The CanItBuy Index
One number every week: in how many of the 17 reference stores with an obstacle an AI agent reaches the checkout. We count only models that got through the control store S01 in the same week, so the index measures obstacles, not outages.
73out of 100 in week 2026-W43
AI agents reached the checkout in 73% of the obstacle stores on average (5 models).
- reading: correct answers in the reading modes
- 71
The index week by week
Models in the latest reading 2026-W43
| Model | Reached checkout | Obstacle stores |
|---|---|---|
| Qwen 3.8 Flash | 71% | 12 / 17 |
| Amazon Nova 2 Lite | 71% | 12 / 17 |
| Claude Haiku 5.5 | 71% | 12 / 17 |
| Gemini 3.8 Flash | 75% | 12 / 16 |
| GPT-6 Luna | 76% | 13 / 17 |
Weekly archive
| Week | Index | Reading | Confirmed obstacles | Models |
|---|---|---|---|---|
| 2026-W43 | 73 | 71 | 10 | 5 |
Embed the chart on your site
The chart updates itself every week. Paste the code into an article or a presentation; the caption with the source is part of the image.
Dark
<a href="https://canitbuy.com/agent-index"><img src="https://canitbuy.com/api/index/chart?lang=en" width="640" height="330" alt="Chart of the CanItBuy Index week by week, latest 73 out of 100."></a>Light
<a href="https://canitbuy.com/agent-index"><img src="https://canitbuy.com/api/index/chart?lang=en&theme=light" width="640" height="330" alt="Chart of the CanItBuy Index week by week, latest 73 out of 100."></a>The CanItBuy Index, canitbuy.com/agent-index, week 2026-W43. Data licensed CC BY 4.0.
How we calculate it
- We take the browser agent results from the Matrix for one week (Monday to Sunday, UTC), the newest result of every store and model pair.
- A model counts only when, in the same week, it reached the checkout in the control store S01 and was checked in at least 6 of the 17 obstacle stores. That way a model or network outage does not pull the index down.
- For each model we take the share of obstacle stores where it reached the checkout. The index is the average over models, from 0 to 100.
- Reading follows the same rule in the page reading, machine-readable data and voice modes: the average share of correct answers in the obstacle stores, when the control scored at least 4 of 5.
- We archive the current and the previous week every day, and closed weeks stay unchanged. The agent never places an order.
- The index is about our reference stores and the Matrix models, not every store on the internet. It shows the direction: whether agents handle obstacles better than a week before.