Google research: LLMs know 95-98% of facts but cannot recall a third of them

Google researchers published a paper on arXiv titled "Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality", reporting that frontier language models encode 95 to 98 percent of the facts in their training data but fail to recall 26 to 34 percent of those facts when asked directly. Search Engine Journal covered the paper on 17 August 2026, and the detail that matters most for anyone working on visibility inside AI answers is the trigger the researchers identified: entity order, meaning which side of a fact gets named first in the question.

Google also published an explainer alongside the paper, titled "Why does recall fail?". The argument in both is the same. When a model gets a fact wrong, the fact is usually not missing from its weights. It is in there, and the model cannot reach it from the phrasing it was handed.

What the paper measured

The researchers separated two failures that look identical from outside the model. One is a model that never learned a fact. The other is a model that learned the fact and cannot retrieve it on demand. The paper's title is the shorthand for that split: empty shelves against lost keys. Across frontier models, 95 to 98 percent of the facts tested were encoded during training. Direct recall then failed on 26 to 34 percent of them.

The method for proving a fact was stored rather than absent is what makes the result difficult to wave away. When a model could not produce an answer in open form, the same fact was put to it as a multiple-choice question, and the model picked the right option. A model does not select correctly at that rate from a list if the fact was never encoded in the first place. The failure sits in retrieval, not in knowledge.

That leads to the number with the most consequence attached to it. Recall failures account for more than 70 percent of GPT-5.2's factual errors, according to the paper. Most of what gets called hallucination in that model, in the tests run here, is a lookup problem rather than an invention problem.

The Oasis example, and why rewording does not rescue it

The worked example in the paper is the fact that Oasis played their first gig at the Boardwalk club. Oasis is the subject, the Boardwalk club is the object. Asked in the order the fact appeared in training data, the model answers. Reverse the two entities so the question starts from the venue and asks which band played its first gig there, and recall degrades sharply. The multiple-choice version of that same reversed question is still answered correctly, which is how the researchers know the fact never left the model.

Rephrasing has minimal effect, the paper reports. Swapping synonyms, restructuring the sentence, asking in a more natural voice, none of it moves the result much. Reversing subject and object order is what breaks it. That distinction is worth sitting with, because a large share of published content advice about answering questions clearly is advice about wording, and wording is the variable the paper says is not the one doing the damage.

Long-tail and rare facts fail hardest. A fact that appears across thousands of pages has been seen from both directions many times during training, so both routes into it exist. A fact that appears twenty times, always phrased the same way, has one route in, and a question that arrives from the other side finds nothing.

The figures reported

Each number below is from the paper as reported by Search Engine Journal on 17 August 2026. None of it is modelled, averaged or extrapolated here.

Extended reasoning helps, and does not solve it

Letting a model think for longer recovers 40 to 65 percent of the facts it could not retrieve directly. That is a big recovery, and it is not free. Extended reasoning burns compute, which shows up as latency and cost, and no consumer search surface is going to spend a long reasoning budget on every routine question. Search Engine Journal reports that 11 to 12 percent of facts stay unretrievable even with thinking switched on.

The researchers also close off the obvious fix. Scaling training data does not remove the recall bottleneck. More data puts more facts on the shelves. It does not, by itself, hand out more keys. If the retrieval path is what fails, adding to the store does not repair the path.

The SEO inference the paper does not test

This is the part to be careful with. The tempting conclusion is that you should write facts about your brand in the subject-object order that people use when they ask, so the stored version of the fact matches the shape of the query. Search Engine Journal is explicit that this idea is "not a finding in the research paper" and that it is not proven. It is offered there as an intuition that may be reasonable, not as a result. It should be repeated the same way here: it is an inference, and the paper does not test it.

The paper studies recall of facts already encoded during pretraining. It does not test whether changing how a fact is written on a public web page changes how a model later retrieves it. Those are two different experiments, and several links in the chain between them are untested. Whether a given page is in the training data at all. Whether the internal representation preserves the subject and object order of the source sentence. Whether a phrasing choice survives the preprocessing that sits between a crawl and a training run. Whether grounded answers, which read live pages at answer time instead of pulling from parameters, behave like parametric recall at all.

The honest position is that entity order in AI answers is a hypothesis worth testing on your own pages, not a practice with evidence behind it. Writing a fact in both orders costs a sentence. It breaks nothing if it turns out to do nothing. That is a reasonable basis for trying it on a handful of pages and watching what happens. It is not a reason to rewrite a site, and it is not something to put in a client deck as a finding.

Who this affects and who it does not

The exposure is uneven, and the paper points at where. Facts about entities that appear constantly in training data, big consumer brands, listed companies, famous places, have been written from every angle already, so a reversed question still lands. Facts about entities that appear rarely have one canonical phrasing and no second route.

