Research pilot · October 2026
How stable is an AI visibility score?
We ran the Free GEO Checker AI test five times on each of 35 websites, from global brands to small startups, to see how much one company’s result moves between identical runs. The answer shapes how any GEO score, ours included, should be read.
- One run is not a verdict. Only 7 of 35 sites got the same AI visibility score in all five runs. The score of 28 sites moved more than 10 points, and 6 moved more than 30. The average standard deviation of a site’s five scores was 7.7 points.
- Technical readiness showed almost no link to visibility in this sample. The correlation between the technical score and the average AI visibility score was 0.05. Patagonia and Microsoft scored in the 50s technically and above 90 on visibility; Notion scored 95 technically and averaged 33 on visibility.
- Third-party profiles were the most cited sources: en.wikipedia.org (78), linkedin.com (65), x.com (58), crunchbase.com (41), github.com (23), zoominfo.com (22), trustpilot.com (19), globaldata.com (18).
- 5 of 40 sites refused an automated request outright (HTTP 403) and could not be checked: lego.com, canva.com, oura.com, swappie.com, doctolib.fr.
What we measured
Each run asks Claude Sonnet 5.5 the same two questions used by the free tool: look up the domain on the web and say what it is, then answer from its own knowledge without searching. A second model compares both answers with a summary of the site’s homepage. The run scores 40 points if search found the site’s own domain, 30 if the search answer describes the business correctly, and 30 if the model already knows it. The full method is on the methodology page.
Sites were chosen in four groups of ten: global brands, SaaS scale-ups, startups and small SaaS, and Nordic and French companies. Technical checks were run once per site. 175 runs were completed on 19 October 2026.
Results by group
| Group | Sites | Technical | AI visibility | Found by search | Known without search | Run-to-run spread |
|---|---|---|---|---|---|---|
| Global brands | 9 | 77 | 87 | 82% | 93% | ±5.7 |
| SaaS scale-ups | 9 | 92 | 75 | 87% | 31% | ±6.3 |
| Startups and small SaaS | 10 | 88 | 80 | 92% | 50% | ±8.6 |
| Nordic and French companies | 7 | 93 | 72 | 74% | 43% | ±10.8 |
Averages. “Known without search” is the share of runs where the model’s own-knowledge answer was judged fully correct. Spread is the average standard deviation of a site’s five scores.
Where the variation comes from
- Whether search returns the site itself. Across all runs, search cited the site’s own domain 85% of the time. For 6 sites this flipped between runs. Airbnb and Notion were never cited from their own domain: search answered from Wikipedia and other profiles instead.
- How complete the description is. Whether an answer counted as fully or partly correct changed between runs for 20 sites. Short answers often leave out a product line, and the grader then gives partial credit.
- What the model remembers. The own-knowledge answer changed grade for 17 sites. Well-known brands were stable; younger companies were not.
Technical gaps across the 35 sites
| Check | Sites failing |
|---|---|
| Question-style headings or FAQ markup | 28 |
| No llms.txt (optional, no points) | 19 |
| No organisation structured data (full or partial fail) | 13 |
| Under 300 words visible without JavaScript | 7 |
| AI crawler user agents refused | 3 |
What this means
- Treat a single AI visibility score as a snapshot that can move several points on an identical re-run. Compare results over several runs.
- Technical fixes make a site easier to read. In this sample they did not on their own predict whether assistants knew the company. Third-party profiles (Wikipedia, LinkedIn, Crunchbase, review sites) appeared in most answers; this study records which sources were cited and does not test what causes a mention.
- Monitoring has to repeat questions. A tool that runs each question once can report a change that is only noise.
All sites
| Site | Technical | Visibility range | Found |
|---|---|---|---|
| spotify.com | 71 | 45–100 | 2/5 |
| tldraw.com | 60 | 30–85 | 1/5 |
| wolt.com | 94 | 45–100 | 1/5 |
| supercell.com | 78 | 45–100 | 1/5 |
| airtable.com | 80 | 45–85 | 4/5 |
| relexsolutions.com | 95 | 45–85 | 4/5 |
| hubspot.com | 100 | 70–100 | 5/5 |
| plausible.io | 88 | 70–100 | 5/5 |
| lemlist.com | 100 | 70–100 | 5/5 |
| aiven.io | 98 | 70–100 | 5/5 |
| apple.com | 90 | 85–100 | 5/5 |
| microsoft.com | 54 | 85–100 | 5/5 |
| ikea.com | 95 | 85–100 | 5/5 |
| patagonia.com | 52 | 85–100 | 5/5 |
| nokia.com | 85 | 70–85 | 5/5 |
| notion.so | 95 | 30–45 | 0/5 |
| monday.com | 93 | 70–85 | 5/5 |
| pipedrive.com | 100 | 70–85 | 5/5 |
| zapier.com | 95 | 70–85 | 5/5 |
| miro.com | 83 | 70–85 | 5/5 |
| cal.com | 80 | 70–85 | 5/5 |
| tally.so | 95 | 70–85 | 5/5 |
| posthog.com | 100 | 70–85 | 5/5 |
| resend.com | 95 | 85–100 | 5/5 |
| raycast.com | 80 | 70–85 | 5/5 |
| fathom.video | 88 | 85–100 | 5/5 |
| alan.com | 95 | 70–85 | 5/5 |
| qonto.com | 88 | 85–100 | 5/5 |
| nike.com | 85 | 100–100 | 5/5 |
| airbnb.com | 65 | 60–60 | 0/5 |
| kone.com | 95 | 100–100 | 5/5 |
| intercom.com | 95 | 85–85 | 5/5 |
| figma.com | 87 | 85–85 | 5/5 |
| buttondown.com | 95 | 100–100 | 5/5 |
| mantox.fi | 100 | 55–55 | 5/5 |
Sorted by the spread between the lowest and highest of five runs. Raw data: research-data.json.
Limitations
- One assistant (Claude). Other assistants use other search indexes and will give other results.
- The correctness grade comes from a language model compared against the homepage, not from human reviewers.
- Five runs per site on a single day. Results will drift as indexes and models update.
- The variation measured combines search, answer generation and the grader; the study does not separate them.
- This is a measurement-reliability pilot. It does not test whether any GEO change improves visibility.
- The sample is small and hand-picked; group averages are indicative, not representative of each segment.
Published 19 October 2026. Free GEO Checker is run by Matias Putkonen. Cite as: Free GEO Checker, “How stable is an AI visibility score?”, October 2026.