How we distinguish evidence, explanation and advice
AAI Kernel’s evidence standard separates what was observed, what was calculated, what is inferred and what is proposed. A retained answer can establish its wording and visible references under recorded conditions. Explaining why that answer occurred, judging its business significance or recommending a response requires additional reasoning or evidence. A source citation must support the specific claim made, and essential limits must remain attached when the claim is quoted.
Four statements, four responsibilities
Consider a hypothetical answer that names a provider but recommends an alternative for a stated need. After identity and role validation, the wording is an observation. Counting such cases across eligible answers is a derived measurement. Proposing that unclear service descriptions explain the pattern is a hypothesis. Advising a change to those descriptions is decision guidance. This example contains no client finding or demonstrated improvement.
Measurement records the observed object using declared rules. Inference extends from evidence to a proposed conclusion; interpretation explains the significance of that conclusion. A hypothesis remains testable and provisional. These distinctions prevent an analyst’s plausible account from being reported as something the system itself proved.
A company-capability statement needs current scope authority and, when claiming an operational delivery, evidence for that delivery. A commercial promise additionally needs an accepted commitment. Neither follows from the existence of an internal method or prototype. Client teams and appointed agencies remain responsible for implementation.
The research method places these separate responsibilities within AAI’s investigation process; it does not turn a procedural description into proof that a particular result has been obtained.
What supports an important claim
An empirical statement needs the research question, entity rules, system and surface, retrieval conditions, observation dates, eligible population, method version and supporting capture or derivation. For a rate or share, identify both numerator and denominator. Explain exclusions and missing evidence where they change the reading. A generic Method link cannot supply proof of a result.
A source-dependent explanation should identify what the source establishes and what remains the analyst’s inference. Official platform documentation supports a statement about documented platform behavior; it does not establish that an intervention improved a company’s position. Vendor benchmarks remain bounded benchmarks rather than universal rules.
A definition needs its canonical owner and operational scope. A procedural statement is an editorial requirement, not an empirical discovery or evidence that an automated process is operating. The glossary keeps definitions stable, while measurement and uncertainty explains how to read results.
For an original public finding, the publication standard specifies the accompanying record. For a report already delivered, the evaluation guide applies the obligations to a material claim and its proposed response.
What observation alone cannot establish
A model’s stated reason is observable output about its response, not a verified record of its internal reasoning. A cited source establishes a visible reference. A source declared important by a model establishes a self-report. Behavioral association and causal support are stronger, separate claims requiring suitable comparisons or controlled interventions. More readable explanation does not automatically mean more faithful explanation.
Agreement between systems is a pattern, not a vote on factual truth. Repetition may reflect shared sources or common errors. A change following a website revision does not isolate that revision from model changes, source changes or sampling variation. Causal explanations can be investigated with an appropriate design; they cannot be inferred from timing alone.
Visibility, citation and recommendation measurements do not by themselves prove traffic, revenue, business-market leadership or a guaranteed future answer. Internal counts of records, analyses or model calls are not evidence of client value or accepted delivery. Judge the supported finding and its relevance to the decision.
When the basis changes, review the dependent claims
Review a claim when its evidence is corrected, its method changes or an important source becomes unavailable. Preserve the old observation and method identity. Correct an error, qualify an overstatement or withdraw an invalid result according to its significance; do not silently replace a historical finding with a newly calculated one.
Review dependent summaries, Answer Units, tables, figures, metadata and translations so the same unsupported meaning does not remain elsewhere. The corrections guidance explains how the public record should make a material change visible.
A research handoff should let a decision-maker recover the observation, interpretation, uncertainty and proposed response separately. A useful explanation can identify a strength, a concern or an important unknown without forcing every finding into an action.
Common questions
- Can an AI visibility audit prove content caused growth?
- Not from a visibility audit alone. A causal growth claim requires a design and evidence capable of separating the proposed cause from alternative explanations.
- Does agreement between models establish truth?
- No. Agreement is an observed pattern, not a truth vote. Check the underlying claim against independent evidence and retain counterexamples.
06 Where this leads
A supported finding can still leave the response undecided. The prioritization guide adds business relevance, control and alternatives to the evidence judgment. If the proposed response depends on a before-and-after improvement claim, comparability and change explains what a follow-up can establish and why timing alone cannot identify the cause.