The four buyer intents in AI prompts

Keyword research organized the search era because keywords were the unit engines matched. Chat prompts refuse that organization: they are long, conversational, context-laden and nearly infinite in phrasing. What survives across the variety is intent. "Best CRM for a small agency", "which CRM should a five-person shop use" and "recommend a CRM that will not cost enterprise money for my tiny team" are three phrasings of one buying moment, and the engine answers all three the same way: with a shortlist.

The evidence says these buying moments now happen inside assistants at scale. G2 found 51 percent of B2B software buyers start research with an AI chatbot more often than Google, with comparing vendor strengths and weaknesses the top use case at 41 percent. Forrester's survey of 18,000 buyers found 55 percent compared vendors inside AI tools during their most recent purchase. Semrush's 2026 AI Visibility Index, built on 126 million U.S. prompts, exists because this prompt volume became worth indexing.

Sorting those millions of phrasings into intents does for GEO what keyword categories did for SEO: it makes the problem finite. Four intents cover the commercially decisive prompts, each with its own answer format, its own winning condition and its own content fix. Reachroller structures every tracked question set around them, which is why a Reachroller report reads as a strategy document rather than a prompt list. The measurement side of that design is covered in what an AI visibility score actually measures.

The best-X prompt is the open casting call: "best email marketing tool for a startup", "top project management apps for remote teams". The buyer names no brands, so the engine drafts the market from scratch, typically three to six names with a line of reasoning each, and increasingly a steer toward one pick for the asker's situation. This is the intent where the invisible majority is most invisible, and where category defaults harvest most of their advantage, since the engine reaches for the brands its most-cited sources repeat.

Winning best-X is a citations game played on pages you mostly do not own. Shortlist answers lean on review platforms, comparison listicles and community threads, which is consistent with McKinsey's finding that a brand's own site supplies only 5 to 10 percent of the sources AI platforms reference. The work is earning honest presence in the list-shaped sources engines already trust, plus publishing your own answer-shaped category page so there is something first-party to retrieve. How engines pick the steer itself is unpacked in how ChatGPT recommends brands.

Best-X also branches into personas, and the branches are more winnable than the trunk. "Best CRM" belongs to giants; "best CRM for a nonprofit running on grants" belongs to whoever answered that situation credibly somewhere an engine retrieves. A challenger tracking only trunk questions will conclude the intent is closed to them, while the persona branches, which real buyers ask far more often than the generic form, sit uncontested.

The payoff justifies the difficulty: best-X is where brands get discovered by buyers who could not have named them. G2 measured 33 percent of B2B software buyers purchasing from a brand they had never heard of before an AI named it. That statistic lives almost entirely in this intent.

"Alternatives to Mailchimp", "something like Notion but faster", "cheaper option than Salesforce for a small team". The alternatives prompt names an incumbent and asks the engine to dethrone it, which makes it the highest-intent prompt a challenger can win: the buyer has budget, a defined use case, and an active reason to leave. The engine responds with a challenger list, usually organized by the dimension of dissatisfaction it infers, price, simplicity, a missing feature.

Notice the asymmetry that makes this intent special for scoring: the prompt is branded, but with someone else's brand. Your appearance in an answer about a competitor's alternatives is earned, so these questions belong in an honest unbranded set. The content fix is equally specific: a truthful alternatives page for each incumbent your buyers flee, stating who should switch and who should stay, plus presence in the third-party alternatives roundups engines retrieve. Engines reward the candor because hedged pages give them nothing to quote; the craft is detailed in comparison pages that win AI answers.

"HubSpot vs Pipedrive for a two-person sales team". The comparison prompt arrives late in the journey: the buyer has a shortlist and asks the engine to referee. Answers follow a stable format, strengths and weaknesses per option, then a verdict conditioned on the asker's situation. G2's data puts this behavior at the center of AI-era buying, with vendor comparison the single most common research use case at 41 percent.

Real conversations blend the intents in sequence, which is worth knowing when you read raw answers. A session that starts with a how-to drifts into best-X when the engine names tools, and a best-X answer invites the follow-up comparison between two of its own suggestions. The engine carries context across those turns, so a brand that entered the conversation at the how-to stage enjoys an incumbency advantage in the comparison three turns later. Tracking still samples the intents as clean single-turn questions, because clean questions are measurable and repeatable, but the strategy should remember that winning the early intents seeds the later ones.

There are two ways to lose a comparison you should win. Absence: the buyer compares two rivals and you are outside the frame entirely, which is really a lost best-X or alternatives question upstream. And misrepresentation: the engine names you but describes stale pricing, a dead limitation or a feature you shipped two years ago, and hands the verdict to the rival on bad data. The second failure is invisible without stored answers, which is why per-question receipts matter and why the correction workflow in how to fix wrong AI answers exists. Your own comparison pages, written with real facts and a clear verdict per persona, become the engine's source material for the matchups you care about.

