Google Research just showed why AI queries are not keywords

A query is a string of words. The need behind it can span several decisions. Google's researchers open with a two-word example: someone searching for camping gear does not want ten slight variations of the same four-person tent. They want a set that works together: a tent, a sleeping bag, a portable stove, a headlamp.

Two words in, several product categories out. That gap is what Retrieve-for-Train sets out to close, and it reaches well beyond camping. The prompts buyers type into ChatGPT, Gemini or Google's AI Mode have the same shape: few words, many implied needs.

What did Google Research publish on 15 September 2026?

Google Research published Retrieve-for-Train (R4T), a framework that trains a compact model to return, in a single pass, a set of complementary retrieval directions for a broad query. The post is signed by Pengcheng Jiang and Judith Yue Li. It summarises an ICML 2026 paper, Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion, first posted to arXiv on 6 March 2026 by eleven authors.

The method runs in three steps. Reinforcement learning first trains a fan-out language model (Gemma3-4B and Qwen3-4B in the experiments, each asked for ten sub-queries per prompt) against a reward that scores the whole set of results instead of each result alone. The trained model is then frozen and generates training pairs offline, with no human labelling. Finally, a diffusion model of 53.9 million parameters learns to go straight from a query to a complete set of targets, without writing out any reasoning tokens.

The reported gains cover quality and speed. R4T beat single-query search, zero-shot expansion and a Best-of-N baseline on both tasks tested. The diffusion retriever ran 12 to 20 times faster than autoregressive approaches. On large batches, autoregressive fan-out climbed toward 50 seconds while R4T stayed between under a second and a few seconds.

What is paraphrastic collapse?

Paraphrastic collapse is what happens when a language model, asked to expand a broad query into several searches, returns near-synonyms of the query instead of complementary directions. Google Research gives the example of “bohemian festival style”. Without careful prompting, a generic model offers “bohemian festival fashion” and “bohemian festival clothes”, where a fashion expert would search for fringe jackets, crochet dresses or suede boots.

The number of searches goes up. The ground they cover does not. Results overlap, the same recommendations come back, and whole parts of the original need stay unexplored. Apparent breadth and real coverage come apart, and that is the entire problem.

Back to camping. “Essential camping equipment”, “best camping gear”, “camping equipment checklist” and “must-have camping gear” are four searches aimed at one spot. None of them goes looking for a stove or a headlamp.

How does Retrieve-for-Train judge a good set of results?

Retrieve-for-Train scores the set as a whole, on three properties that pull against each other: groundedness, diversity and alignment. Groundedness penalises any sub-query that drifts away from the database, so every direction points at something that exists. Diversity is measured with the Vendi Score across the full set, which pushes the model toward real semantic breadth. Alignment ties each sub-query back to the original prompt, so breadth does not turn into drift.

The published ablation shows why all three are needed. Optimised for groundedness alone, the model gamed its reward with degenerate strings that happened to land on database coordinates. With alignment added, it fell back into paraphrasing the user's prompt. Only the diversity term closed both shortcuts.

For anyone working outside Google, that shift is the useful part. Relevance judged one result at a time does not guarantee a good answer. Ten results can each be relevant and still form an incomplete set, if all ten cover the same dimension of the need.

Does Google AI Mode use Retrieve-for-Train?

No public source shows that Google AI Mode uses Retrieve-for-Train. The framework was evaluated on two specialised tasks: text-to-image retrieval over a large dataset of user-curated fashion outfits, and text-to-music retrieval over a proprietary dataset of expert-built playlists. Neither is web search.

What is documented is narrower. Since March 2025, Google has described AI Mode as using a “query fan-out” technique: it runs several related searches across subtopics and data sources at the same time, then assembles the results. The Google Research post links the term to that AI Mode announcement. Neither document says which fan-out method runs in production, how many sub-queries it generates, or which criteria it optimises.

A GEO method that blurs the two loses the ability to separate what it observes from what it assumes. That separation is the difference between a diagnosis and a claim.

Why do many AI visibility panels measure wording instead of the decision?

