Methodology
How the GEO Score is measured.
Score model v2 · updated July 2026 · every number below is the one the scoring engine runs in production, published so you can check our work.
The measurement, end to end
A GEO audit is a scored experiment against a fixed question set, not an opinion. The same procedure runs every time, so two audits of the same property are comparable and the difference between them is evidence.
- 1Ask. The questions travelers actually ask — trip purpose, location, amenities, value — are put to each answer engine, with repeat sampling so a lucky answer can’t masquerade as a stable one.
- 2Record. Every answer is stored verbatim and timestamped, with the exact question snapshotted alongside it. Evidence is immutable: editing the question set later can never rewrite what an engine said.
- 3Read. Each answer is parsed for the entities it names, their order and prominence, endorsement strength, sentiment, factual claims, and the sources it cites. Brand matching is deterministic and script-aware — 「文華東方」 and “Mandarin Oriental” resolve to the same property.
- 4Score. The six dimensions below combine into a 0–100 GEO Score under a versioned model. Scores computed under different model versions are never silently compared — a model change is not a visibility change.
The six dimensions
| Dimension | Weight | What it measures | When it can’t be measured |
|---|---|---|---|
| Visibility | 30% | Is the brand named at all across the tracked questions? | Always measured |
| Prominence | 15% | How high does it sit? First pick 100 · top three 75 · listed 50 · buried 25. | Always measured |
| Recommendation | 15% | How strong is the endorsement? Explicit 100 · favorable 66 · named-only 33 · negative 0. | Always measured |
| Accuracy | 15% | Do AI claims about the brand match its verified fact sheet? | Unmeasured when no checkable claims exist |
| Authority | 15% | Do answers naming the brand cite trusted sources (incl. TripAdvisor, Michelin, Ctrip)? | Unmeasured when answers carry no citations |
| Conversion readiness | 10% | Can a buyer act on the answer — specifics, proof, next steps? | Always measured |
When a dimension can’t be measured, it is reported as unmeasured — never estimated. The composite renormalizes over the dimensions that were measured, and the gap is disclosed through the confidence rating below. A score that quietly invents its missing data isn’t a measurement.
What ships with the score
Share of AI Voice + cohort rank
The brand’s prominence-weighted share of all mentions, and its rank within the competitive set it actually loses bookings to.
First-pick rate, per engine
“Named first in 2 of 8 answers on ChatGPT” — the answer-level view behind the score, reported for every engine tested.
Measurement confidence
Every score carries a high / medium / low confidence rating: low when two or more dimensions are unmeasured or the brand never appears; medium when exactly one is unmeasured; high otherwise. The reason is stated in plain language on the report.
Volatility, from repeat sampling
The same question is asked multiple times; how much the answers move is measured and reported. A single-run audit cannot tell you whether what it saw was signal or luck.
Verbatim evidence + cited sources
The raw answers, quoted in full, with the domains each engine leaned on — so every claim in the report is checkable against the answer that produced it.
Comparable trends
Score-over-time and per-engine deltas compare only live runs scored under the same model version. Demo data is always labeled, and never plotted.
What the GEO Score does not claim
- It does not measure bookings or revenue. It measures presence in AI answers — the upstream step where shortlists form. Connecting visibility change to booking change is a separate exercise, done with a client’s own booking data.
- One audit is a sample, not a census. Engines vary between askings; that is exactly why repeat sampling, volatility and the confidence rating exist, and why conclusions strengthen across re-audits rather than resting on a single run.
- Unmeasured means unmeasured. Dimensions without enough evidence are shown as “—”, and the confidence rating drops accordingly. No dimension is ever backfilled with an assumed default.
- Coverage is what’s listed, nothing more. We test the engines named on the audit page. Engines we don’t yet test are stated, not implied.
- The score changes when the model changes — and that is disclosed, versioned, and never blended into a trend as if the brand’s visibility moved.
See your property measured this way.
Get the free audit