Accuracy
How often Verifeed gets it right
We run a fixed test set of 30 research questions with known, sourced answers through the same service you use, and publish the totals here: full reports and quick answers separately.
Full reports
Not measured yet.
The first numbers appear once the whole test set has run through the live service. Nothing is shown here before it was measured.
Quick answers
Not measured yet.
The first numbers appear once the whole test set has run through the live service. Nothing is shown here before it was measured.
The test set
- 30 questions: 20 in English and 10 in Arabic, across Saudi Arabia, the UAE, Egypt, Qatar, Kuwait, Jordan, Lebanon, Bahrain, Morocco and Oman. Company facts, regulators, market events and official statistics.
- 89 expected facts. Each has a primary source (the company, a regulator, an exchange, a statistics office, a filing or a law) or two sources on different sites, with a quote checked against the page.
- 12 known conflicts, where sources disagree: a good report marks them disputed and shows both sides.
- 3 traps: claims repeated online that a primary source contradicts. A good report never confirms them.
References checked on 14 Oct 2026.
What these numbers don't tell you
- The references are reference answers, not ground truth. A report that confirms a different figure with stronger sources may be right.
- 5 of the 89 facts have weak references; a miss on those is read with care.
- A confirmed claim outside the expected facts counts neither for nor against: scoring everything else in a report needs labels we don't have.
- An AI model only decides whether a report's claim says the same thing as a reference. The claim's status, set by checking its sources, decides the score.
- 30 questions is a small sample. Your questions may be harder or easier.