In August 2026, Google Research and the Technion published a study with a quietly unsettling finding: the newest large language models already know almost everything you would test them on — they encode 95 to 98 percent of the facts in a benchmark — yet they still fail to directly recall roughly a quarter to a third of those same facts when you ask. The knowledge is in the model; the model just cannot always reach it. And the single sharpest predictor of whether it reaches a fact is entity order. Every fact links two entities in a fixed order — a subject that appears first and an object that comes after ('Oasis played their first gig at the Boardwalk'). A direct question asks for the object; a reverse question flips the frame and asks for the subject. Models handle direct questions far better than reverse ones, because a fact learned in one direction is not automatically recallable in the other — the long-known reversal problem, now measured on frontier systems including Gemini and GPT-5. This guide explains what subject-object entity order is, what the study actually found and where its numbers stop, why recall is order-sensitive at all, and the honest bridge from a finding about a model's internal memory to how you should write for AI answers — which is narrower and more useful than the hot takes suggested. The practical core: state the facts that matter most about you consistently, and in the direction people ask them, everywhere a model reads you.
In August 2026, Google Research and the Technion published a study with a finding that should reframe how you think about AI answers: the newest large language models already know almost everything a benchmark would test them on, and still cannot reliably say it. In the paper 'Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality,' frontier systems like Gemini-3-Pro and GPT-5 encoded 95 to 98 percent of the tested facts — the knowledge was demonstrably in the model — yet they failed to directly recall 26 to 34 percent of those same facts when asked. The empty shelf is not the problem anymore; the lost key is. And the single most reliable predictor of whether the model finds the key turns out to be entity order.
Every fact joins two entities in a fixed order. One appears first — the subject — and one comes after — the object. Ask for the object and you are asking a direct question; flip the frame and ask for the subject and you are asking a reverse question. Models handle direct questions markedly better than reverse ones, because a fact absorbed in one direction is not automatically retrievable in the other. This is the long-known reversal problem, now measured on frontier models rather than the small research systems where it was first described. This guide covers what subject-object entity order actually is, what the study measured and — just as important — where its conclusions stop, why recall is order-sensitive at all, and the honest bridge from a finding about a model's private memory to how you should write for AI answers. That bridge is narrower than the headlines implied, and more useful for it.
Strip the jargon and a fact is a small, directional statement linking two things. 'Oasis played their first gig at the Boardwalk club.' 'Marie Curie discovered polonium.' 'Kompozy generates and publishes content across many platforms.' In each, one entity is stated first and one second, and the model reads that pair as an ordered relationship, not a symmetric one. The entity that appears first is the subject; the one that follows is the object. That order is not incidental to how the fact gets stored — it is part of the storage. The model does not file 'Oasis and Boardwalk' as a two-way link you can walk from either end; it files something closer to a one-way arrow pointing from the subject to the object.
Question type follows directly from that. A direct question asks for the object while giving you the subject: 'where did Oasis first play?' — you supply Oasis, the model returns the Boardwalk. A reverse question asks for the subject while giving you the object: 'who played their first gig at the Boardwalk?' — you supply the Boardwalk, the model has to walk the arrow backward. Same fact, same two entities, opposite direction of travel. The whole finding rests on this asymmetry: to the model these are not the same question with the answer swapped, they are two different retrieval tasks of different difficulty, and the reverse one is harder. This is a sharper, more mechanical version of the shift from keywords to whole questions covered in how AI search behavior is replacing keywords — here the direction of the question, not just its wording, changes what comes back.
The idea that a model trained on 'A is B' fails to learn 'B is A' predates this study. It was named and demonstrated years earlier on smaller models: a system taught that a certain person's parent is a certain celebrity could not answer the mirror question about who that celebrity's child is, even though a human who knows one fact knows the other automatically. The explanation is that gradient-based training updates the association in the direction it appears in the text; nothing forces the model to also install the reverse edge. For a person, 'is the parent of' and 'is the child of' are the same relationship read two ways. For a model trained one direction at a time, they are two separate things to learn, and only one of them got taught.
