> Quick Summary: Recent Google research reveals that frontier AI models encode 95% to 98% of tested facts, yet fail to directly answer questions about 26% to 34% of them. The bottleneck is not missing training data, but “subject/object entity order”—meaning models struggle to retrieve facts when a user’s query reverses the directional phrasing of the original source text.
A model can know a fact and still fail to answer a question about it
Google published research this month with a finding that cuts against the usual assumption about why AI chatbots get facts wrong. The instinct is to blame missing training data: the model never saw the fact, so it guesses. Google’s data says that’s mostly not what’s happening. Testing frontier models including Gemini-3-Pro and GPT-5, the research found the models had actually encoded 95% to 98% of the facts being tested, near-total coverage, and still failed to directly answer questions about 26% to 34% of them. The bottleneck isn’t knowledge. It’s retrieval.
Subject first, object second
The mechanism Google identified has a name that sounds more academic than it needs to: subject/object entity order. When a model learns a fact from text, it encodes the entities in the order the source text presented them. The entity that appeared first in that sentence is the “subject”; the one that follows is the “object.” The research’s example is a music-trivia fact: “Oasis played their first gig at the Boardwalk club.” Trained on that sentence, the model has Oasis as the subject and the Boardwalk club as the object, in that order.
Ask the model a question in the same direction the fact was learned, something like “where did Oasis play their first gig,” and it answers correctly most of the time, because the query matches the order the fact was encoded in. Reverse the direction, asking something closer to “what band’s first gig was at the Boardwalk club,” and recall accuracy drops substantially, even though the underlying fact and the underlying knowledge are identical. The information is there. The retrieval path the model needs to reach it from that direction isn’t reliably built.
Why recognition outpaces recall
The detail that makes this a retrieval problem rather than an ignorance problem is what happens when the same models are given the answer as one of several options instead of being asked to produce it from scratch. In multiple-choice format, the models correctly recognize the right answer far more often than they can generate it unprompted in the reverse direction. Recognition and recall are different cognitive operations even for a language model, and this research shows they can diverge sharply depending purely on which direction a fact needs to be pulled from, not on whether the model was ever trained on it.
Why this matters for anyone trying to get cited by an AI answer
If a business’s content only ever states a fact in one direction, the model trained on that content may only reliably retrieve it from that same direction. A hosting company’s page that always writes “AllCloudHost offers KVM VPS hosting with NVMe storage” but never phrases the reverse relationship, something closer to “NVMe storage is included with AllCloudHost’s KVM VPS plans,” may be feeding the model a fact that’s hard to retrieve when a user’s question is phrased starting from “NVMe storage” instead of starting from the company name. The fact is technically present either way. Whether an AI answer engine can actually pull it back out at the moment someone asks depends on whether the phrasing lines up with how the question gets asked.
This is a genuinely different lever than the ones most AI-SEO advice already covers, like adding schema markup, front-loading direct answers, or citing sources. Google’s research is not proposing a formal fix, and the article covering it is explicit that the practical implication, ordering entities to match likely query patterns, is inferred rather than a tested recommendation from Google itself. That distinction matters: this is a real, measured finding about how retrieval fails, with a plausible but not yet proven mitigation, not a confirmed best practice with a checklist behind it.
Adapting content strategy for bidirectional retrieval
Short of waiting for a formal best-practices guide that doesn’t exist yet, the practical move is to state important facts in both directions somewhere on the page rather than assuming one phrasing covers every way a user might ask. A page about VPS hosting plans benefits from stating the relationship starting with the product name (“AllCloudHost’s VPS plans include…”) and also starting with the specific feature or spec a searcher might lead with (“NVMe storage and KVM virtualization are included in every AllCloudHost VPS plan”). It reads slightly redundant to a human skimming quickly, but redundancy in phrasing direction is exactly the gap this research says matters for AI retrieval, and it costs a sentence or two, not a rewrite.
FAQ sections are a natural place to apply this without making the main body repetitive, since a well-built FAQ can pose the same fact as a question from multiple directions (“What does AllCloudHost’s VPS plan include?” and “Which hosting plans include NVMe storage?”) without reading as unnatural the way repeating it mid-paragraph would. Comparison tables help for the same reason: a table row pairing a feature with a product name effectively states the fact from both directions at once, in the exact terse form the research’s own experiments show models handle well in recognition tasks.
None of this replaces the fundamentals of writing content that answers a real question clearly. What Google’s research adds is a specific, measurable reason why a fact you’ve written accurately can still be invisible to an AI answer depending on which way the question comes in, and a low-cost way to hedge against that: don’t just get the fact right, state it from more than one direction.
Where this shows up most: comparison and spec pages
The pattern is easiest to see on the kind of page a hosting company writes constantly: a plan comparison or a spec sheet. A table naturally states a fact from both directions at once, which is exactly the format the research’s own multiple-choice tests suggest models handle best.
| If the question starts with… | The answer needs to connect back to… |
|---|---|
| “What does AllCloudHost’s VPS plan include?” | NVMe storage, KVM virtualization, dedicated IP |
| “Which hosting plan includes NVMe storage?” | AllCloudHost’s VPS plan |
| “Does PrestaShop hosting include a free SSL certificate?” | Yes, AllCloudHost’s PrestaShop hosting |
| “What comes with AllCloudHost’s PrestaShop hosting?” | Free SSL certificate, optimized PHP config |
A prose paragraph tends to only state one of those two directions, whichever felt more natural to write. A table row states both, because a table doesn’t have a “starting” entity the way a sentence does; a reader (or a model) can enter from either column. That’s a low-effort structural fix that happens to align with what the research found actually matters for recall, without requiring anyone to think in terms of “subject” and “object” while writing.
Evaluating the implications for search visibility
This research is new, and the article covering it is upfront that the practical fix, restating facts in both directions, is an inference from the mechanism Google described, not a benchmarked recommendation Google itself has published. It’s a reasonable hedge given what’s known, not a guaranteed ranking or citation boost. Treat it as one more input alongside the more established practices (clear direct answers, genuine citations, accurate structured data) rather than a silver bullet that replaces them.
The underlying takeaway holds regardless of how the specific mitigation eventually gets validated: near-total knowledge coverage inside a model doesn’t guarantee a reliable answer at query time, and the direction a fact travels from source text to a user’s question is a real variable in whether that answer surfaces at all. For a business whose visibility now depends partly on AI answer engines citing it correctly, that’s worth building into how content gets written going forward, not treated as a one-time audit.