Buyer-intent prompts vs SEO keywords. The two query distributions barely overlap.

SEO keyword volume measures what people type into Google. AI search measures what buyers ASK an LLM in full sentences. The two query distributions overlap less than most SEOs expect, and the gap is the reason traditional keyword tools mismeasure the AI search opportunity.

Industry-data studies from 30-million-citation panels and large public AEO analyses consistently find that AI search query distribution skews commercial, comparative, and full-sentence. The buyer typing "best landing page analyzer for marketing agencies" into ChatGPT is asking a different question than the buyer typing "landing page analyzer" into Google. Google returns ten ranked links. ChatGPT returns one synthesised answer that names two or three brands. The measurement frame that worked for SEO ranks pages against keyword volume. The measurement frame that works for AEO ranks brands against buyer-intent prompts.

This post is about the structural differences between the two query distributions, why your existing keyword tools are blind to most of the AI search opportunity, and what to measure instead.

A buyer typing into ChatGPT does not type a keyword. They type a sentence with context, qualifiers, and an evaluative frame baked in.

An SEO keyword looks like "landing page analyzer". Three words. No qualifier. The buyer is somewhere in their journey, but the keyword does not say where. Google Search returns ten ranked links and the buyer reconstructs the missing context from the page titles.

A buyer-intent prompt looks like "what's the best landing page analyzer for a marketing agency running CRO audits for SaaS clients with limited budget". Eighteen words. Three qualifiers. The evaluative frame is "best for me, given my situation". The buyer has typed the missing context up front because the LLM can use it, and the buyer knows the LLM can use it. The answer they get back names two or three tools by name and explains the trade-offs in two paragraphs. The buyer picks one and moves on.

The query is no longer a navigational instruction. It is a request for a recommendation, and the user expects the engine to behave like a knowledgeable friend, not a directory. This shift changes what "ranking" means. There is no top of the SERP. There is the small set of brands the LLM names in the answer, and the much larger set it does not.

The keyword databases SEO tools use to estimate monthly search volume are built from Google's autocomplete and clickstream data. They under-count AI search queries by design.

SEMrush, Ahrefs, Moz, and the public keyword research tools all derive their volume estimates from Google. Their corpora skew toward short, navigational, informational queries because that is what Google's surface has historically optimised for. Long-form, full-sentence, evaluatively-framed queries are the long tail and are estimated at near-zero volume even when the underlying AI search behaviour is substantial.

The result: a keyword like "landing page analyzer" shows a healthy monthly volume in the SEO tools, while the buyer-intent prompts that dominate AI search show as zero-volume. The SEO tools cannot see that "what's the best landing page analyzer for a marketing agency" is being asked thousands of times per month across ChatGPT, Perplexity, Claude, and Gemini, because the SEO tools never captured the underlying behaviour. The query exists on a different surface.

Using SEO keyword volume to size the AI search opportunity is like using car-park-occupancy data to size the bike-lane opportunity. The methodology is fine. The category is wrong.

Using SEO keyword volume to size the AI search opportunity is like using car-park-occupancy data to size the bike-lane opportunity. The methodology is fine. The category is wrong.

AI search queries skew commercial. The proportion of buyer-intent queries that include commercial language is substantially higher on LLM search surfaces than on Google.

Industry data from 30-million-citation panel studies (the largest publicly documented AEO analyses at the time of writing) consistently finds that buyer-intent prompts to LLM search products skew toward commercial intent: "best X", "X for Y use case", "X vs Y", "alternative to X", "cheapest X", "X with feature Z", "compare X and Y". The user is past the informational stage. They are evaluating, comparing, or shortlisting.

The explanation is structural. A user who needs an informational answer can get it from the LLM directly without leaving the chat. They do not need to click through to a vendor's site. The user who DOES click through to a vendor's site (or who is influenced by which vendor the LLM names) is, by selection, further down the funnel. The mix of queries that produce vendor-name citations is more commercial than the mix of queries that produce educational paragraphs.