The reasonable assumption was that scale would quietly dissolve this — that a big enough model, having seen a fact stated many ways by many authors, would end up with both directions covered. The 2026 study's contribution is to test that assumption on the frontier and find it only partly true. On today's best systems the reversal gap is smaller than it was on the toy models, but it is still there and still measurable: reverse questions, where the sought answer is the subject, are recalled less reliably than direct ones. Popularity helps — facts stated constantly across the web tend to appear in both orders often enough to close the gap — but the long tail of facts that mostly appear in one direction inherits that direction's blind spot. Scale narrowed the reversal problem; it did not retire it.
The study's method is what makes its numbers worth trusting rather than repeating loosely. The researchers built a benchmark called WikiProfile — 2,150 facts drawn from Wikipedia, each probed by ten separate tasks: two that test whether the fact is encoded at all (can the model reproduce it given its training-style context), four knowledge-evaluation tasks, and four multiple-choice variants. That design lets them separate two things that usually get blurred together: whether a model has a fact in its weights, and whether it can actually recall the fact on demand. They ran it across 13 models, from the Gemma 3 family up through Gemini-2.5-Pro, Gemini-3-Flash, Gemini-3-Pro, and GPT-5, so the results span a real capability range rather than a single system.
The headline split is stark. On the frontier models, 95 to 98 percent of the benchmark's facts were encoded — present and reproducible in a training-style context. But those same models failed to directly recall 26 to 34 percent of facts when questioned normally. In other words, up to a third of what the model provably knows, it cannot reliably retrieve on request. The paper's framing is that for advanced models, recall failures account for the large majority of factual errors — the bottleneck has shifted from acquiring knowledge to using it. That is a meaningful reversal of the usual story about AI hallucination: the fix is less 'train on more data' and more 'get the model to reach what it already has.'
Two conditions make recall fail most often, and they compound. The first is reverse framing: when the question asks for the subject rather than the object, recall degrades relative to the direct version of the same fact. The second is rarity: for facts that appear infrequently in training data, the encoding gap stays modest but the recall gap widens sharply — the model still technically knows the rare fact, it just cannot find it. The study also tested a partial remedy. Letting a model use inference-time 'thinking' — extended internal reasoning before answering — recovered 40 to 65 percent of the encoded-but-not-directly-recallable facts, and helped most on exactly the hard cases, the reverse and long-tail questions. But it did not close the gap entirely (11 to 12 percent stayed unrecalled even with thinking), it is computationally expensive, and there is no clean signal for when it is needed. Recognition held up better than recall throughout: shown the right answer among choices, models could often identify it even when they could not produce it unprompted — the classic signature of a retrieval failure rather than a knowledge gap.
It helps to hold the intuition rather than the mechanism. A model's parametric knowledge is not a database you query with a key; it is a web of learned associations, and retrieval means starting from the entities and phrasing in your question and letting activation flow to whatever is most strongly connected in that direction. When your question matches the direction, phrasing, and context in which the fact was learned, that flow lands on the answer easily. When your question diverges — flips the subject and object, uses different words, drops the context the fact usually travels with — the model has to reach the same stored fact from an angle it was not reinforced from, and sometimes the activation simply does not get there. The fact is on the shelf; you approached from the wrong aisle.
This is why the study frames the failure as the query diverging from the training-time context, phrasing, or ordering in which the fact was encountered. Entity order is the most legible instance of that divergence, but it is one case of a general rule: a model recalls best along the grooves its training wore in, and worst when you cut across them. That general rule is the real, portable lesson — reversal is just its cleanest demonstration.
Here is where careful reading matters, because the tempting inference is bigger than the evidence supports. The study is about parametric factuality — facts baked into a model's weights during training and recalled with no external help. Google's AI answers are not purely that. AI Overviews and AI Mode are retrieval-augmented: they run searches, pull live sources, and generate an answer grounded in what they retrieve, complete with citations. A well-written, clearly-stated page can be retrieved and quoted regardless of how some base model happened to store a related fact in its weights. So the loud version of the takeaway — 'reorder your sentences and you will rank in AI Overviews' — overreaches. The finding does not say that, and presenting it as a proven ranking lever would be exactly the kind of first-mover inaccuracy that is worse than a vaguer, truer statement.
