AI search news ·
Why does AI know my business but never recommend it? Google Research says frontier models encode 95–98% of facts and fail to recall 26–34% of them — and the reverse question is the hard one
A business owner who asks ChatGPT "is [my business] any good?" and gets a warm, accurate paragraph, then asks "who do you recommend for [my category] in [my city]?" and is not named, has just reproduced — on one business — the finding Google Research published on 2026-08-12. In "Empty shelves or lost keys? Recall is the bottleneck for parametric factuality" (Nitay Calderon and Gal Yona; paper arXiv 2602.14080), the team profiles 2,150 Wikipedia-derived facts across 13 models and separates three things standard accuracy collapses: whether a fact is encoded in the weights, whether the model can recall it unprompted, and whether it can recognise it among alternatives. Frontier models (Gemini-3-Pro, GPT-5) encode 95–98% of the facts, yet fail to directly recall 26–34% of them; with thinking on they still miss 11–12%. Two failure modes are systematic: rare facts are encoded at close to the rate of popular ones but recalled far less, and reverse-direction questions — the model learned "A is B" and is asked "what is B?" — are recognised as easily as direct ones in multiple choice but generated much less often. "The reversal curse is a recall problem." A small business is a rare fact, and a buyer's question is the reverse direction.
What Google Research measured
The instrument is a benchmark the team calls WikiProfile: 2,150 facts extracted from Wikipedia as ordered (subject, object) pairs, each paired with ten tasks — two that probe encoding (can the model reproduce the fact in its pre-training context), four open-ended recall questions in direct and reverse form ("if A is B, a direct question asks 'What is B?', while a reverse question asks 'What is A?'"), and four multiple-choice recognition variants. Questions were generated, refined and filtered by Gemini-2.5-Pro, checked against a search engine to discard anything with multiple answers, then manually validated. Each fact is then classified into one of five profiles — encoding failure, recall failure, direct recall, recall with thinking, inference without encoding — so the unit of analysis is the fact, not the question. The paper covers 13 LLMs; the blog quotes Gemini-2.5-Pro, Gemini-3-Pro and Flash, GPT-5, and the Gemma 3 family. Search Engine Journal carried the story on 2026-08-17.
The numbers
| Finding (Google Research blog, 2026-08-12) | Reading |
|---|---|
| Facts encoded, Gemini-3-Pro and GPT-5 | 95–98% |
| Facts NOT directly recalled, same models | 26–34% |
| Facts still not recalled with thinking on | 11–12% |
| Encoded-but-unrecalled facts recovered by thinking (thinking-optimised models) | 40–65% (vs 5–15% of non-encoded facts) |
| Rare vs popular facts (bottom 20% vs top 20%) | encoding gap "relatively modest"; recall gap "larger" |
| Reverse vs direct questions | no harder (often easier) in multiple-choice recognition; "consistently harder" in open-ended generation |
| Scaling (Gemma 3 family) | encoding failures fall sharply with size; recall failures persist and become a larger share of remaining errors |
The authors' own framing: "Many rare facts are not absent from the model's parameters. They are present, but difficult to access. The bottleneck has shifted from knowledge acquisition to utilization." And on direction: "the fact appears to be encoded, and even recognizable, but difficult to recall when the query direction departs from how the fact was encountered during training." Recall, they conclude, "is tightly coupled to the conditions under which a fact was learned. When the query diverges from the training-time context, phrasing, or ordering in which the fact was encountered, recall becomes harder."
Our data has this exact shape
This is a parametric-memory paper — no web search, closed book — so it does not describe the retrieval path that fetches a snippet of your page at answer time. But the recall-vs-recognition split is the same split we published on real businesses in July: across 168 local businesses in 41 US metros, the engines recognised the business in 98.8% of by-name answers ("is [business] a good choice?") and named it in 19.8% of buyer-question answers ("who do you recommend for [category] in [city]?"), with 58% of businesses appearing in none of their four buyer answers. The by-name question is the direct, recognition-shaped query — the entity is in the prompt. The buyer's question is the reverse, generation-shaped one: the model must produce the entity from its attributes. Google's paper gives that gap a mechanism: the business is encoded (the warm paragraph proves it), the buyer's question runs against the direction the fact was learned in, and small businesses are the rare tail where recall drops most. It also predicts what our own weekly self-audit keeps finding — that the questions we do get named on are the ones whose wording matches a page of ours almost verbatim: the query converging on "the training-time context, phrasing, or ordering" of the fact.
Why this matters if you care about AI visibility
- "AI knows about us" is the wrong test. A by-name check measures encoding and recognition, which frontier models pass at 95–98%. The number that decides whether a customer hears your name is recall under the customer's phrasing — and that is where the same models drop 26–34% on facts far better documented than any local business. Check the buyer's question, not your own name.
- Write the fact in the direction the question comes from. Recall degrades when the query's ordering departs from how the fact was learned. A page that says "[Business] is a family dentist in Brookline" teaches A→B; the customer asks B→A. Text that also states the reverse — "looking for a family dentist in Brookline? [Business] …" — puts the fact on the web in the order it will be queried. Search Engine Journal's read of the paper reaches the same practical conclusion: order entities the way queries order them.
- Thinking modes recover recall — which is one more reason mode matters. Thinking recovers 40–65% of encoded-but-unrecalled facts. Combined with RESONEO's finding that thinking mode is also where ChatGPT actually opens pages, the same buyer question can produce a different shortlist depending on which button the user pressed — the instability we measure every week is partly the model, partly the mode. Measure across repeats and engines, then ask the engines your buyer's question and see who they name.
Why does ChatGPT know my business but not recommend it?
Because knowing and recalling are different operations. Google Research (blog 2026-08-12, arXiv 2602.14080) finds frontier models encode 95–98% of tested facts but fail to recall 26–34% of them unprompted, with rare facts and reverse-direction questions the hardest to recall. A by-name question puts your business in the prompt (recognition); a buyer's question asks the model to produce it from its attributes (recall). In our July data on 168 local businesses, engines named the business in 98.8% of by-name answers and 19.8% of buyer-question answers.
What is the reversal curse and does it apply to businesses?
The reversal curse is when a model that learned 'A is B' cannot answer 'What is B?'. Google Research's 2,150-fact study finds models recognise reverse facts as easily as direct ones in multiple choice but generate them far less often — 'the reversal curse is a recall problem'. A business's web presence is mostly written A→B ('[Business] is a plumber in Austin'); the customer asks B→A ('who is a good plumber in Austin?').
See your number
See which businesses AI names when your client's buyers ask.
Running this for clients? The $249 agency 5-pack audits five businesses, white-labeled.
Who runs this
- Built and operated by Sensara LLC, Atlanta, Georgia — about us and how the audit works.
- See what the report looks like before you run anything — score per engine, the competitors AI names instead of you, and a fix plan.
- We run the same audit on ourselves every week and publish the result: in the latest run AI named AskedAbout in 1 of 144 answers. We report our own numbers the way we report yours.