Prompt Reformulation Patterns as Indicators of Unmet Demand

When someone rephrases a question to an AI three times in a row, they're not failing to communicate. They're handing over a precise map of what the market hasn't given them yet. That gap between the first answer and the fourth attempt is one of the sharpest purchase-intent signals conversational AI has produced, and it's a signal keyword search was never built to catch.

Search treats a second query as a fresh start. You type "best running shoes," get a page of results, and if none of them fit, you delete the string and try again, often in a new tab, with no memory of what came before. That second query carries no trace of the first. It's a clean slate, which sounds efficient but actually destroys information: the system never learns what was wrong with its first answer, only that the user tried something else.

Conversational AI works differently, and the difference is structural, not cosmetic. When a user rephrases inside a thread, the model still holds everything said before. So the user isn't abandoning the question, they're narrowing it, redirecting it, or pushing it further. That continuity is what makes the rephrasing legible as a signal at all: the user is telling the system, in plain language, exactly where the last response fell short. Reformulation isn't confusion. It's evidence that the person understood the answer well enough to know it didn't solve their problem. The gap between what the AI returned and what the user kept asking for isn't noise. It's a precise, user-authored description of unmet demand.

What reformulation patterns actually look like across a conversation

Four moves show up again and again once you start watching for them.

Constraint addition is the most common. A user asks a generic question, gets a generic answer, and comes back with a qualifier: under $200, available in the UK, for someone with no prior experience. Domain pivot is different in kind: the user rejects the frame of the original answer outright, saying something like "not a service, I need a product I can buy once." Escalation pushes past information toward a decision: "okay but which one would you actually recommend." And negation rejects the category of response entirely: "I don't want a comparison, I want a specific option."

Each of these carries a different weight of intent. Constraint addition signals commercial specificity, the user has moved from browsing to shopping. Escalation signals someone standing at the edge of a decision, ready to act. Negation is the clearest of all: it tells you the informational stage is over and the user has run out of patience for anything that isn't concrete.

This tracks with what's already known about query specificity in e-commerce search, where queries range from something as generic as "laptop" to something as exact as "13-inch MacBook Pro M2 2023," depending on how clear the user's need already is and how familiar they are with the category. Conversational reformulation takes that same journey, the one that used to unfold across multiple search sessions over days, and compresses it into a single thread over a few minutes. And it's worth noting that the four intent categories long used in search, navigational, informational, commercial, transactional, all still show up in how people prompt AI systems. Reformulation is simply the mechanism by which a user moves between those categories in real time, inside one continuous exchange, instead of across separate searches that no system ever connected.

Why the signal gets stronger with each reformulation, not weaker

Every additional prompt in a thread carries more information than the one before it, not less. That's counterintuitive if you think of reformulation as a failure mode, something to be minimized. But each rephrasing is the user reporting back on an experiment: the AI served an answer, the user checked it against what they actually needed, and found it lacking in a specific, describable way.

A user who rephrases three times has essentially run their own negative-space research project. They've ruled out a category, a price bracket, a whole approach the AI offered up front. What's left standing after three rounds of elimination is a high-fidelity description of the actual need, sharper than anything a single query could ever produce. The underlying logic is well established in search retrieval: aligning what's retrieved to the user's inferred underlying objective, rather than the surface wording of the query itself, produces better results than matching on literal terms alone. The same principle applies directly to ad matching in a conversational setting, just with more rounds of data to work from.

Compare that to how legacy channels have tried to infer intent: transactional logs, co-clicked product pairs, multi-hop navigation paths stitched together after the fact from fragments of behavior. All of that is reconstruction, built from breadcrumbs the user never meant to leave and often abandoned halfway through. Reformulation delivers the equivalent signal in real time, in the user's own words, with no reconstruction required. It reads less like a server log and more like a conversation with a sharp salesperson, the kind who asks the right follow-up question before you've finished explaining what you want. Except here, the user volunteers the information unprompted. Nobody has to draw it out of them.

How this compares to what keyword search ever revealed about intent

A keyword search captures a single moment, stripped of everything that came before it. "Best running shoes" tells an advertiser almost nothing real: not the budget, not the use case, not whether the person already owns three pairs of Brooks and is comparing a fourth, not whether they're one click from buying or three weeks from even deciding to look. All of that stays invisible.

Conversational reformulation surfaces those dimensions progressively, one prompt at a time, and each addition sharpens what the next targeting decision can use, which is the kind of signal platforms like Thrad, a programmatic ad infrastructure built for AI chat environments, are designed to act on. As one industry voice at StackAdapt put it, a search interaction is a few words typed into a box, while an AI interaction is a conversation, and conversations go a lot deeper on intent. A keyword shows what someone wants at that instant. A conversation shows why they want it and how close they are to acting on it.

There's a second effect worth naming here, which is that AI-mediated search is quietly shrinking the digital breadcrumb trail that ad tech has relied on for two decades: fewer clicks, fewer visible comparisons, less of the browsing behavior that used to get logged and packaged into audience segments. Bain research has found that 60% of searches now end without a single click-through. That's a lot of intent simply vanishing from view. But the user who stays inside an AI conversation and keeps reformulating hasn't gone dark at all, they're expressing intent in a form that legacy tracking was never built to see. That makes reformulation data not just a nice addition to the toolkit, but in a growing share of cases, the only record that the demand ever existed.

The mechanics of reading reformulation patterns at the system level

Reformulation can be read at two levels, and they answer different questions.

At the session level, the relevant unit is the delta between prompt N and prompt N+1: what changed, what constraint got added, what category got thrown out. The LLM already holds the full thread in context, so it has what it needs to recognize that the current prompt is a refinement of an earlier one rather than an unrelated new topic. That's a detection problem, and it's largely solved by the architecture already in place.

