Prompt Is Not the Query: What Actually Happens Before an LLM Cites a Brand

Most GEO conversations start with the same question: what are users asking AI? It sounds like the right place to begin. Identify the prompts, optimize for them, appear in the answers. Clean logic.

But there's an assumption quietly embedded in this approach — that what a user types and what an LLM actually processes are the same thing. They aren't. And understanding that gap changes how you think about brand visibility in AI responses entirely.

What a Prompt Actually Is

When someone opens ChatGPT, Perplexity, or Gemini and types a question, they're not issuing a structured command. They're expressing a need in the most natural way available to them at that moment. The same person searching for the same thing on different days might write it three different ways — sometimes a sentence, sometimes two words, sometimes a full paragraph with context. None of those phrasings is "the query." They're all imprecise surface expressions of an underlying intent.

This matters because LLMs don't use the literal prompt to retrieve information or form a response. They interpret it. Before anything is generated, the model builds an internal representation of what the user is actually asking — stripped of idiosyncratic phrasing, filled with implied context, and expanded to cover the realistic scope of the intent.

Information retrieval researchers have studied this transformation process extensively, though in the context of search systems rather than brand citation directly. The findings are instructive when applied to GEO. Dhole and Agichtein at Emory University's IR lab found that LLM-driven query reformulation — where the model rewrites a user's input before retrieval — outperformed using the raw query by up to 24% on standard benchmarks. The implication for GEO is direct: what the LLM processes internally is meaningfully richer than what the user typed. Optimizing for the typed input is optimizing for the wrong layer.

For brand visibility, this has a direct consequence: if you're trying to appear in AI responses by targeting specific prompt phrasings, you're optimizing for the surface layer of a process that's actually happening one level deeper.

The Intent Cluster Problem

The most useful way to think about this is through the concept of an intent cluster — the full semantic space surrounding a user need. Every distinct need a person might bring to an AI has a cluster: all the different phrasings, vocabularies, levels of specificity, and contextual framings that point to the same underlying intent.

A user asking "best large capacity air fryer" and a user asking "which air fryer models have the biggest cooking capacity in 2025" are expressing the same intent. An LLM recognizes this. It doesn't treat them as separate signals requiring separate content; it treats them as two entry points into the same semantic space. The response — and the brands cited — will be drawn from sources that have established authority across that space, not sources that happened to match one specific phrasing.

This is the fundamental shift GEO requires. The question is not which prompts should we target but which intent clusters are we actually present in, and with what depth of authority. These are different questions that lead to different strategies.

At CiteVista, when we began mapping prompt clusters for brands across different product categories, this distinction became impossible to ignore. We found that many brands had strong content for a handful of specific phrasings but almost no visibility in the broader semantic territory around those phrasings. The LLM would encounter their content for exact-match style queries but couldn't draw on them confidently across the range of ways users actually expressed the same intent. The result: inconsistent visibility that felt random from the outside but was structurally predictable once the intent map existed.

How Persona Shapes the Query

There's a second layer that makes this more complex: the same underlying intent doesn't produce the same internal query for every user.

A first-time buyer exploring a product category phrases questions differently than someone who has bought in that category before and knows what they're evaluating. Someone making a value-conscious decision uses different language and signals different constraints than someone looking for a premium option. A professional with deep domain knowledge asks in a way that implies different background information than a general consumer.

LLMs pick up on these signals — sometimes stated explicitly in the prompt, sometimes inferred from vocabulary, framing, and implied context. The interpreted intent shifts accordingly. And because the intent shifts, the sources consulted and the brands cited can shift too.

This means a brand can be visible to one type of user and completely absent for another, even within the same product category and the same underlying question. It's not a matter of ranking — it's a matter of which version of the intent space your content actually covers.

A 2025 survey on LLMs in information retrieval published in Datenbank-Spektrum describes how LLMs "assist in query understanding by reformulating queries to bridge the gap between user intent and indexed information" — noting that this reformulation is sensitive to contextual signals in ways traditional keyword matching never was. That research was written about search systems, not brand visibility. But the mechanism is the same. For GEO, the practical implication is that persona isn't just an audience consideration for tone and messaging. It's a structural input into what queries get generated internally, and therefore what brands get retrieved.

Why You Can't Know This Without Simulation

Here's the practical problem: you have no direct visibility into how LLMs are reformulating user prompts about your category. There's no Search Console equivalent, no query log you can access, no keyword tool that tells you the internal representations being generated.

What you can do is simulate.

Simulation means generating a realistic set of prompts — across different phrasings, different persona profiles, different funnel stages — running them against the major LLMs, and observing what comes back. Not to find a single winning prompt. To map the territory: where does your brand appear, in what context, alongside which competitors, cited from which sources.

The output of a properly structured simulation tends to be surprising. Intent areas where brands assumed strong visibility often show gaps. Adjacent topics that were never explicitly targeted sometimes show up consistently because of third-party mentions. Certain persona types find the brand reliably while others never encounter it. The distance between assumed visibility and actual visibility is usually significant — and the only way to close it is to first see it clearly.

This is the actual starting point for a GEO content strategy. Not keyword research applied to AI, not guessing which prompts to write about, but an empirical map of where you stand across the intent landscape that matters for your category.

What Changes Once You See the Map

Once you're working from an intent cluster map rather than a list of target prompts, several things shift.

Content planning becomes less about covering specific questions and more about building genuine authority in semantic territory. That means identifying which intent clusters are strategically important, understanding what "authority" looks like in each one (which sources the LLM is already citing, what those sources have in common, what your content does or doesn't share with them), and developing content that earns recognition across the full cluster rather than matching a subset of its phrasings.

The persona dimension also reframes content differentiation. Writing for an expert audience and writing for a novice audience aren't just tone adjustments — they're targeting different query-generation patterns that may require different content architecture altogether.

And source diversity matters more than most GEO frameworks currently acknowledge. Because LLMs are drawing on a broader evidence base than your owned content, appearing only on your own site — even with excellent content — provides weaker citation signals than appearing across multiple sources that the model already treats as authoritative.

A Note on How We Approach This at CiteVista

The questions this piece raises — which intent clusters matter for your brand, how different persona types reach you, what queries LLMs are actually generating from user prompts — are the core of what we've built CiteVista to answer. The platform simulates prompt clusters across persona profiles and funnel stages, maps the queries LLMs generate from those inputs, and tracks where brands appear (and don't) across the resulting output.