For brands, this is good news. The traffic AI search drives skews lower-funnel by composition. Earlier-stage Exposure-Ninja data found AI referral traffic converts at roughly 14.2% versus 2.8% for organic search. The conversion-rate gap is plausible because the user arrived already mid-evaluation, with the LLM's recommendation as ambient pre-trust.

"X vs Y" and "alternative to X" are not exotic edge cases. They are the structural backbone of buyer-intent prompting.

Comparison queries are easy for an LLM to answer well because the answer format is built into the question. The user has asked for a comparison. The LLM produces a comparison. The named brands in the answer are typically the two or three the model has the most third-party corroboration for, drawn from review sites, comparative listicles, Reddit threads, and earned media coverage.

The keyword equivalents (typing "X vs Y" into Google) exist but underestimate the underlying behaviour. The Google version is short, navigational, and the user usually wants the comparison page itself. The LLM version is long, evaluative, and the user wants the synthesised opinion. The brands that show up in the second case are not always the brands that rank for the first case. Comparative listicles drive a documented 21.9-46% of AI citations across platforms (Goodie 2025); G2, Capterra, and similar review platforms drive substantial single-platform shares depending on the AI surface.

The strategic implication: optimising for "X vs Y" on Google is necessary but not sufficient. Brands also need to exist in third-party comparison content (listicles, review sites, Reddit threads) where the LLM gathers the corroboration that produces the citation. The vs-page on your own site is one input. The brand mention in someone else's comparative article is the higher-leverage input.

The right measurement frame for AI search is cited rate on buyer-intent prompts, segmented by engine. Keyword volume is the wrong unit of analysis.

Cited rate is the proportion of buyer-intent prompts (across the engines you care about) on which your brand is named in the synthesised answer. It is the AI-search analog of organic ranking, except the unit of competition is the brand-name mention rather than the URL ranking position.

Per-engine cited rate matters because the engines disagree with each other meaningfully. ChatGPT Search overlaps with Perplexity on only about 11% of cited domains. Each engine pulls from a different substrate (Bing index, Reddit, Google Search plus structured data, etc.). A brand that is cited well on Perplexity may be invisible on ChatGPT Search and vice versa. Aggregating across engines into a single number throws away the actionable information.

The prompt bank itself is the load-bearing design choice. A keyword-derived prompt bank ("landing page analyzer", "CRO audit tool") under-measures the commercial-intent and comparison segments that dominate real AI search behaviour. A buyer-intent prompt bank designed from buyer language (sales calls, support tickets, Reddit posts, customer interviews) captures the queries the LLM actually receives. The two banks are not interchangeable. Substituting one for the other produces a measurement that looks plausible and is structurally wrong.

AI Radar ships a buyer-intent prompt bank by default, segmented by buyer persona and funnel stage, with per-engine cited-rate tracking.

The default prompt bank is segmented across buyer personas (CRO agencies, marketing agencies, web designers, founders) and funnel stages (problem-aware, solution-aware, vendor-comparing, ready-to-buy). The vendor-comparing and ready-to-buy stages get the largest weight because those are where the AI search opportunity concentrates. Each prompt is a full sentence written in buyer language, not an SEO keyword paraphrase.

Each prompt × engine combination runs weekly on every paid tier. Agency runs it 3 times per check and averages the result into a confidence band (covered in the companion post on why single-shot AI citation tools present noise as signal). The dashboard reports cited rate per engine, top-cited competitor domains per prompt, and full-win vs partial-win attribution (the engine cited your domain, but was it the intended page or a different page on your site).

Customers can also add their own prompts. The most common addition is the brand's most-asked sales-call question, paraphrased into LLM-search form. That is the prompt the customer wants ranked for and that the SEO keyword tools showed as zero volume. Adding it to the prompt bank surfaces, for the first time, whether the brand is being named on it.

For the broader picture (what AI citability is, the 8 signals AI engines weigh, and the 5-move 90-day playbook to lift citations), see the AI Citability Playbook and our explainer on answer engine optimization.

To see the buyer-intent prompt bank in action against a real domain, see WhyIQ AI Radar.