geo-prompt-architecture

# GEO Prompt Architecture Build GEO prompt systems that fit the client’s real business, not generic keyword lists. ## Overview Use this skill to generate and audit AI visibility monitoring prompts for GEO programs. It turns a client brief into a `topic -> prompt` architecture across non-brand discovery, competitor comparison, and brand defense, then helps translate monitoring results into concrete optimization actions. When the work needs structured product inputs or outputs, use the JSON schemas in `schemas/`. When the client model is unclear or highly verticalized, use the examples in `examples/` and the playbooks in `references/`. ## Best For - GEO software teams onboarding new clients - GEO agencies building prompt sets at scale - operators who need better prompt coverage by topic, product line, and funnel stage - teams that want to rebalance prompt libraries away from brand-heavy bias - teams that want monitoring prompts tied to later content and asset optimization ## Start With ```text Use $geo-prompt-architecture to generate GEO monitoring prompts for this client. ``` ```text Use $geo-prompt-architecture to review this prompt set and rebalance brand vs non-brand prompts. ``` ```text Use $geo-prompt-architecture to turn these monitoring results into prompt and content recommendations. ``` ## External Access And Minimum Credentials This skill can work from a pasted brief, screenshots, exports, or a website URL. - no private credentials are required for basic prompt generation or review - live browsing is helpful when the client website, topics, product lines, or competitor overlap must be validated - do not assume access to analytics, Search Console, CRM, AI monitoring dashboards, or private docs unless explicitly provided ## Core Model Always frame GEO prompts as a `topic-first` system: 1. `Topic map` Decide which problem spaces, categories, use cases, trust questions, competitor clusters, channels, and seasonal themes deserve monitoring. 2. `Non-brand discovery` Users do not know the brand yet. These prompts measure whether the brand can enter new answer spaces. 3. `Competitor comparison` Users are comparing brands, alternatives, or solution routes. These prompts measure competitive visibility. 4. `Brand defense` Users already know the brand and are validating fit, quality, pricing, sizing, shipping, returns, or worth. These prompts measure narrative control and decision-stage performance. Topic sources can be: - user-provided priority topics - product lines turned into topic seeds - inferred topics generated from the website, business model, use cases, competitors, channels, and weak AI surfaces Default pack size: - `5` topics - `50` prompts total - `10` prompts per topic Default pack mix: - `30-32` non-brand discovery prompts - `12-15` competitor comparison prompts - `5-8` explicit brand prompts Recommended per-topic starting shape: - `6` non-brand discovery prompts - `3` competitor comparison prompts - `1` brand defense prompt Default target mix: - `60-70%` non-brand discovery - `20-25%` competitor comparison - `10-20%` brand defense Do not let brand prompts dominate unless the user explicitly asks for a brand-defense-only set. ## Workflow ### 1. Reconstruct the client model Before generating prompts, identify: - business model - market and language - target customer - user-provided topics, if any - core product lines, if any - conversion path - key competitors - weak AI surfaces, if provided Useful business-model labels: - SaaS / software - ecommerce / DTC - services / consultancy - marketplace / aggregator - manufacturer / supplier - content / media If inputs are incomplete, infer carefully and label the inference. If the user wants a standard onboarding shape, use [schemas/client-brief.schema.json](schemas/client-brief.schema.json). If the business model is ambiguous, read [references/vertical-templates.md](references/vertical-templates.md) and compare against the sample cases in: - [examples/coofandy-topic-first-output.md](examples/coofandy-topic-first-output.md) - [examples/trip-com-consumer-travel-marketplace.md](examples/trip-com-consumer-travel-marketplace.md) - [examples/movinghead-stage-lighting.md](examples/movinghead-stage-lighting.md) ### 2. Build the topic map Do not jump straight into prompts. First, build a topic map that explains what the monitoring system should cover. Priority order: 1. normalize user-provided topics 2. turn product lines into topic seeds 3. infer missing topics from: - use cases - audience segments - competitor overlap - trust and evaluation questions - channels and marketplaces - seasonality and trend patterns Useful topic types: - product/category - use-case - audience/segment - competitor/alternative - trust/evaluation - channel/marketplace - seasonal/trend Every output should make it clear whether a topic is: - `provided` - `derived-from-product-line` - `inferred` If