Google research: entity order limits LLM fact recall
A new Google research paper finds that encoding isn’t the bottleneck for LLM fact recall — frontier models including Gemini-3-Pro and GPT-5 encode 95-98% of facts but fail to directly recall 26-34% of them, per Search Engine Journal. The paper ties part of the gap to subject/object entity order: models trained on a fact phrased one way struggle to answer when a question reverses that order, and recall failures account for more than 70% of GPT-5.2’s errors. Extended “thinking” recovered 40-65% of the otherwise-unrecalled facts, per the paper.
Why it matters: If phrasing shapes which subject/object order a fact gets encoded in, how a fact is worded — not just whether it's present — may affect whether an AI answer engine can retrieve it, though the researchers stop short of confirming this as a content lever.
Posted to the wire August 17, 2026.