GEO audit: what is checked, what a score means and what it does not
GEO stands for Generative Engine Optimization: the question of whether generative AI systems name a company in buying queries and whether the website gives them usable signals for it. A GEO audit measures both separately. This text explains the measurement using QueryHalo as the example, so you can interpret a score instead of believing it.
QueryHalo is a one-time AI visibility audit: 6 neutral buying questions go to 3 AI models (GPT, Claude, Gemini) and all 18 answers are documented with the competitors they name. The website check is free; the full report costs €29 once (3 audits €59, 10 audits €149), usually takes 1–3 minutes and arrives as PDF and JSON.
Two measurement levels that should not be mixed
The first level is the model answers: do GPT, Claude and Gemini name the company when asked neutrally for providers? That is the actual AI visibility. It can only be observed, not derived from the website.
The second level is the website: does the public homepage give a crawler or model clear, machine-readable and citable information? That is the technical basis. It often explains why a company is not named, but it is no guarantee that it will be.
A serious audit reports both values separately. Whoever shows only a website score has not measured AI visibility; whoever only counts mentions cannot derive actions.
How the AI visibility value is produced
QueryHalo builds six neutral buying and recommendation questions from offer, audience and market. The company name appears in none of them. The same six questions go separately to three model families – currently GPT 5.6 Luna, Claude Sonnet 5 and Gemini 3.5 Flash Lite via Vercel AI Gateway. That yields 18 answers.
For each answer we record: Is the company named (company name or domain)? At which position does it appear in the model's brand list? Which other brands are named? AI visibility is the share of the 18 answers with a mention: 4 mentions out of 18 give 22%. In addition, the report shows per model how many of the 6 answers name the company, because 4 mentions in GPT and 0 in Claude mean something different from 1–2 mentions everywhere.
How the website score is produced
The website check reads the server HTML of the homepage – what a crawler sees without executing JavaScript – and scores eight criteria totalling 100 points: page title (14), meta description (12), primary heading (12), structured entity data in JSON-LD (20), brand name in visible text (12), citable facts (20), reachable robots.txt (8) and llms.txt (2, optional and always granted).
The weighting follows a simple logic: the two signals that give AI systems the most to hold on to weigh most – an unambiguous, machine-readable entity and concrete numbers that can be cited. The thresholds are deliberately low: a title from 20 characters, a description from 80 characters and four facts with units earn full points. The score measures fundamentals, not excellence. The exact rules are in the technical checklist.
The overall score and its weighting
Overall score = 65% AI visibility + 35% website score. Observed visibility weighs more because it is the outcome; the website is a means. A company with 22% visibility and 72 points in the website check receives 0.65 × 22 + 0.35 × 72 = 39.5, rounded to 40.
| Overall score | Band | Typical situation |
|---|---|---|
| 0–34 | low | Few or no mentions; entity data and facts are often missing on the website too. |
| 35–69 | medium | Occasional mentions in one or two models; website basis present but patchy. |
| 70–100 | good | Regular mentions across several models with a solid technical basis – rare for small providers. |
What a score means
- It is a snapshot: measured on one day, with that day's model versions, without live web search.
- It is comparable with itself: the same audit with the same inputs at a later date shows whether the answers changed.
- It is a diagnosis: the gap between visibility and website basis tells you whether the problem is awareness, website structure or both.
- It is evidenced: all 18 answers are open in the report; every number can be traced.
What a score explicitly does not mean
- No ranking: there is no position 1–10 in ChatGPT to “reach”. Answers vary between calls.
- No search volume: the audit does not measure how many people ask these questions. Whoever reports prompt volumes is estimating.
- No causality: a higher website score does not automatically lead to mentions. Large, heavily linked providers are named even with weak websites.
- No live visibility: with web search enabled or in AI Overviews, answers can differ; that is a separate measurement.
- No statement about quality: not being named is the normal case for smaller providers and no verdict on the offer.
How to work with the result
Read the competitive picture first: which providers do the models name in several answers, and with what reasoning? Then check whether your website could deliver that reasoning at all – prices, audience, metrics, comparisons. Implement the actions from the prioritized plan, starting with the low-effort ones. Repeat the audit with identical inputs after a few weeks. Only the second data point turns a score into a statement about development.
Frequently asked questions
Why 65 to 35 and not 50 to 50?+
Because the mention is the outcome and the website a means. A company that is named often despite a weak website has a smaller problem than one with a perfect website and no mention. The weighting is a choice, not a natural constant; that is why the report also shows both values individually.
Can I see the website score without buying?+
Yes. The eight-criteria website check is the free first step; it uses no AI models. The 18 model answers and the plan are part of the paid report from €29.
Why are client-rendered pages underestimated?+
The check reads server HTML. Content that only appears in the browser via JavaScript is invisible to it – much like to many crawlers. If your homepage is built that way, the low score is a hint to render server-side or at least deliver title, description, H1 and JSON-LD in the HTML.
How does this differ from an SEO audit?+
An SEO audit evaluates rankings, crawling, load time and backlinks for search engines. A GEO audit observes recommendations of generative models and checks the signals models need for entities and facts. Much overlaps; a good SEO foundation helps but does not replace observing the answers.