query.
●All SiteNexis scores are labelled as estimates in the UI. We do not claim to know exactly how any specific AI system internally weights retrieval signals. We model probability from measurable content characteristics. The distinction between "this is what we measured" and "this is what the AI decided" is communicated explicitly everywhere in the product.
Determinism as a Feature
Determinism in scoring is not a limitation — it is a design principle. A score that changes every time you measure it is not a score; it is noise. A score that stays constant for the same content and moves predictably when the content changes is what makes improvement trackable. When you improve your chunk boundary quality, your Chunk Stability Index goes up. When you add sameAs links, your external validation score improves. When you fix entity inconsistencies, your Machine Trust Score increases. The change is attributable, measurable, and directly connected to the specific action that caused it.
When AI APIs Are Used
SiteNexis does use Anthropic's Claude API in specific, bounded contexts where algorithmic analysis is insufficient. Contradiction detection across semantically similar but not textually identical claims requires natural language understanding. Complex entity disambiguation in ambiguous content requires semantic reasoning. AI extractability scoring — measuring whether a human-readable passage is also machine-comprehensible — benefits from a model that can assess both simultaneously. In every case where the Claude API is called, the result is cached by content hash, the prompt template is in the codebase (not inline), and the output is parsed through a typed schema parser. The AI assists the analysis; it does not replace the measurement.