The prompt is not the query
Our second paper on arXiv follows the first. It measures how a user's request actually spreads across the turns of an AI conversation. The short version: the final prompt carries a median of only 36% of the session's user vocabulary, and about half the time a need stated earlier never reaches the last prompt at all. Read the full preprint on arXiv (2607.22392).
Bottom line
Most AI-search tools treat the prompt as a stable query: something you can count, classify, and replay on its own. Inside a real conversation that assumption breaks. Across 670 commercial multi-turn conversations and 7,463 public PRISM conversations, the final prompt held a median of just 35.6% and 36.4% of the session's unique user vocabulary. A stated need appeared in the history but not the final prompt in 50.3% and 44.8% of conversations. Among conversations that carried any needs, the final prompt restated the full set in only about 26%, while 17.9% and 19.3% added a brand-new need at the end. The endpoint is not a summary of the request, it is often another move in it.
Source: Tannenbaum, B. (2026), “The Prompt Is Not the Query”, arXiv:2607.22392 [cs.IR]. Two figures per stat are commercial and PRISM cohorts. Vocabulary shares are read as information availability, not semantic drift (length-matched nulls).
This paper describes how a request is distributed across the turns of a conversation. It does not estimate the causal effect of that history on the model's answers, and it is not a claim about answer quality. It is the measurement case for tracking AI demand at the session, which is how Aiso's AI-search visibility platform works.
Why the prompt stopped being the query
Keyword-era measurement had a convenient property: the query was the whole ask. One search box, one line, one intent to classify. AI-search evaluation inherited that habit and still treats a prompt as a stable unit that can be counted and replayed in isolation. But a conversation makes that unit questionable. Each user turn can add a constraint, revise an assumption, ask for evidence, or point back to an alternative raised earlier. The request is built up over the session, not declared in one line.
This paper replaces the fuzzy idea of latent “intent” with something you can observe and count: conversation-conditioned request state. Transparent rules read each user turn for concrete moves that update the request, and the study measures how that state is spread across the turns. It reuses the frozen rules and the governed cohort of the preceding search-density study, then adds a large public replication, so the pattern can be checked twice.
Two cohorts, one pattern
| Cohort | Conversations | Vocab in final prompt | At most half in final prompt | Need dropped from final prompt |
|---|---|---|---|---|
| Commercial multi-turnEnglish, discovery + replication design | 670 | 35.6% | 68.4% | 50.3% |
| PRISM (public)1,389 participants | 7,463 | 36.4% | 74.3% | 44.8% |
All figures are medians or shares of conversations, reported in the paper for the commercial discovery-replication cohort (670) and the public PRISM cohort (7,463 conversations, 1,389 participants). The two corpora were collected differently and still land in the same place.
The vocabulary result: the last turn is a fragment
The simplest measure is lexical: of all the distinct content words a user contributes across a session, how many are in the final prompt? The answer is about a third. The median final prompt carried 35.6% of the session's user vocabulary in the commercial corpus and 36.4% in PRISM. In roughly seven conversations in ten the final prompt held at most half.
Before reading drift into that, the paper checks the obvious confound. Length- matched null models show that low coverage is largely a consequence of how short the last turn is, not the user switching topics. So the vocabulary numbers are read as information availability, how much of the request is even present in the final turn, rather than semantic drift. The categorical results below are the stronger evidence.
Share of the session's user vocabulary that reaches the final prompt
The needs result: dropped, incomplete, or updated
Vocabulary is a blunt measure. The sharper one is categorical. The paper's transparent rules detect specific request-state dimensions, a constraint, a revised assumption, a request for evidence, a reference to alternatives, and then asks where in the conversation each one shows up. The final prompt turns out to be a poor mirror of the whole request in three distinct ways.
It drops needs. A dimension appeared in the history but not the final prompt in 50.3% of commercial and 44.8% of PRISM conversations. It rarely restates the full request. Among conversations that carried any dimensions, the final prompt reproduced the complete observed set in only 26.1% and 26.2%. And it adds needs. In 17.9% and 19.3% the final prompt introduced a dimension that had not appeared before. An endpoint that adds something new is not a summary and not a reference, it is another state update.
| What the final prompt does | Commercial | PRISM |
|---|---|---|
| Reproduces the full set of stated needs | 26.1% | 26.2% |
| Adds a need not stated earlier | 17.9% | 19.3% |
| Drops a need stated earlier | 50.3% | 44.8% |
Shares of conversations, from the paper. “Drops a need” and “reproduces the full set” are computed over conversations that carry at least one detected request-state dimension.
Aiso analysis: four readings that matter for demand
These are our readings of the paper's own figures, framed for anyone who measures AI-search demand or brand presence. They are directional, and they respect the paper's caution that this is a description of how a request is distributed, not a causal claim about model answers.
| Reading | Figure | Why it matters |
|---|---|---|
| Most of the ask is not in the last line | ~64% | The final prompt carries a median 36% of the session's user vocabulary, so most of what the person actually said lives in earlier turns. |
| Half the time the last prompt drops a need | ~50% | A stated request-state dimension appears in the history but not the final prompt in 50.3% of commercial and 44.8% of PRISM conversations. |
| The endpoint rarely restates the request | ~26% | Only about a quarter of dimension-bearing conversations reproduce the full observed set of needs in the final prompt. |
| The endpoint is often another update | ~18% | In 17.9% of commercial and 19.3% of PRISM conversations the final prompt adds a need never stated before, so it is not a summary or a reference. |
Aiso analysis of figures reported in arXiv:2607.22392. Percentages combine or round the commercial and PRISM cohorts for readability; the exact per-cohort figures are in the tables above.
