How NextGenIQ measures selection

The observation surface. Coverage across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews at the answer engine level. Persona-conditioned prompt rotation. Geo and language segmentation. The output is a continuous stream of events tagged with model, prompt, persona, geo, time, source set, and position.

The interpretation layer. Decomposed visibility surface across presence, position, framing, and authority. Bayesian probability calibration on every score. Sample sizes and confidence intervals visible to the customer.

The reasoning layer. Probabilistic estimation of expected lift from each candidate action. Causal reasoning over the dominant lever behind a selection outcome. Constraint filtering against brand policy, budget, content velocity, and risk tolerance.

The verification layer. Every action is tagged with intended mechanism, target outcome, and verification window. Confirmed lift, partial lift, and null lift feed back into the recommendation engine, which uses the feedback to recalibrate over time.

Twenty-one category prompts run against five answer engines on a daily cadence: ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews.

A deterministic parser scans every engine response for brand selection events. Each match is logged with the surrounding context, viewable on click. Quotes only, never paraphrases.

Domain citations count only when the engine actually included a URL in its response. Hallucinated URLs get filtered before they reach the dashboard.

Bayesian probability over the observed mention history. Every score displays alongside its sample size and confidence interval, never alone.

Drops and gains are attributed to observed events on the customer graph. Templated causes are not generated. When confidence is low, the system says so.

Every recommendation carries a target prompt cluster, an expected mechanism, and a verification window. Post-action, the tracking engine measures the actual lift.

Below a minimum sample size, the dashboard shows insufficient evidence rather than a fake score. Honesty over false precision, every time.

Recommendations that produce null lift get demoted from the action vocabulary. The system learns when its hypotheses were wrong, so it stops repeating them.

We separate observation from interpretation

Engine responses, mention events, and citation extractions are stored as raw observations. No paraphrasing, no LLM-generated narratives at the data layer. The customer can always trace any claim back to the source.

We use LLMs as classifiers, not inventors

Mention detection, source classification, and recommendation selection are classification tasks with finite output spaces. The LLM picks from a known vocabulary; it does not generate novel claims.

We ground every causal explanation in observed events

The recommendation engine attributes every drop and gain to specific events on the customer graph. Templated causes are not produced. When the system cannot find a high-confidence cause, it says so explicitly.

We close the loop with verification

Every recommendation is tracked through a verification window. Post-action, the system measures whether the predicted lift happened. Recommendations that consistently produce null lift get demoted from the action vocabulary.

We surface honest uncertainty

Insufficient evidence beats a fake score. When sample size is below threshold or signal is too noisy to interpret, the dashboard says so. Customers never see a confident number that the data does not support.

We make every claim auditable

Every score, mention, citation, and recommendation links back to the underlying data points that produced it. If the chain breaks, the claim does not display. Hallucinations cannot survive a provenance check.