the system identifies more than 5 valid topics, choose the top 5 by: - business value - monitoring value - GEO leverage - competitor pressure - channel fit ### 3. Map the funnel Prompt outputs should use the marketing-funnel labels your product shows: - `TOFU` - `MOFU` - `BOFU` Use this default mapping from the older buyer-journey model: - `Problem awareness` -> `TOFU` - `Solution education` -> `TOFU` - `Category evaluation` -> `MOFU` - `Brand comparison` -> `MOFU` - `Purchase decision` -> `BOFU` - `Use / implementation / expansion` -> `BOFU` Commercial-intent override: - if a prompt is clearly procurement-led, product-spec specific, supplier/vendor selection oriented, or near-term purchase oriented, prefer `BOFU` even if it would otherwise look like category evaluation or comparison Read [references/prompt-framework.md](references/prompt-framework.md) when you need the full generation framework. ### 4. Generate prompt sets by topic Generate prompts inside each topic. Keep the layers separate: - non-brand discovery prompts - competitor comparison prompts - brand defense prompts If product lines exist, use them as one grouping dimension, but do not treat them as mandatory. Some clients need prompt sets grouped by: - topic - business problem - audience segment - marketplace channel - competitor cluster Prompt rules: - write natural-language user questions, not SEO fragments - prefer prompts that fit AI conversations and recommendation flows - include scenarios, constraints, audiences, budgets, regions, or channels when useful - avoid low-value navigational brand variants - keep explicit brand-name prompts sparse in the default 50-prompt pack ### 5. Add GEO judgment, not just prompts For each prompt, include enough structure to make the set operational. Default fields: - prompt - topic - topic_source - topic_type - layer - funnel stage (`TOFU` / `MOFU` / `BOFU`) - category - product line - target customer - business value - GEO priority - monitoring value - likely answer-entry mode - why it matters If the user wants a compact output, keep the fields but shorten the explanations. If the user wants a product-ready response shape, use: - [schemas/prompt-set-output.schema.json](schemas/prompt-set-output.schema.json) - [schemas/prompt-scorecard.schema.json](schemas/prompt-scorecard.schema.json) ### 6. Audit and rewrite existing prompt sets When reviewing an existing prompt list, do not regenerate everything by default. For each prompt: - keep - optimize - downgrade - delete - replace Common failure modes: - no topic map before prompt generation - too many topics with too few prompts per topic - too many brand prompts - no comparison prompts - no true non-brand discovery prompts - off-funnel or synthetic phrasing - prompts that fit search engines better than AI answers - prompts that mismatch the client’s real product line or market - prompts that cluster around one topic while ignoring the real topic surface ### 7. Reverse-optimize from monitoring results When the user brings AI monitoring results, use them to improve both content and the prompt library. Track at least: - was the brand mentioned? - how was it mentioned? - which brands replaced it? - what source types were cited? - what loss reason best explains the miss? Then propose: - content actions - page / asset actions - evidence / entity actions - prompt-set changes Read [references/reverse-optimization.md](references/reverse-optimization.md) when you need the loss-reason model or the reverse-optimization loop. Read [references/scoring-model.md](references/scoring-model.md) when the user wants prompt-set QA, scorecards, or benchmark-style review. ## Output Patterns Default output order: 1. client model summary 2. topic map 3. prompt strategy by layer 4. prompt set by topic 5. priority prompts 6. optional reverse-optimization actions When auditing, prefer tables like: | Original | Action | Final | Reason | |---|---|---|---| | Prompt A | Keep | Prompt A | Fits the topic, product line, and funnel | | Prompt B | Optimize | Better Prompt B | Original is too generic or too brand-heavy | | Prompt C | Delete | — | Low monitoring value | ## Guardrails - Do not treat prompt generation as generic keyword research. - Do not skip topic generation just because the client did not provide topics. - Do not over-index on brand terms. - Do not collapse every prompt into bottom-funnel buying language. - Do not invent product lines, topics, channels, or competitors without labeling the inference. - Do not assume every prompt should become an article; some should map to category pages, comparison pages, FAQs, reviews, or marketplace listings. - When the user asks for monitoring prompts, bias toward prompts that can reveal visibility movement over time. - Do not apply an ecommerce prompt pattern to a marketplace, SaaS, or industrial manufacturer without checking business-model fit first.