How we measure AI search visibility

GEO Tool measures two different things, and it’s important to understand both:

1. AI Readiness (the free audit). A deterministic, heuristic score based on signals we extract directly from your public page — robots.txt rules, structured data, content extractability, freshness signals, and technical health. This runs in seconds and costs nothing because it requires no live AI API calls. It tells you whether your page is positioned to be found by AI crawlers and parsers.
2. Measured Visibility (paid tier). Live grounded queries sent to the actual AI engines (ChatGPT, Perplexity, Gemini, and Claude) to measure whether your brand is actually appearing in answers. This is a retrieval-based measurement, not a training-memory check. See Section 4 for the full detail.

The distinction matters. A Readiness score tells you about the technical prerequisites for AI visibility; Measured Visibility tells you the actual outcome. Think of Readiness as “is the door unlocked?” and Measured Visibility as “are people actually walking through it?”

The Readiness score aggregates six signal categories. Each is independently measurable from your public page with no AI API call.

Each AI engine weights the six signal categories differently because they have genuinely different retrieval architectures. Rather than presenting a single “AI score,” we compute a per-engine Readiness score and then average them for the overall headline number.

ChatGPT weights crawler access highly because 87% of its citations come from Bing’s top-10, and Bing requires OAI-SearchBot access. Perplexity weights freshness and authority most because its six-stage RAG pipeline aggressively filters for recency and factual density. Gemini most values structured data and authority signals, reflecting Google’s E-E-A-T framework. Claude puts extractability first because it is the most selective engine, with 97.8% of citations from established, clearly structured sources.

Any engine that is blocked in robots.txt receives a 55% penalty on its blended score (the score is multiplied by 0.45), reflecting that blocked engines cannot index or cite the page at all.

The overall Readiness score is the arithmetic mean of the four per-engine scores, each clamped to [0, 100].

Readiness scores are a necessary but not sufficient condition for AI visibility. Measured Visibility goes further: we send real queries to the actual AI engines and measure whether your brand appears in the grounded, cited answers.

Why grounded queries, not training-memory probes

For all major consumer AI engines — ChatGPT Search, Perplexity, Google AI Overviews, Copilot, and Claude with web search — answers are grounded in live web retrievalat query time. Training memory decides whether an AI “knows” a brand exists; live retrieval decides whether it cites that brand in answers. GEO is primarily a retrieval problem.

We do not use base-model prompting (asking the AI what it “knows” about a brand without web search). That tests frozen training memory, is highly non-deterministic, prone to hallucination for less-known brands, and does not reflect what users actually see in grounded consumer engines.

Query set design

We build a curated query set covering three buckets: discovery queries (category-based, no brand names), comparison queries (evaluation-stage), and brand-direct queries. The mix is roughly 60/30/10 — discovery is weighted highest because that is where new audience is captured. A minimum of 30 queries per audit provides meaningful signal; the full standard audit uses 50.

Repetitions and statistical treatment

A single run of a single query is nearly meaningless. LLM non-determinism means the same query can produce different citations across runs, even at temperature=0 (batch-size variance at the GPU level cannot be eliminated). The Princeton GEO paper found 60–100 repetitions are needed to stabilise visibility percentages.

We run a minimum of 5 repetitions per query per engine. Citation rate is calculated as the fraction of runs where the brand was cited (brandCited = true). We use the Wilson score interval to compute a 95% confidence interval on this proportion.

API vs. consumer UI fidelity

We use each engine’s programmatic API with web search enabled. This is the industry standard and the only scalable approach, but it introduces known gaps:

• Consumer ChatGPT.com may include personalization and browsing history that APIs do not replicate.
• Gemini API with Google Search grounding and Google AI Overviews share only ~38.5% of top cited domains — a structural gap we explicitly disclose on every Google score.
• Perplexity documents potential divergence between its API and consumer UI.

We flag these limitations in every report that includes a Measured Visibility score. We believe transparency about measurement limitations is more valuable than false precision.

AI visibility scores are estimates, not facts. Three sources of volatility affect any measurement:

1. LLM non-determinism. Even at temperature=0, API responses vary due to GPU-level batch-size variance. A point-in-time score with no repetitions may vary ±30% in a rerun. Our Measured Visibility engine uses multiple repetitions and reports Wilson-score confidence intervals so you can see how stable each number is.
2. Model and algorithm updates. OpenAI, Google, and Perplexity update their models continuously, often without public announcements. A score from six weeks ago may not reflect current reality. We show the measurement date on every score and flag scores older than 30 days.
3. Geographic and personalization variation. The same query produces different answers across locations, languages, and account types. Our measurements are run from standardised locations. Results for other geographies may differ.

The Readiness score (free audit) is much more stable than Measured Visibility because it tests deterministic signals — robots.txt, structured data, HTML content — not stochastic AI outputs. A Readiness score change reflects a real change to your page.

We think intellectual honesty about measurement limits is what separates a useful tool from a convincing-looking score generator. Here is what this tool explicitly cannot tell you:

Our methodology is built on published academic research and primary practitioner studies. The field moves fast; we re-verify key statistics quarterly. Evidence strength labels: Proven = controlled study or replicated across multiple studies; Probable = single study or strong practitioner consensus with plausible mechanism; Folklore = widely repeated, no study.