The defensible bridge is narrower and still worth acting on. Both stages of an AI answer punish ambiguity about who-did-what. The retrieval-and-extraction stage looks for a passage that cleanly and unambiguously states the fact in the direction the query asked — a page that says 'Acme Plumbing repairs tankless water heaters in North Denver' is a cleaner extraction for that query than one where the same fact is buried, hedged, or stated backward. And to whatever extent a model leans on its own parametric memory to frame, corroborate, or fill gaps in an answer, that memory is order-sensitive in exactly the way the study shows. So the practical instruction is a bias, not a trick: state the facts that matter most about you in clear, subject-first form, phrased the way people actually ask the question, and keep that phrasing consistent everywhere — so neither the retrieval step nor any parametric recall has to work against the grain to attribute the fact to you. That is the same specificity-and-clarity discipline argued in why specific, detailed content earns AI citations and clear, unambiguous messaging as a ranking input, sharpened by a new reason to care about direction as well as clarity.
Translate the bridge into edits and it is unglamorous and cheap. Lead your key claims with the subject, not the object: 'Kompozy publishes to nine platforms' recalls and extracts more cleanly than 'nine platforms are supported by, among others, Kompozy.' Phrase the load-bearing facts the way your audience phrases the questions — mirror the question's subject-object order in the sentence that answers it, so the fact and the query point the same way. State each fact in plain, complete form somewhere prominent rather than assuming it can be reconstructed from scattered pieces; a model that has to reverse-engineer a fact from fragments is doing the reverse-question work that the study shows it does worst. And repeat the important associations across your content and across surfaces, because popularity — the same fact appearing in many places, in the order that matters — is what closed the reversal gap for well-known facts in the first place.
Now the discipline of not overclaiming, because this is a young topic where models will cite whatever confident thing they find. Do not rewrite your whole site around sentence direction; the effect is a marginal bias, and the dominant factors in whether an AI answer names you are still the ones covered in why AI recommends your competitor and content that performs in AI search — genuine relevance, corroboration across sources, and being the most credible answer to the specific question. Do not treat entity order as a substitute for having the substance. And do not present the reordering as a guaranteed lever to clients or readers — the honest framing is that it removes friction the model would otherwise have to overcome, which is a good bet precisely because it costs nothing and aligns with everything else that makes content extractable. Trust the mechanism enough to adopt the habit; do not inflate the habit into a promise.
Read the study as an operating instruction and it says one thing louder than the rest: state the facts that matter about you the same way, in the same direction, many times, everywhere a model reads. Both halves of that are hard by hand. 'The same way' fails the moment a human writes the same claim three different ways across a blog post, a caption, and a video script. 'Many times, everywhere' fails at the volume a single team can produce. Those two failure modes — inconsistency and thin coverage — are precisely the ambiguity the recall and retrieval steps have to fight through. Kompozy exists to remove both, which is why the finding maps onto it almost line for line.
Kompozy is a full AI content generation and multi-platform publishing engine, not a repurposing add-on, and the mechanism-specific leverage is consistency by construction. Every asset it produces is governed by one Persona Brief that pins your core claims — who you are, what you do, for whom, where — in one canonical, subject-first form. So the load-bearing facts about you appear in identical, same-direction phrasing across blog articles, text and image posts, Carousel Posts that restate the specifics in extractable text, and Persona Shorts where an on-camera presenter says the same claim aloud — with HyperFrames keeping the styling brand-exact so it reads as one recognizable source everywhere. That is the opposite of the drift-across-surfaces problem the retrieval step gets tripped up by: the association a model reads on your blog is the same association, in the same order, it reads on your posts and hears in your video.
