AI accuracy — how well data recognition works

The Accuracy report measures the quality of AI data recognition, primarily the extraction of product attributes from descriptions and files.

Step by step

Open the report and set the date range

1. In the sidebar, under Dashboard, expand Reports → AI and select Accuracy.
2. The screen is titled AI Extraction Accuracy, with the subtitle Monitor AI extraction performance and confidence scores.
3. The information icon beside the title explains two terms: Confidence score — Time-weighted accuracy. Higher scores mean more reliable extractions. — and Auto-apply — Enabled when confidence >= 95% and there are at least 20 samples.
4. Set Date range to Last 7 days, Last 30 days (default), Last 90 days, This year, or Custom (which displays From and To). Click Refresh to recalculate.

The date range applies only to part of the screen. The tiles, Accuracy trend chart, By source, and AI Models tab use it. Attribute details, Issues, the Attribute accuracy chart, and the Best attributes and Need improvement cards always use the last 30 days, regardless of your selection.

Read the metrics

1. Six tiles each have their own caption: Accuracy (Accepted without changes), Acceptance rate (Accepted + Modified), Total extractions (In period), Ready for auto-apply (High-confidence combinations), Avg. response time (Time to respond), and Rejected (False positives).
2. Below them are Accuracy trend (two series: Accuracy and Acceptance rate), By source, and Attribute accuracy, with a 90%+, 70-90%, <70% legend.
3. The Attribute details tab is a table with Attribute, Total, Accepted, Modified, Rejected, and Accuracy columns. This is where you can see that a few specific attributes are dragging down the overall result. Rows are not clickable; there is no detail dialog.
4. The Issues tab opens with the warning Attributes with a high rejection rate — consider improving prompts or training data for these attributes. Its columns are Attribute, Total, Rejected, and Rejection rate. If nothing stands out, it says No issues detected!
5. The AI Models tab compares models: AI Model, Extractions, Accepted, Rejected, and Accuracy. A model with no name is labelled (Unknown model).
6. At the bottom of the Overview tab are two cards for a quick takeaway: Best attributes and Need improvement.

Export results

1. Click Export in the upper-right corner.
2. Choose Excel (.xlsx) or PDF report. The export uses the selected Date range.
3. The file is named ai_accuracy_report.xlsx or ai_accuracy_report.pdf, and its contents — headings, sheet names, and report title — are in English, regardless of the interface language.
4. If nothing happens after you click, the file was not downloaded: a failed export does not show any message; only the spinner runs briefly.

Empty report despite having permission

1. Symptom: the screen shows No accuracy data with the text Accuracy statistics will appear here once you start using AI extraction and rating the results and a Go to products button, even though you know extractions have taken place.
2. Go to Configuration → Team and access → Roles and permissions and open the role with the pencil icon (Edit role).
3. On the Module permissions tab, check Read access to the Integrations module; it protects this report’s data.
4. On the Field permissions tab, under Report, the AI - Accuracy entry determines only whether Accuracy appears in the menu. (This entry has no Polish translation and stays in English; its counterpart is AI Usage.)
5. Select the missing field, click Update role, and confirm the code in the 2FA verification required dialog.
6. Without the Integrations module permission, the report opens but stays empty; no access error appears. The Usage report does not have this mismatch; its protection is consistent.

Metrics

There is also a trend chart and three tabs: overview, attribute details, and issues, with a list of false positives. Results can be exported.

Accuracy versus acceptance

These two metrics are easy to confuse, but they mean different things:

• Accuracy — whether AI recognised the data correctly.
• Acceptance — whether a person accepted it.

High accuracy with low acceptance usually means the suggestions are correct but unnecessary, or that the approval process is too cumbersome and people reject items in bulk.

The reverse is worse: high acceptance with low accuracy means errors are getting through.

Attribute details are the most useful part

The attributes tab shows performance per attribute. Usually, a few specific features turn out to be dragging down the overall result; those are the ones to improve in prompts or exclude from automatic application (Configuring the assistant — integration, model, prompts, and costs).

Empty report despite a permission grant

This screen has a permission mismatch:

The menu item requires permission for the AI accuracy report, but the data is fetched under permission for the Integrations module.

A role with only the report permission will see the menu item and an empty screen, without an access error (Report permissions — two layers and four mismatches).

The usage report is protected correctly; this mismatch affects accuracy only.