The conversation is the real AI search query

Search has traditionally been measured as a sequence of discrete queries. Conversational AI breaks that unit of analysis.

A person can begin with a broad question, reveal constraints through follow-ups, compare alternatives introduced by the model, and end with a short instruction such as “which one would you choose?” The final answer may resolve a dozen facts, but neither the opening nor the closing prompt contains the full request.

Our three papers examine a different part of that process. The first looks inside the answer. The second follows the request across turns. The third tests what happens when the model loses that history. Read together, they point to the same conclusion: the useful measurement unit is the conversation.

Three measurements, one story

Layer one: an answer is compressed search work

The first paper introduces answer-reconstruction search density, a way to estimate how much query and source work is compressed into one conversational answer. Across 183 information-seeking conversations containing 1,994 retained answer units, the median answer held 11 retrievable units. A median of three structural facets covered 80% of those units.

Those three facets are a lexical diagnostic, not three observed Google searches. A separate six-task live-web calibration found a median of 1.5 queries and two source pages to reconstruct 80% of an answer. The pilot is a feasibility check, not a population estimate. The important point is structural: one visible interaction can settle several questions and draw on several evidence paths.

That is why keyword volume and raw prompt counts are weak proxies for demand in AI systems. They count the expression, not the information work performed. Our paper-specific analysis of search density covers the full method and its limits.

Layer two: the request develops after the first prompt

The second paper examines 9,212 human-LLM conversations across commercial, values, and controversial domains. In the direct cross-domain multi-turn analysis of 8,133 conversations, the opening turn proved to be an incomplete representation of what users eventually asked.

In the 670-conversation English commercial cohort, the opening prompt contained a median 50.0% of the unique content terms that appeared across the eventual conversation. In 41.0% of those conversations, at least one observable intent dimension appeared only after the opening turn. The public PRISM cohort showed the same general pattern: 36.4% median opening coverage and a 41.7% later-added dimension rate.

A prompt panel built only from polished, standalone questions removes the process through which people clarify what they want. That matters for research, evaluation, and commercial visibility. You can read the companion analysis, The prompt is not the query, for the turn-level distributions.

Layer three: context changes the answer

The first two papers describe information structure. The third tests an outcome. For 180 paired multi-turn conversations, we compared answers produced with the full conversation against answers produced from the final prompt alone. We also tested a compact reconstruction of the preceding context, capped at 160 words.

Full-context and isolated answers differed materially in a weighted 44.7% of cases, with a 95% confidence interval of 33.8% to 56.1%. Full context also improved satisfaction by 0.49 points on a 0-to-4 scale. Adding the compact reconstruction reduced the material-difference rate to 30.8% and nearly closed the satisfaction gap, but did not make the answers equivalent.

The context effect varied by cohort: 68.5% in the commercial cohort and 35.4% in PRISM. That difference is a warning against assuming one universal effect size. It is also evidence that real-world commercial conversations may be particularly sensitive to information accumulated across turns.

What this changes in practice

For AI visibility measurement

Do not treat a list of isolated prompts as a complete market model. Use prompt tracking as a controlled benchmark, then add multi-turn scenarios that preserve evolving constraints and evaluate the resulting answers at the claim and source level. Sampling discipline still matters, which is why we separately publish guidance on how many prompts to track and how often to run them.

For content and brand teams

Map the decision, not only the head question. Publish evidence for the criteria, constraints, comparisons, objections, and follow-up questions that emerge after discovery. A page can matter even when its language never appears in the opening prompt, because it may support a later facet of the answer.

For product and privacy teams

Context has utility, but retaining everything is not the only option. The compact reconstruction in the third study recovered much of the satisfaction benefit while using a bounded summary. That creates a practical design space between context-free answers and unlimited retention, provided teams measure what is lost.

What these studies do not prove

The results do not establish a universal conversion effect, causal persuasion effect, or a single search-query equivalent for every AI answer. The web reconstruction used only six synthetic tasks. The context experiment covered 180 paired conversations and did not test persistent memory across separate sessions. All three measures are policy- and sample-dependent.

What the series does establish is narrower and useful: a prompt is an incomplete observation of conversational demand. Measuring the opening line alone misses later intent. Measuring the final line alone misses prior state. Counting the interaction alone misses the work compressed into the answer.

If AI search is conversational, its measurement has to be conversational too.