How we generate prompts

The quality of AI visibility measurement depends on the quality of the prompts being measured. Cited builds a custom prompt library for every brand it tracks — anchored in real brand context, written the way real customers ask AI assistants, and classified by both intent and topic so you can read your dashboard along either axis.

​

How the pipeline works

When a brand is onboarded — or when its prompt library is refreshed — the pipeline runs in three broad stages. Stage 1 — Brand intelligence. Cited reads the brand’s website and a set of declared and detected competitor sites to understand product positioning, features, pricing, category context, and use cases. The result is a structured brand profile that grounds the downstream prompt synthesis in real context, not generic assumptions. Stage 2 — Prompt synthesis. Cited uses the brand profile to draft candidate prompts the way real customers ask AI assistants — natural, conversational, sometimes messy — rather than the way keyword-stuffed search queries are written. Each candidate is assigned an intent type and a topic at the time of generation. Stage 3 — Quality filtering. Every candidate prompt passes three gates:

1. Consumer authenticity — does this read like something a real customer would actually type into an AI? Candidates that read like marketing copy are dropped.
2. Banned jargon — a curated stop-list rejects marketing-style phrasing (“innovative solution”, “cutting-edge technology”, etc.). This is a hard gate, not a quality nudge.
3. Semantic deduplication — near-duplicates are collapsed so the library doesn’t waste budget asking the same question two slightly different ways.

The pipeline iterates Stages 2 and 3 until your plan-tier prompt target is met.

​

Intent classification

Each generated prompt is tagged with an intent type that captures where the customer is in their decision journey. The dashboard lets you filter and break down every metric by intent — so you can see, for example, whether your brand is strong on problem-led prompts but weak on direct comparisons. These six types are the production tags applied during prompt generation. They map to — but are more granular than — the four-type intent taxonomy used in higher-level reporting.

​

Topic classification

In addition to intent, each prompt is also assigned to a topic within your category — the subject the customer is asking about. Topics are derived from the brand intelligence in Stage 1, so they reflect the real surface area of your category, not a generic dictionary. For a skincare brand the topics might be oily skin, anti-ageing, sensitive skin, and sunscreen; for an HR SaaS brand they might be payroll, onboarding, attendance, and compliance. The dashboard lets you read every metric by topic, so you can see where your brand is winning, where competitors are dominating, and where the category itself is under-represented in AI answers.

​

Why anti-jargon filtering matters

LLMs respond very differently to marketing-style prompts than to consumer-style prompts. “Which CRM is easiest to set up” surfaces a different brand list than “which CRM leverages innovative AI capabilities.” Measuring on jargon-stuffed prompts would tell you what AI says about brochures, not what AI says to customers.

​

Library size by plan tier

Custom prompt libraries are generated per brand using the pipeline above. Plan-tier sizes:

• Starter: 25 prompts
• Pro: 75 prompts
• Scale: 125+ prompts

Custom libraries include a broad mix — branded prompts, comparison prompts, use-case prompts, and category prompts — designed to give you the fullest picture of how AI describes your brand across the surfaces real customers use.

​

Related concepts

• Intent taxonomy — the conceptual framework behind the intent tags
• How we extract mentions — what happens after prompts run
• Non-determinism — why the same prompt produces different results across runs

​

Frequently asked questions