The endpoint is neither a summary nor merely a reference: it is often another state update. A prompt counted, classified, and replayed on its own is a fragment of the request, not the request.
What it means for measuring AI demand
Do not score AI visibility on the last prompt alone
The final prompt holds a median of about a third of the session's user vocabulary. Replaying head prompts in isolation, the way most keyword and prompt trackers do, ignores most of what the person asked. If you audit AI visibility one short prompt at a time, you are measuring a fragment.
The unit of measurement is the session
Request state accumulates across turns: a constraint here, a revised assumption there, a request for evidence, a reference to an alternative raised earlier. The paper shows that state is spread across the conversation, not concentrated in any single line. Measure the conversation, not one prompt.
A deciding need can be invisible in the final prompt
In roughly half of conversations, a need stated earlier never reaches the last turn, yet it can still shape the answer the model gives. A brand, spec, or constraint can decide the outcome without ever appearing in the prompt you would have replayed.
The endpoint is a state update, not a recap
About one conversation in five ends by adding something new. Treating the final prompt as a summary of the whole request is wrong often enough to matter. The end of the conversation is another move, not a tidy restatement.
The companion paper
This is the second paper in a pair on the same governed cohort. The first, answer-reconstruction search density, looks at the other side of the conversation and asks how much query and source work a single AI answer compresses: a median of 11 retrievable facts across 3 facets. Read together, they say the same thing from both ends. The answer packs far more than a keyword tool would count, and the request is spread far wider than a single prompt.
How it was measured
- Commercial corpus: 670 English commercial multi-turn conversations in a discovery-replication design, the governed cohort carried over from the search-density study.
- Public replication: 7,463 PRISM conversations from 1,389 participants, collected independently, used to check the pattern holds outside the commercial data.
- Request state is measured by transparent rules over the user turns, an observable construct standing in for latent intent, so the detection is reproducible.
- Vocabulary coverage is the share of a session's unique user-side content tokens present in the final prompt. Length-matched null models test whether low coverage is just a function of turn length.
What the numbers do not say
- This is a description of how request state is distributed across turns. It does not estimate the causal effect of conversation history on the model's answers.
- Low final-prompt vocabulary coverage is largely explained by turn length. It should be read as information availability, not as the user drifting off topic.
- Request-state dimensions are detected by transparent rules, not human intent labels. They are an observable proxy for intent, not intent itself.
- The commercial corpus is selected for commercial relevance and is not representative of all AI use. PRISM is public and broad but collected under its own protocol.
Frequently asked questions
What is conversation-conditioned request state?
It is an observable stand-in for user intent. Instead of guessing at a latent goal, the paper reads each user turn for concrete moves that update the request: adding a constraint, revising an assumption, asking for evidence, or referring to an alternative established earlier. Request state is the set of those moves, detected by transparent rules over the user turns, and measured across the whole session rather than inferred from a single prompt.
How much of a conversation is in the final prompt?
Not much. Across 670 commercial multi-turn conversations the final prompt held a median 35.6% of the session's unique user-side content vocabulary, and across 7,463 public PRISM conversations the median was 36.4%. The final prompt contained at most half of that vocabulary in 68.4% and 74.3% of conversations, respectively. Most of what the user said sits in earlier turns.
Why isn't the final prompt a good proxy for what the user wants?
Because it routinely leaves things out and sometimes adds new ones. Transparent rules found at least one request-state dimension in the history but not in the final prompt in 50.3% of commercial and 44.8% of PRISM conversations. Among conversations that carried any dimensions, the final prompt reproduced the full set in only 26.1% and 26.2%. At the same time it introduced a previously unseen dimension in 17.9% and 19.3%. The endpoint is neither a summary of the request nor a mere reference back to it, it is often another update.
Does low final-prompt coverage mean the conversation drifts off topic?
No. The paper builds length-matched null models and shows that low lexical coverage is largely a consequence of how short the final turn is, not the user changing subject. So the vocabulary results are read as information availability, how much of the request is present in the last turn, rather than semantic drift. The categorical dimension results are the stronger evidence that the final prompt is an incomplete view of the request.
What does this mean for measuring AI search demand?
It means the session is the right unit of analysis, not the prompt. AI-search evaluation commonly treats a prompt as a stable query that can be counted, classified, and replayed in isolation. This study shows that assumption breaks down inside real conversations: the request is distributed across turns, the final prompt is partial, and it can even introduce fresh needs at the end. To measure demand and brand presence faithfully, track request state across the whole conversation. The study describes how request state is distributed, it does not estimate the causal effect of history on the model's answers.
Measure the conversation, not the prompt
Aiso tracks the real conversations customers have with AI assistants, the needs that build up across turns, and where your brand shows up in the answer, across ChatGPT, Claude, Gemini, Perplexity, and Copilot. See the whole request, not just the last line.