Because they are built like keyword lists: many phrasings of one selection question, each counted as a separate need. Picture a 30-prompt panel for a fertility clinic in Bangkok that includes “best fertility clinic Bangkok”, “top IVF clinic Bangkok”, “recommended fertility clinic Bangkok”, “best IVF centre Thailand” and “leading IVF clinic Bangkok”. The answers will differ a little. The decision under test is the same every time: which provider to choose.

That is paraphrastic collapse, moved from the retrieval system to the measurement of it. The panel looks broad and covers a single territory. Compare it with a panel built on the distinct needs around the same decision.

A clinic could appear in eight out of ten repetitions of the selection question and still be missing when the patient asks about cost, accreditation or IVF after 40. Another could appear less often on the generic question and hold strong, credible evidence on expertise, safety and comparison. The two visibility profiles differ in kind, and a single mention rate hides the difference.

Paraphrase panels still have a use. To test how sensitive an engine is to wording, near-synonyms are exactly the right instrument. They measure wording stability. They do not measure coverage of a decision, and a report should say which of the two it measures.

What is a query territory?

A query territory is a coherent information need, or decision dimension, that buyers can express through several different questions. A query is the exact string submitted to an AI engine. The territory is the part of the decision that string belongs to. KAIS uses the territory, not the query, as the unit of measurement.

Take the commercial intent “find an AI search visibility agency in Thailand”. The territories around it include discovery (which firms offer GEO and AEO), expertise (who has documented experience measuring AI search), methodology (how mentions, citations and recommendations are measured), evidence (who publishes original research or verifiable results), comparison (how service scopes differ), trust (which independent sources support the claims) and selection (who fits a given organisation's requirements). They are related. They are not synonyms, and a company can be well represented in some while being absent from others.

In practice, a KAIS panel is built in layers: the client's actual activities, the words its customers use, four intents (finding a provider, understanding how something works, comparing options, checking reputation), the people who ask and the events that trigger the question, objections, and each market language as its own axis. A core of 20 to 30 questions, chosen for commercial value, stays frozen and is re-measured at every wave so the deltas mean something. Questions that name the brand never count in the score: a query that names the brand always finds it.

The territory is also the unit of KAIS's exclusivity model: one client per strategic territory, defined by a named list of questions annexed to the contract.

Why does the evidence around a brand matter more than repeating the target phrase?

Because each need behind a question can lead to a different set of documents, publishers and entities, and those documents rarely repeat the words of the original prompt. A clinic can publish excellent pages about its treatments and still be thin in the independent sources that cover accreditation or the experience of international patients. A competitor with a weaker service page can hold stronger evidence in a specialist directory or an independent publication. Same prompt, different information paths.

Take a company that wants to appear for “best AI search agency in Bangkok”. The keyword reflex is to publish more pages on variants of that phrase. An evidence-led approach asks what a system would need to find before it could reasonably put the company in an answer. Accessible content that describes its services accurately. Documented expertise. A method and a scope written down. The same identity, location and capabilities from one source to the next. Independent publications that discuss its work. Something that sets it apart from the alternatives.

Those answers sit on different sites and in different kinds of source. A technical article supports a question about expertise. An original research report supports a question about method. An independent publication corroborates the work, a professional profile anchors identity and scope, and a genuine client review speaks to the service itself. None of them has to contain the phrase “best AI search agency Bangkok” to count. Two engines can even lean on different parts of that evidence for the same brand, as Same brand, different proof shows.

The goal is not to repeat the query across the web. It is to make accurate, useful and credible evidence available for each need that sits behind it.

What should AI visibility measurement track beyond exact prompts?

AI visibility measurement should keep five readings separate instead of folding them into one number too early.

• Exact-query visibility: does the brand appear when a specific prompt is submitted? It remains the base for controlled benchmarks and for comparing one wave with the next.
• Territory coverage: across which distinct needs does the brand appear, and where is relevant evidence missing?
• Cross-territory consistency: does the engine describe the entity the same way when the question moves from discovery to expertise, comparison or trust?
• Evidence diversity: which types of source support the brand across those territories: its own site only, or credible independent publications as well?
• Recommendation persistence: does the brand remain a candidate when the question adds a constraint on budget, location, expertise or regulation?

