AI visibility methodology¶

AI answers are a real local-discovery surface. They are also the easiest surface to measure badly.

This page explains what the scorecard records, what it refuses to pool, and why a single mention is not a score.

Surfaces¶

V1 supports:

Platform	Surface	`source_type`	Method
Google	AI Mode	`consumer_surface`	SERP AI Mode observation
Google	AI Overview (future)	`consumer_surface`	Not implemented in V1; validate the request shape first
ChatGPT	ChatGPT via LLM Mentions	`model_api`	LLM mention observation (US/English only per provider)
Gemini, Perplexity, Claude	Future validated adapters	`model_api`	Not in V1

A consumer/search surface observation is not the same as a model/API response. They are recorded separately and never pooled into one opaque metric.

The ChatGPT adapter refuses a market or language the provider does not support. A US-only observation is never silently reused for a business in another country; the audit records the limitation or returns INSUFFICIENT EVIDENCE if the required surface cannot be collected.

Only validated surfaces contribute to the core score. Experimental sources can be reported as diagnostics without score impact.

What is recorded¶

Every AI observation records:

{
  "platform": "google",
  "product_surface": "Google AI Mode",
  "source_type": "consumer_surface",
  "method": "serp_ai_mode",
  "provider": "dataforseo",
  "prompt": "Who can help with an emergency plumbing problem in Example County today?",
  "market": "Example County, CA",
  "language": "en",
  "repeat": 2,
  "model_version": "available-model-version",
  "target_mentioned": true,
  "target_recommended": false,
  "target_cited": false,
  "target_entity_accuracy": "accurate",
  "cited_domains": ["yelp.com", "example-county-news.example"],
  "competitor_mentions": []
}

The three target fields are separate:

Mentioned — the business appears in the answer.
Recommended — the answer recommends it, not merely lists it.
Cited — the answer cites its domain or entity as a source.

A mention inside a directory list is not a recommendation. A recommendation without a citation is not a citation. Do not reduce AI visibility to "AI visibility": "yes".

Repeats¶

AI answers vary. The framework does not pretend otherwise.

Each prompt-platform pair is observed R times. The default is three repeats, determined by fixture work and documented in configuration. Every repeat is recorded separately.

Aggregation rule:

pair mention/recommendation/citation is true when it appears in at least ceil(R/2) repeats;
repeat stability is the share of pairs where all repeats agree on mention status;
unstable prompts are reported separately.

A live audit should increase repeats before drawing a firm conclusion. A single run is a snapshot, not a trend.

Metrics¶

Metric	Definition
Mention coverage	Share of validated prompt-platform pairs where the target is mentioned
Recommendation coverage	Share where it is recommended
Citation coverage	Share where its domain or entity is cited
Prompt-intent coverage	Share of canonical prompt intents with at least one mention or recommendation
Repeat stability	Share of pairs where every repeat agrees on mention
Surface breadth	Share of validated surfaces with at least one mention
Competitive share	Target recommendations ÷ target + business-competitor recommendations

Directories and publishers are excluded from the competitor denominator. They are recorded as SERP occupiers and citation sources, not as contractors.

Top cited source domains are counted from the per-pair citation lists and shown in the full report. They are diagnostic context, not a separate score.

Where prompts come from¶

AI prompts derive from the same intent model as the search queries. They are not a separate GEO list.

INTENT: water heater repair
Search queries:
  water heater repair
  water heater repair Example County
  water heater not working
AI prompts:
  Who repairs tankless water heaters in Example County?
  Which local company should I call when my water heater stops working?

The prompt templates are versioned, deterministic where possible, reviewed, tied to intent IDs, and locked before results collection.

Judgment¶

Most AI checks aggregate structured booleans. Three checks require review:

LV-AI-04 — service relevance;
LV-AI-05 — location relevance;
LV-AI-06 — entity accuracy.

The judgment layer accepts a score, evidence, and source references. It cannot invent a source ID or rewrite a deterministic fact. See Judgment layer.

What the score does not claim¶

It does not guarantee that an AI recommendation produces a visit, a lead, or a job.
It does not treat one AI answer as a durable ranking.
It does not pool ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode into a single "AI score."
It does not claim that a citation is an endorsement.
It does not use a model's free-form prose as the score.

AI referral traffic and AI-origin leads are owner-data questions. They are reported separately and never reduce the public score.

Reproducibility¶

The audit records the prompt, platform, surface, method, provider, market, language, repeat, and model version where available. The verifier checks that the observation's source_type matches its surface and that consumer/search and model/API observations are not pooled.

A historical AI score can be reproduced from the audit artifact and the rubric. The external system may produce different answers tomorrow; that is acceptable when provenance is preserved.

Scorecard family boundary¶

This page covers discovery. The GBP Scorecard diagnoses profile trust and action; the Website Scorecard diagnoses website conversion. Read The Scorecard Family for the ownership matrix and the no-duplicate-scoring rule.