Google Study: Entity Order Trips Up AI Fact Recall

A research paper published by Google finds that frontier language models encode 95 to 98 percent of the facts they are tested on, yet cannot produce a direct answer for 26 to 34 percent of those same facts when asked. Search Engine Journal reported the paper Monday, and it isolates a specific pattern behind the failures: questions that reverse the order in which a fact’s two entities first appeared in training. That distinction separates a knowledge gap from a retrieval failure, and the two problems call for different fixes.

Parametric information is what a model stores during training, drawn from web pages, books, code, and everything else in its training set. The researchers set out to explain why models sometimes cannot answer questions about facts they were trained on. They found encoding is close to saturated in frontier systems, meaning models including GPT-5 and, separately, Gemini-3-Pro already hold nearly all the facts being tested. Recall, not storage, is the bottleneck. The paper attributes more than 70 percent of GPT-5.2’s errors to recall failure, a share the researchers say grows in stronger models.

The researchers define subject and object by position in the original training text. Whichever entity is named earlier in that text becomes the subject; the one introduced afterward becomes the object. A question asking for the object is a direct question. A question asking for the subject, reversing that original order, is a reverse question. Google’s own explainer illustrates the idea with a short fact about the band Oasis and the venue of an early gig: naming the venue is easy for a model that learned the fact with the band mentioned first, while naming the band from a description of the venue is harder, even though the underlying fact is stored.

Models can still identify the correct answer to a reverse question when it appears as one option among several in a multiple-choice format, even when they cannot produce it unprompted. The researchers treat that as evidence the fact is encoded and recognizable rather than lost. They do not explain why recognition succeeds where free recall does not.

The team also tested whether reworking the wording of questions changed recall rates, and found it made little difference. What mattered was the entity order, not how else a question was phrased. Rare, long-tail facts showed a related pattern: the gap between how well common and rare facts were encoded stayed small, but the gap in how well they could be recalled grew larger, another sign that retrieval drives the failures more than storage does.

Prompting models to reason longer before answering recovered 40 to 65 percent of the facts they could not recall directly, though at added computational cost, and the researchers flag no established way to know when that extra reasoning should trigger. They also found that further scaling of model training does not close the recall gap, ruling out the simplest brute-force fix.

Roger Montti, who reported the study for Search Engine Journal, floated a related idea for content strategy: ordering subject and object entities the way queries commonly phrase them might help. He is explicit that this is his own hypothesis, not a conclusion the paper reaches, since the research measures recall in general question answering rather than how any search system selects or cites a page.

One way to read the mechanism, stated here as inference rather than as the paper’s conclusion: if a model’s ability to retrieve a fact depends on the direction a relationship was expressed, a page stating that relationship only one way is betting the question will arrive from the same direction. That is a testable idea, not an established ranking factor.

Search teams should not treat this paper as evidence that sentence order changes rankings or AI citations, because it does not test that claim. What it supports is narrower: whether a fact phrased only one way on a page is harder for a model to retrieve when a user’s question runs the other direction, a hypothesis worth checking against a site’s own content before acting on it.

Roger Montti reported this research for Search Engine Journal on August 17, 2026, citing a Google research paper and Google’s own explainer on recall.