Two limits apply to all five, and they belong in any report. A list of sub-questions drawn up by an analyst is not proof that an engine ran those sub-queries internally: it models the need, it does not trace the retrieval. And the sources an answer displays are not a complete record of what the engine retrieved. From the outside, retrieval is visible only through what the engine chooses to show.

How can a brand test its query territory coverage?

A brand can test its coverage by taking one commercial question, building two controlled sets of prompts around it, and running both repeatedly on the AI engines its market uses. The first set holds alternative phrasings of the selection question. The second holds the distinct needs behind the same decision. For a company in Thailand, the starting question might be “which AI search agency should we choose?”.

For each answer, record:

1. whether the brand appears;
2. its role in the answer: mentioned, considered, or explicitly recommended;
3. its position, when the answer ranks options;
4. the sources visibly cited or surfaced;
5. whether its attributes are described accurately;
6. which competitors appear;
7. whether it persists from one territory to the next.

Repetition matters. A single pass cannot tell a stable absence from an occasional one, so KAIS labels one pass per question an exploratory reading and runs three passes for a diagnostic. The second set does not reproduce any engine's internal fan-out, and it does not need to. Its job is to show how visibility changes when a buyer explores different sides of the same decision. Thirty near-identical questions tell you about wording stability. Thirty complementary questions tell you about the breadth of your presence. Both are useful. They are not the same measurement.

Where does query fan-out fit in the KAIS AI Visibility Chain?

Query fan-out sits between the user's question and the second stage of the KAIS AI Visibility Chain, Be Retrieved: it decides which parts of the information space get explored before anything can be cited.

The later stages stay separate. A source can be discoverable and never retrieved for a given question, retrieved and not cited, cited without shaping the answer (the gap measured in Cited is not used), and understood correctly without being chosen. The chain also has a documented exception: an engine can name a brand from what its model already learned, with no retrieval at all. Fan-out research says nothing about that path.

This is why adding more variants of a target keyword rarely fixes an absence. The page may be unreachable, as in the firewall case described in Your robots.txt says yes. Your firewall says no. The evidence may not answer the buyer's specific requirement. The engine may describe the company wrongly. Independent corroboration may be missing. Or the brand may simply not be selected in that response. Each gap calls for a different intervention, and the diagnosis has to come before the tactic.

The prompt is short. The decision is not.

Google Research did not announce the end of keywords, and Retrieve-for-Train is not a new ranking factor. Keywords still describe how people phrase demand, and they remain a sound unit for conventional search.

What the work shows is narrower and more useful. When a need is broad, a set of complementary results serves it better than a pile of relevant, redundant ones. Buyers ask AI engines short questions. The information they need to decide runs much longer.

So the question for a brand shifts. It is no longer only “do we appear when someone submits this exact prompt?”. It is also “when the buyer explores the different sides of that question, does the available evidence keep us in the answer?”

Sources

• Google Research — Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train, Pengcheng Jiang and Judith Yue Li, 15 September 2026. The camping gear and bohemian festival examples, paraphrastic collapse, the three-step pipeline, the composite reward (groundedness, diversity, alignment), the 53.9M-parameter diffusion retriever, the 12 to 20 times speedup and the latency of nearly 50 seconds for autoregressive fan-out at scale.
• Jiang, Li, Ryu, Hu, Su, Wan, Hebert, Peng, Han, Kuzmin and Boutilier — Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion, arXiv:2603.06397, submitted 6 March 2026, ICML 2026. Fashion and music benchmarks; query-time fan-out latency reduced by an order of magnitude. The results concern the tested framework and do not establish deployment in any Google product.
• Google — Expanding AI Overviews and introducing AI Mode, Robby Stein, 5 March 2025. AI Mode's query fan-out technique: several related searches run concurrently across subtopics and data sources, then combined.