![]()
Slite tested two AI search agents against three versions of one knowledge base. Clean documents beat the smarter agent every time.
PARIS, FRANCE, November 18, 2026 /EINPresswire.com/ — AI search accuracy depends more on the state of a company’s documents than on the sophistication of the search, according to an evaluation Slite published today as part of its knowledge framework.
Slite ran the same set of questions through two AI search agents against three versions of the same knowledge base: one messy, one put through a single automated cleanup pass with no human review, and one with every error fixed.
On the messy knowledge base, simple AI search answered 50% of questions correctly. One cleanup pass took it to 76%, a gain of 26 points with no change to the search itself. With every error fixed it reached 87%.
The multi-step agent scored 63% on the messy knowledge base, 87% after one cleanup pass and 96% fully fixed. That agent can search repeatedly, judge what it finds and follow links before answering. On an outdated knowledge base, it was still wrong about one time in three.
Jason Phillips, PhD, Staff Engineer at Slite, who has worked in machine learning and language AI for nearly two decades, built the test on WixQA, a public benchmark. He planted four kinds of error common in real company documentation: duplicate pages, newer pages with no link back to the ones they replace, facts updated in one place but not the others, and plain contradictions. Both agents answered an identical question set against all three versions.
Grading was strict. An answer that gave the correct figure next to an outdated one was counted wrong, because that is the answer a person would act on.
The finding runs against the prevailing bet in enterprise search, where companies buy accuracy with more expensive frontier models or with multi-step retrieval architectures. Both raise the cost of every answer, and in this evaluation both left most of the available gain on the table. Cleaning the documents moved the number further.
The full evaluation sits inside Slite’s knowledge framework, which sets out how teams keep documentation accurate enough for AI agents to act on and proposes a metric called the verified answer rate. The underlying knowledge base research is published alongside it.
About Slite: Slite is a self-maintaining knowledge base for teams and the AI tools they use that turns fragmented company knowledge into one maintained source of truth.
Milena Petrovic
Slite
Legal Disclaimer:
EIN Presswire provides this news content “as is” without warranty of any kind. We do not accept any responsibility or liability
for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this
article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
![]()
Media gallery