Most B2B firms, regional service businesses, clinics, agencies, local software vendors and non-English brands sit in the rare category. That is a mechanistic explanation for something that has been observed without explanation for a while now: a brand can be present in a model's training data, can be identified correctly when named directly, and still never come up when the question is asked from the category side. Ask a model what a company does and it answers. Ask which companies do the thing and the same company is absent. Under this paper, that is a recall failure, not an absence.

One boundary worth marking. This research is about parametric recall, which is what a model produces from its own weights with no live lookup. AI Overviews, ChatGPT search modes and Perplexity all retrieve documents at answer time for many queries. Grounded answers of that kind are a different mechanism, and the paper does not claim to describe them. Where a system reads your page live, the classic work of being crawlable, current and quotable still governs the outcome.

What a marketer can actually check

Nothing below depends on the untested inference. These are checks that stand on the paper's own findings about which facts are fragile.

• Ask a model the questions from both directions. Name your brand and ask what it does, then describe what it does and ask which brands do it. A gap between those two answers is the recall pattern the paper describes.
• Do the same in Thai and in English separately. A fact that resolves in one language may not resolve in the other, and the coverage gives no basis for assuming they behave alike.
• List the facts you actually need retrieved: what you sell, where you operate, who you serve, what you are certified in. Those are the sentences worth auditing, not the whole site.
• Check whether each of those facts exists anywhere outside your own domain. A fact stated only on your website has one source and one phrasing behind it, which is the fragile profile the paper describes.
• Separate the two visibility problems before you spend on either. Being absent from a grounded AI answer that cites live pages is a crawling and citation problem. Being absent from an unsourced model answer is closer to what this paper is about.

Work on how a brand is described across the wider web sits inside generative engine optimisation, and the crawl and citation side of it sits closer to conventional AI SEO. The paper is a reason to keep those two workstreams labelled separately rather than merging them into one deck.

What the coverage did not say

Being clear about the gaps is part of reporting this accurately.

• No product change was announced. There is no ranking update, no AI Overviews change and no rollout date attached to this research.
• The reporting did not state which languages were tested, so no claim can be made here about Thai-language behaviour specifically.
• The more-than-70-percent error share is attributed to GPT-5.2. It is not presented as a figure that holds across every model.
• The paper does not say how many times or in how many orders a fact needs to appear before both retrieval routes exist. There is no threshold in the reporting.
• Nothing in the coverage says publishers can influence recall. That connection is the inference, and it remains untested.

What this means for Thai marketers

Thai brands are, in the terms of this paper, almost entirely long-tail entities. A Bangkok clinic, a Thai manufacturer or a domestic SaaS product appears in training data a small number of times compared with a global brand, and often in one fixed phrasing. That is the exact profile the paper identifies as failing recall hardest, which offers a plausible reason why a Thai brand can be perfectly well known in its own market and invisible in an English-language AI answer about its category.

The source says nothing about Thailand. It does not test Thai, and no figure in it can be pointed at the Thai market. What can be said is that if the mechanism holds across languages, brands with thin and repetitive coverage are the ones most exposed, and Thai brands are disproportionately in that group.

The practical read is unglamorous and does not depend on the untested part. Getting a brand's core facts stated in more places, by more third parties, in more than one phrasing, is a link building and digital PR problem before it is a copywriting problem. Writing subject and object order to match how Thai users actually ask is a cheap experiment to run on top of that. It is worth trying. It is not worth billing as a proven tactic.

FAQ

Does this mean AI models are making up facts they never learned?

Mostly no, according to this paper. Recall failures account for more than 70 percent of GPT-5.2's factual errors, meaning the model had the fact and could not retrieve it. The researchers confirmed this by putting failed questions back as multiple choice, where the same model picked the correct answer.

Is it worth rewriting pages so facts appear in the same order people search in?

There is no evidence that this works, and Search Engine Journal says the idea is not a finding in the paper. The research covers recall of facts already inside a model, not whether editing a live web page changes what the model retrieves later. Trying it on a few pages costs almost nothing, so treat it as an experiment rather than a recommendation.

Will more content fix the problem?

The paper says scaling training data does not fix the recall bottleneck, because the facts were already encoded. Publishing more pages on your own domain repeats the same fact in the same phrasing from the same source. Getting the fact stated by other publishers, in different phrasings, is a different move, though the paper does not test that either.

Does this apply to AI Overviews and ChatGPT search results?

Not directly. The paper is about parametric recall, which is what a model answers from its own weights without looking anything up. Systems that fetch live pages before answering work through a different mechanism, and the coverage makes no claim about them.

Did Google announce a change to Search because of this?

No. This is published research plus an explainer, with no product announcement, no rollout date and nothing for site owners to configure.

Where this leaves the work

The useful part of this paper is not a tactic. It is a better diagnosis. A brand that is in the model and still never recommended has a retrieval problem, and retrieval problems are fixed by the fact existing in more places and more shapes, not by more words on one domain. If you want to work through what that means for a specific brand, the AI search visibility service page is the place to start.

Source: Search Engine Journal.