Then the volume half, which is what turns a clean statement into a well-corroborated one. Autopilot schedules and publishes that spread across the supported surfaces — eight social platforms plus blog and email — from one queue, behind a per-post review gate so a person signs off before anything ships. Because the reversal gap closes fastest for facts stated often, in the order that matters, across many places, the recurring cadence is the mechanism rather than a nicety: it is how a fact about your business goes from stated-once to reinforced-everywhere. Be precise about the limit, because it is what keeps the rest credible — Kompozy cannot reach inside a model's weights or force an AI answer to name you, and no tool can. What it does is make the honest version of this finding executable: your key facts, in clear subject-first form, phrased the way people ask, published consistently and often on every surface an AI reads — so that whether the model retrieves you or recalls you, it is working with the grain instead of against it.
The 2026 recall study reframes AI factuality: the best models already know almost everything you would test them on and still cannot always retrieve it, and entity order — which of a fact's two entities you approach from — is one of the clearest predictors of whether they succeed. Facts learned in one direction are recalled worst when the question flips that direction, a reversal problem that scale narrowed but did not erase. The disciplined lesson for content is not to game sentence order for rankings; AI answers are retrieval-augmented and the substance still decides. It is to remove friction: state the facts that matter most about you in clear, subject-first form, phrased the way people ask, and keep that phrasing consistent and frequent across every surface a model reads — so neither retrieval nor recall has to work against the grain to attribute the fact to you. It is a free habit that aligns with everything else that makes content extractable, which is exactly why it is worth adopting and never worth overselling.
Every fact connects two entities in a fixed order: a subject that appears first and an object that comes after. In 'Oasis played their first gig at the Boardwalk,' Oasis is the subject and the Boardwalk is the object. A direct question asks for the object ('where did Oasis first play?'); a reverse question asks for the subject ('who first played at the Boardwalk?'). Large language models recall facts more reliably when the question follows the same subject-object order the fact was learned in, and struggle more when the order is reversed.
In 'Empty Shelves or Lost Keys?', published on the Google Research blog on August 12, 2026 (with the Technion), researchers built the WikiProfile benchmark of 2,150 facts and tested 13 models. Frontier systems like Gemini-3-Pro and GPT-5 encoded 95-98% of the facts but still failed to directly recall 26-34% of them; even with extended thinking they failed on 11-12%. The conclusion: for the best models, factual errors now come less from missing knowledge than from knowledge that is stored but not reliably retrievable — recall, not encoding, is the bottleneck.
The reversal problem is the finding that a model trained on 'A is B' does not automatically learn 'B is A' — it can answer in the direction it saw the fact but stumbles when the question flips the entity order. The 2026 study confirms a version of it persists in frontier models: reverse questions, where the answer is the subject rather than the object, are recalled less reliably than direct ones, and inference-time thinking recovers many of them. So the effect is smaller than in older models but has not disappeared.
Be careful here — the research is about a model's internal, trained-in memory, not directly about which page an AI Overview retrieves and quotes. AI Overviews are retrieval-augmented: they pull live sources and generate from them, so a clean, quotable page can be cited regardless of how a base model stored a fact. The reasonable, honest implication is narrower: state your key facts in clear, subject-first form, phrased the way people actually ask, and consistently across surfaces — so that both the retrieval step and any parametric recall have the least ambiguity to resolve. It is a bias worth having, not a proven ranking lever.
The study rewards one behavior above all: stating the facts that matter about you the same way, in the same direction, many times, everywhere a model reads. That is a production-and-consistency problem, and it is exactly what Kompozy is built for. It is an AI content generation and multi-platform publishing engine that governs every asset with one Persona Brief — so your core who-does-what claims appear in identical, subject-first form across blog articles, image posts, carousels, and short-form video — then publishes them across eight social platforms plus blog and email on autopilot. It turns 'say it consistently and often' from an intention into an operating system.
Google entity order refers to whether a fact's subject or object comes first — every fact links two entities in a fixed order. A 2026 Google Research study found frontier models encode 95-98% of tested facts but fail to directly recall 26-34%, and recall breaks hardest on reverse questions that flip the subject-object order the fact was learned in. The takeaway for content is narrow but real: state key facts in clear, subject-first form, phrased the way people ask, consistently across every surface.
Get started → · ← All guides · Compare Kompozy vs other tools