An approved claim is a reviewed customer-facing statement that can be reused across the website, sales, support and AI answers.

A source of truth is the canonical record a team turns to when pages, decks, docs or AI answers disagree.

Knowledge drift is the gap that opens when published statements, internal documentation and actual product behaviour stop matching.

A retrieval boundary is a rule limiting which sources an AI system may reference when answering for a given audience.

On the FreshQA benchmark, no model answering without retrieval scored above 15% on fast-changing questions.

Adding one outdated passage alongside a current one dropped Llama-3.1-70B’s accuracy from 89% to 73%.

The same test dropped Qwen-2.5-7B from 76% to 34%, and its misleading answers rose from 4% to 22%.

Asked about changed facts, leading models gave outdated answers between 15.7% and 31.3% of the time.

SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts across AI engines, published 28 January 2026.

AI engines name the same brand only about one time in five.

In April 2025, Cursor’s support assistant told users about a device limit that did not exist.

Our Stale Answer Benchmark tests sixteen B2B companies, four claims each, across three AI engines.

Cross-engine disagreement is the share of claims where at least two engines answer materially differently.

This article is labelled a 12-minute read.

This article carries distinct published, updated and last-verified dates.