"How do I automate follow-up emails after a demo?" The buyer asked for a method and the engine answers with one, but methods run on tools, so the answer names them: "in a tool like X, set up a trigger when...". The how-to prompt is commercially underrated because it does not look like shopping. It is the earliest intent in the journey, it fires constantly, and the tool the method assumes collects a quiet endorsement before the buyer knows a category exists.

How-to is also the intent a small brand can win fastest, because engines cite genuinely useful task content wherever it lives, and most incumbents write documentation for existing customers rather than answers for searching strangers. The Princeton GEO study's finding applies directly: content with statistics, quotations and cited sources lifted generative visibility by up to 40 percent, and a concrete tutorial with real numbers is exactly that shape. Write the task page so the method is complete and honest, with your product as the worked example, and the full craft is in how to write content AI engines actually cite.

The intents are stages as much as categories, and reading them in journey order explains where each brand's losses concentrate. How-to fires earliest, often before the buyer knows a product category exists: the problem has a task shape and the assistant supplies the method. Best-X fires when the category crystallizes and the buyer wants a field of options. Alternatives fires when a known option disappoints, either an incumbent they use or a leader they priced. Comparison fires last, when the shortlist is set and the decision needs a referee. A buyer can walk the entire sequence inside one assistant across a few sessions, which is the concrete meaning of Forrester's finding that 55 percent of buyers compared vendors inside AI tools during their latest purchase.

The journey mapping tells you where losses cost most. Losing how-to costs awareness you never see. Losing best-X costs shortlist entry, the expensive one, since it gates everything downstream. Losing alternatives costs the switchers, the highest-intent traffic in the taxonomy. Losing comparison costs deals you were one answer from winning, and it is the only loss that can happen through misdescription rather than absence. Weighting your tracked set toward the stage where your funnel actually leaks turns the taxonomy from a filing system into a diagnosis.

Category type shifts the weights too. Developer tools and product-led SaaS see heavy how-to volume, because their buyers are practitioners solving tasks. Considered B2B purchases skew toward comparison and best-X, where committees need defensible verdicts. Commodity categories live on alternatives, where price dissatisfaction churns constantly. The right starting split is rarely equal quarters; it is a hypothesis about your buyer that the per-intent data then corrects.

The construction process takes an afternoon. Start from five real buyer situations, ideally lifted from sales calls or onboarding surveys, and write each intent's natural question for each situation: what would this person ask an assistant at the task stage, the field stage, the switching stage, the verdict stage? That grid yields twenty candidate questions before padding, phrased the way buyers phrase them, context included, because assistants get full sentences with constraints rather than keyword fragments.

Then prune with three tests. Would a real prospect plausibly type this, or is it marketing language wearing a question? Is it unbranded, with your name absent, though competitor names are fine in alternatives and comparison prompts? And is it decision-relevant, meaning the answer could plausibly add or remove a vendor from consideration? Questions that fail any test dilute the score with conversations that never mattered. Keep phrasing variants of the same intent deliberately, two or three per high-stakes question, because engines sometimes answer variants differently and the spread itself is information.

Freeze the set before measuring, and version every later change. A question set edited casually becomes a cherry-picking instrument, since retiring losing questions inflates the score without any answer changing. The set should evolve when your product or market does, with the change noted in that period's report so the trend line keeps its meaning back to the baseline.

"What is Acme and is it any good?" The branded prompt is real, buyers ask it constantly, and it must never touch a visibility score. The reason is construction: a question containing your name obligates the answer to discuss you, so a branded mention measures the engine's reading comprehension rather than your presence in the market. The Victorious study quantified the gap, with 96 percent of brands described accurately on direct questions while 89 percent never appeared in category research answers. Blend the two and a brand can report 50 percent visibility while being absent from every conversation that could win it a customer.

This is the most common inflation pattern in AI visibility tooling, partly because it is easy to commit innocently: branded prompts are the first thing every founder types, and a question generator seeded with your domain produces them by default. The defense is structural, a question set built from the four intents with branded prompts quarantined into their own accuracy-monitoring track. That track has real value, catching the stale pricing and wrong claims engines repeat about you, but it is a different instrument answering a different question. The full argument, with worked examples, is in branded vs unbranded prompts.

Reachroller enforces the quarantine by design: tracked question sets are built on the four buyer intents, branded questions are excluded from the headline score, and every mention links to the stored raw answer that produced it. Sampling is repeated because answers move between runs, SparkToro measured under 1 percent consistency across identical ChatGPT prompts, and the per-intent breakdown shows which of the four conversations you are losing. Starter is $29 per month for 25 tracked questions and 400 credits, and when the taxonomy exposes a lost intent, the platform writes the publish-ready fix page for it. The intents are how buyers actually ask; the honest score is how you find out whether the answer is you.