A new Google research paper points to a strange gap inside frontier language models: they encode almost everything they read, but they can’t always say it back. Gemini 3 Pro and GPT-5 have 95–98% of tested facts encoded in their parameters, yet they fail to directly recall 26–34% of those same facts when asked a question. The bottleneck isn’t missing knowledge. It’s retrieval — and one of the clearest triggers for retrieval failure is something SEOs can actually influence: the order in which a subject and an object appear in the text a model learns from.
What “parametric information” actually means
Parametric information is everything a model absorbed during training — web pages, books, code, lyrics, documentation, all of it compressed into the model’s weights rather than looked up at query time. The researchers set out to answer a simple question: when a frontier model fails to answer something it was clearly trained on, is that a training gap or a retrieval problem?
Their finding: encoding is saturated, but recall is not. The paper states it directly: “Encoding is saturated; recall is not. For frontier LLMs such as Gemini-3-Pro and GPT-5, factual encoding is near saturation, with 95–98% of facts encoded. Yet these models fail to directly recall 26–34% of the facts, or 11–12% even with thinking.” Recall failures account for more than 70% of GPT-5.2’s errors, and an even larger share in stronger models — which suggests the problem gets worse, not better, as models scale.
Why subject/object order breaks recall
The more useful discovery for SEO purposes is what causes recall to fail. Facts are extracted from source text as a subject entity and an object entity, in whatever order the sentence presented them. The paper defines it plainly: “the subject is the entity that appears first in the text, and the object appears subsequently.” A question whose answer is the object is a direct question; a question whose answer is the subject, reversing that order, is a reverse question.
Google’s own example makes this concrete: “Oasis played their first gig at the Boardwalk club.” Here “Oasis” is the subject and “the Boardwalk club” is the object. If most of the training text that pairs these two entities always puts Oasis first, the model struggles when a query reverses that order — asking, in effect, “which band’s first gig was at the Boardwalk club?” instead of “where did Oasis play their first gig?”
What makes this stranger still: when the reversed pairing is presented as one option among several in a multiple-choice format, the model recognizes it correctly. The fact is in there. It just can’t be pulled out cleanly through open recall in the order it wasn’t trained on. The researchers don’t offer a mechanism for why recognition succeeds where recall fails — they use the gap as proof the knowledge exists but isn’t reliably retrievable.
Rephrasing didn’t help. Order did.
The team also tested whether simply rewording a question changed recall rates. It largely didn’t. What moved the needle was entity order specifically, not sentence structure or vocabulary. That’s a useful distinction for anyone assuming that varied phrasing on a page is what earns AI visibility — phrasing matters less here than the underlying entity relationships do.
Rare, long-tail facts showed a related pattern. The gap between encoding common facts and rare facts was small — models learn rare facts too — but the recall gap for rare facts was noticeably larger. Again, the information was there; pulling it back out under real-world query phrasing was the harder part.
More reasoning helps, but it’s not free
Giving models more “thinking” time recovered a meaningful chunk of otherwise-unrecallable facts — 40–65% of previously unrecoverable answers came back with extended reasoning. That’s a real fix, but an expensive one: more thinking means more compute per query, and the paper flags a second unsolved problem — knowing in advance which queries actually need it. The researchers also note that scaling model size further doesn’t resolve the underlying recall bottleneck. This isn’t a problem more parameters fix.
What this means for how you write facts
Google’s paper is a model-behavior study, not an SEO recommendation, and it stops short of telling publishers what to do about it. But the mechanism it describes lines up with something entity-focused SEO has been arguing for a while: how clearly you structure entity relationships shapes whether a system can act on what it knows about you, not just whether it was exposed to the information once.
The practical hypothesis — untested by this paper, but reasonable given its findings — is that facts about your brand, product, or claims should appear in text in more than one entity order: subject-then-object and object-then-subject, so that a query phrased either way has a matching pattern to draw from. That’s consistent with the case for entity optimization that goes beyond schema markup alone, and with efforts like structured, machine-readable views of how your entities relate to each other, which reduce a model’s dependence on inferring order from prose in the first place.
It also reinforces a point raised in work on building entity authority for AI search: authority isn’t just about being mentioned near the right keywords. It’s about the relationships between entities being stated clearly and repeatedly enough that a model can retrieve them under whatever framing a real user’s question happens to take.
Frequently asked questions
What is parametric information in an LLM?
It’s the knowledge a model absorbed and stored in its weights during training, as opposed to information it looks up live. The Google paper found this store is nearly saturated in frontier models — most tested facts are encoded — but that encoding doesn’t guarantee the model can retrieve the fact on demand.
What’s the difference between a direct and a reverse question?
A direct question asks for the object entity, matching the order the fact appeared in training text. A reverse question asks for the subject entity instead, requiring the model to retrieve the fact in the opposite order from how it was originally presented.
Does rephrasing a question improve recall?
Not significantly, according to this research. Entity order was the factor that affected recall, not how the question was worded.
Does adding schema markup fix subject/object recall issues?
The paper doesn’t test schema markup directly, and it isn’t a study of SEO tactics. But structured, explicit statements of entity relationships — including markup that avoids relying on a single sentence order — are a reasonable way to reduce ambiguity for systems that otherwise have to infer relationships from prose.
Does making models bigger solve the recall problem?
No. The researchers explicitly found that scaling frontier LLMs further does not resolve the recall bottleneck, and that recall failures make up a larger share of errors in stronger models, not a smaller one.