Open data
The aggregate results behind our AgentScore studies, ready for a spreadsheet, a chart or your own analysis. When you use them, please cite the study as shown below. We never name a site or release results for one.
Online stores: 255 sites, scanned October 20, 2026
600 websites drawn at random from the online shops listed in Wikidata, so the sample leans toward established retailers from many countries. 467 could be scanned, and 255 were stores, with products or a cart. Scored with AgentScore 2026.10-v8. Read the study.
| File | What it holds | Rows |
|---|---|---|
summary.json | Everything below in one file, with the headline figures, the method and the citation | |
bands.csv | How many stores scored in each AgentScore band | 4 |
categories.csv | The median store's share of the points in each AgentScore category | 3 |
problems.csv | The share of stores that fell short on each check, where it was checked | 10 |
leaders.csv | How often the top and bottom quarter of stores passed each check | 6 |
How to cite this data
Ghost Agent Lab, AgentScore benchmark of 255 online stores, scanned October 20, 2026 (scoring 2026.10-v8). https://ghostagentlab.com/blog/agentscore-store-study-2026/
Method and limits
- Every site gets the same automated AgentScore scan of its public pages. It never fills in a form or places an order.
- "Falling short" counts a warning or a failure. Each check counts only the sites where it could run, and the study lists only checks that covered at least 50 sites.
- Results from different scoring versions aren't directly comparable.
- Percentages are rounded to whole numbers, so a column may not add up to exactly 100.
The same figures, one sentence each, are on the statistics page. See how AgentScore is scored for every check and its weight. For everything else in the Learning Center as data, see the content API.