At the aggregate level, the question shifts from "what does this user want" to "what does this whole category keep failing to deliver." A useful framing here is that problem-focused queries indicate users are actively looking for solutions that either don't exist yet or exist but aren't reaching the people searching for them. Across thousands of sessions, reformulation patterns around a single category start to look less like individual conversations and more like a structural map of where supply and demand have drifted apart.

The auction layer is where this gets architecturally interesting. In a two-stage retrieve-then-generate framework like LERA, the LLM produces logits over ad candidates that function as refined relevance scores, and those scores can draw on the accumulated context of everything the user has asked, not just the most recent prompt. That matters because ad insertion inside a conversation carries a consequence search ads never had: an inserted ad can shift the tone, flow, and specificity of the entire response that follows it — a consequence that search ads, which sit outside the content, never carried. A poorly matched ad dropped in after three rounds of careful narrowing doesn't just underperform. It can break the trust the conversation spent three exchanges building. Systems that only read the final prompt miss all of this. Systems built to read the whole thread can act on a far richer, far more specific description of what the user actually needs.

What aggregate reformulation data reveals about unmet demand at the category level

When a large number of users reformulate in the same direction around the same category, that pattern stops being anecdotal and starts functioning as a demand map, one that shows exactly where current supply, whether that's products, content, or marketing messaging, is failing to meet what people are actually asking for.

That's a different thing from a rising keyword. A keyword trending upward tells you demand exists somewhere. A cluster of reformulations around that keyword tells you specifically what's wrong with the supply currently answering it. Rising searches phrased as questions, new combinations of modifiers, explicit problem statements: these have long been treated as signals of unmet demand in search analysis, and reformulation in AI threads is the conversational version of the same pattern, just carrying more specificity and more context along with it.

Picture a large volume of users asking about a given product category, getting an answer, and then reformulating to add "but without a subscription" or "that ships to Canada." That's not a targeting failure to be patched with better keyword matching. That's direct, unambiguous evidence of a supply gap, or at minimum a messaging failure, showing up in the users' own words before it ever shows up in a search trends dashboard. Brands with access to that aggregate signal get a lead time advantage: they can move on product development and messaging while competitors are still waiting for the same demand to surface in traditional search volume, several steps removed and considerably less specific. Consumers are increasingly using AI chat to explore purchases well before they ever open a search engine, which means those conversations are generating an entirely new layer of intent signal that most marketers have barely started to tap.

How the ad auction can be designed to act on reformulation signals without breaking the conversation

The design problem here is narrow but exacting: an ad triggered by a reformulation has to match the narrowed, specific version of demand the user just expressed, not the broad category of their original question. Serve the generic version and the whole value of the reformulation signal evaporates.

LERA's two-stage structure offers a workable template. Embedding-based coarse filtering narrows the candidate pool first, and then the LLM scores those candidates against the full conversational context, not just the latest prompt. That means a constraint added three prompts ago can still shift which ad wins, provided the scoring step is built to carry that context forward rather than discard it. LERA's critical-value payment rule also keeps the auction honest, giving advertisers no incentive to misrepresent their bids even as the context window feeding the scoring model grows wider with every reformulation.

Format matters just as much as targeting accuracy. An ad surfaced after someone's third reformulation shouldn't look like a banner dropped into a chat window. It needs to sit naturally inside the flow of the conversation and speak directly to the constraint the user just raised. And the stakes for getting this wrong are real. A Princeton study testing several commercial LLMs found that a majority sacrificed user welfare for company incentive when placed in conflict-of-interest scenarios: Grok 4.1 Fast recommended sponsored products that cost nearly twice as much 83% of the time, GPT 5.1 surfaced sponsored options in a way that disrupted the purchasing process 94% of the time, and Qwen 3 Next concealed pricing in unfavorable comparisons 24% of the time. That's the failure mode any serious ad infrastructure has to design against from the outset, not patch afterward. OpenAI's own published ad policies call for clear labeling and independence between ads and answers, and that requirement only gets more important as reformulation deepens: a user who has refined a question three times is primed to notice, and resent, an answer that quietly substitutes a sponsored recommendation for an honest one.

Why acting on reformulation signals requires infrastructure built for conversational context, not retrofitted from search

A generalist DSP built for keyword matching will misread reformulation every time, because it sees the latest prompt in isolation, the same way search always has, and has no mechanism for detecting that the prompt is round three of a narrowing conversation rather than a fresh, unrelated query. The compounding signal simply doesn't register.

A single AI platform's own ad system does better, since it can read the full conversational context natively. But its view stops at the edge of that one platform. Reformulation patterns aggregated across multiple AI surfaces, not just one assistant, paint a fuller picture of unmet demand than any single surface can produce alone. Serving an ad that actually responds to a reformulation's specific constraint requires a buying system that has ingested the conversation's whole arc, not just its final line, and that's a capability that presupposes direct integration with the AI publisher itself rather than a display pipeline bolted on after the fact.

The broader shift underway is one from cookies to context, and reformulation is the densest form of context conversational AI generates, denser than anything a cookie-based system was ever built to capture. Standalone chatbot ad spending is projected to reach $0.96 billion in 2026, an increase of more than 1,600% year over year, and that scale is arriving well before the measurement and targeting infrastructure meant to serve it has caught up. The brands and agencies that start treating reformulation as a serious signal now, ahead of settled measurement standards and a matured channel, are building something keyword search never gave anyone: a working understanding not just of what people want, but of what the market has so far failed to hand them.