How AEOArc measures AI visibility

How we collect AI answers

Live collectors configured here: Google AI Overviews (SERP), Google AI Mode (SERP), Bing Copilot (SERP). Available when connected: ChatGPT Search (web), ChatGPT (simulated), Perplexity Sonar, Gemini (Google Search grounded), Gemini (simulated), Claude. Each stored answer records the engine, model, timestamp, region, and collection mode.

Every result is labeled with how it was collected. "Search-grounded" means the engine performed live web retrieval before answering. "SERP-observed" means we recorded what a search results API surfaced (for example an AI Overview), not a private chat transcript. "Model-simulated" means a plain model completion without live search — useful as a directional signal, but weaker evidence, and always labeled as such. Collectors that are feature-flagged off or missing API keys are marked disabled or not configured and are excluded from overall system health.

Live collector snapshot

Overall system health: degraded (0/3 active collectors operational). Named engines below reflect this environment only — we never claim an engine is live when it is disabled or not configured.

ChatGPT Search (web): disabled · ChatGPT (simulated): disabled · Perplexity Sonar: not_configured (available when connected) · Gemini (Google Search grounded): not_configured (available when connected) · Gemini (simulated): not_configured (available when connected) · Claude: not_configured (available when connected) · Google AI Overviews (SERP): unavailable · Google AI Mode (SERP): unavailable · Bing Copilot (SERP): unavailable

How prompts are built

Each project gets a prompt set across 12 buyer-intent types (best-of, comparison, local, pricing, review, alternatives, and more), generated from your business profile and competitors. Paid plans track 35–150 prompts per scan; the free scan runs 10.

Prompts are versioned, so week-over-week comparisons always compare like with like.

How the visibility score works

The 0–100 AI Visibility Score is a weighted composite: mention frequency (35%), citation rate (20%), answer position (15%), sentiment (10%), intent coverage (10%), and competitor gap (10%). Each answer is analyzed for brand mentions, cited URLs, competitor mentions, and sentiment; a deterministic fallback parser handles analysis failures rather than dropping samples silently.

Why we show confidence intervals, not single numbers

AI answers are stochastic: identical prompts return different answers across runs. Published research (arXiv 2603.08924) shows single-run citation metrics are statistically indistinguishable from noise in many cases, and industry studies measured 40–60% of cited domains changing within a month.

So we sample repeatedly, aggregate across runs, and report mention probability with confidence ranges plus a volatility level. When the sample count is too low for a reliable interval, we say so — the report shows an evidence grade (A–F) reflecting sample size, recency, region coverage, and collection modes.

What "verified outcome" means

An improvement is only marked verified when the follow-up measurement clears the statistical confidence check against the baseline — not when a single re-run happens to look better. Everything else is labeled directional.

What we will never claim

AEOArc cannot force any AI engine to recommend your business, and neither can anyone else. AI answers vary by engine, prompt wording, region, and time. We improve the signals that help AI systems discover, understand, cite, and recommend you — and we measure honestly whether it worked.

Known limitations

Live coverage depends on which provider keys and feature flags are configured for your deployment. Engines without keys appear as not configured; engines with keys but feature flags off appear as disabled and do not degrade overall health. Free scans prioritize Google AI Overviews (SERP-observed) when OpenAI billing is unavailable so you still get a real measurement path. Consumer AI apps personalize answers per user, so your customers may see answers that differ from our samples — which is exactly why we report distributions, not promises.

Live system health is published at /status and as machine-readable JSON at /api/public/engine-status.