Map the prompts that shape the shortlist
Model how buyers ask AI for advice, comparisons, recommendations, and validation throughout the purchase journey.
Build a controlled set of 30-50 realistic buyer prompts, the questions people actually type into ChatGPT and Perplexity, so you can plan content against them and measure visibility on them, month after month.
You will have a controlled prompt set for content planning and repeatable visibility measurement.
AEO cannot be measured with one vanity prompt. Real buyers vary their role, company, constraint, task, and follow-up. A prompt universe captures those patterns without pretending to know exact prompt volume.
IF YOU SKIP IT / Without a prompt set you are optimizing for prompts nobody types and measuring with screenshots. One lucky answer becomes "AEO is working", until it silently is not.
A pattern is not a diagnosis
Use this sequence throughout the course: observed signal, plausible explanations, verification, confidence, reversible action, retest. Prompt results can narrow an investigation but cannot prove why a system responded as it did.
- Prompt
- The full sentence a person types into an AI assistant, context and constraints included. Longer and more specific than a keyword.
- Archetype
- A repeatable shape of question (recommend, compare, validate...). Coverage across archetypes is what makes a prompt set representative.
- Cohort / tier
- A group of related prompts judged together. You win or lose cohorts, never single prompts.
- Discovery vs validation
- Discovery prompts never name your brand and test whether AI brings you up. Validation prompts name you and test whether AI understands you.
- Frozen set
- A prompt list you stop editing so month-over-month results stay comparable. Edits go in the next version, not the current one.
Each phase feeds the next. Don't skip ahead, the artifact at the end is only trustworthy if every phase ran.
- 01Seed
Start with demand families, call questions, reviews, RFPs, and customer tickets.
- 02Parameterize
Vary role, company type, size, stack, location, constraint, and urgency.
- 03Chain
Add the likely follow-up after each first answer.
- 04Classify
Tag stage, archetype, importance, expected brands, and evidence requirement.
- 05De-duplicate
Keep prompts that test meaningfully different answer conditions.
- 06Map
Assign an owned asset and a trusted-source opportunity to every Tier 1 family.
- 07Freeze baseline
Record exact prompt, engine, model/surface, location, date, and account state for repeatability.
- ✓You now have
Prompt universe
Build prompt archetypes
Start with repeatable question shapes.
Nobody types 'b2b reporting software' into ChatGPT. They type 'recommend a reporting platform for a 15-person agency that needs white-label dashboards and pulls from HubSpot'. Full sentences, with a role, a constraint, a stack. If your prompt set does not look like that, you are measuring a world that does not exist.
The eight archetypes give you coverage without guessing: explain, solve, recommend, compare, alternatives, validate, implement, troubleshoot. Build them from the attribute list you wrote in Module 01, and keep your brand out of every discovery prompt. You are testing whether engines bring you up on their own. Naming yourself in the prompt is answering your own exam.
- 01
Create archetypes for explain, solve, recommend, compare, alternative, validate, implement, and troubleshoot.
- 02
Add the buyer role, company context, constraint, and desired outcome to each archetype.
- 03
Keep prompts natural; avoid inserting your brand into discovery prompts.
- Open the attribute list from Module 01.6 and write one prompt per archetype per priority attribute, in a sheet with columns: prompt, archetype, attribute, source.
- Write them the way people talk: role, company size, constraint, stack ('recommend a reporting platform for a 15-person agency on HubSpot that needs white-label').
- Sanity-test five in a ChatGPT temporary chat. If the answer format surprises you (an essay where you expected a shortlist), rewrite the prompt, not your expectation.
All eight archetypes are represented, every prompt reads like a human typed it, and discovery prompts never contain your brand name.
- PRO TIP
The best prompts are questions prospects asked on sales calls, verbatim. You collected them in Module 01; spend them here.
- WATCH OUT
One brand name in a discovery prompt invalidates it. The test is whether AI brings you up unprompted, and you just prompted it.
OPTIONAL / AI CO-PILOTRun this lesson with AI, the prompt, sources and expected output+
Choose the business priority, interpret exceptions, protect confidential data, challenge weak evidence, and approve the final decision.
Clean, classify, compare, calculate, and draft rows from supplied evidence. AI may surface patterns; it does not own strategy.
Gong / CRM / support, plus captured answer set from ChatGPT / Gemini / Perplexity.
| Seed question | Source | Buyer role | Stage | Observed frequency | Priority |
|---|---|---|---|---|---|
| Best reporting software for agencies | 6 sales calls | Founder | Compare | High | P1 |
CLAUDE / CHATGPT PROMPTAnalyze the evidence without outsourcing the decision+
Remove personal or confidential data. Attach the named exports, explain every column, and tell the model when the dataset was collected.
You are assisting a human B2B SaaS SEO and AEO operator with Module 05: Map the prompts that shape the shortlist. Current lesson: Build prompt archetypes Objective: Start with repeatable question shapes. Required artifact: Prompt universe BUSINESS CONTEXT I WILL PROVIDE - Product, category, target market, pricing model, sales motion, and geography - The priority customer segment and the commercial outcome for this 90-day cycle - Seed prompt library exported from Gong / CRM / support - Captured answer set exported from ChatGPT / Gemini / Perplexity - Definitions for any internal fields, stages, scores, and abbreviations TASK 1. Create archetypes for explain, solve, recommend, compare, alternative, validate, implement, and troubleshoot. 2. Add the buyer role, company context, constraint, and desired outcome to each archetype. 3. Keep prompts natural; avoid inserting your brand into discovery prompts. REQUIRED OUTPUT Return a table using these exact columns: Seed question | Source | Buyer role | Stage | Observed frequency | Priority. For every recommendation, cite the source row, URL, call note, or data point that supports it. Add a confidence column in your analysis: High, Medium, or Low. List missing evidence separately instead of guessing. Finish with a section named HUMAN DECISIONS REQUIRED. RULES - Do not invent search volume, revenue, customer statements, product capabilities, or competitor facts. - Do not treat correlation as causation. - Preserve contradictory evidence and explain why it conflicts. - Do not make the final priority or publishing decision. Prepare the evidence for a human owner. - Use this example only as a format reference, not as evidence: “Recommend reporting platforms for a 15-person agency that needs white-label dashboards and HubSpot data.”
At least eight archetypes are represented. A named human owner must verify this before the lesson is complete.
“Recommend reporting platforms for a 15-person agency that needs white-label dashboards and HubSpot data.”
Add meaningful variations
Small wording changes can produce different shortlists and sources.
Change one word in a prompt and the shortlist changes. Add 'UK', swap 'agency' for 'in-house team', mention a different CRM, and half the recommendations rotate out. That volatility looks like noise. It is actually a map of which contexts you win and which you lose.
Reading the map requires lab discipline: vary one dimension at a time. Change the company size and keep everything else. Then the geography. Then the stack. Chemists do not change two variables per experiment, and starting now, neither do you. Add follow-up prompts too, because real buyers push back with 'which is easiest for non-technical people', and engines answer follow-ups differently than openers.
Recommend reporting platforms for a 15-person agency that needs white-label dashboards and HubSpot data.
“An in-house team of 3, in the UK, on Salesforce” will also change the shortlist, and you will not be able to say which dimension moved it. The row becomes an anecdote.
- 01
Vary one dimension at a time: company size, industry, role, budget, integration, geography, or risk.
- 02
Create follow-ups that ask why, what evidence supports the choice, and which option does not fit.
- 03
Separate branded validation prompts from unbranded discovery prompts.
- Build the parameter matrix in the sheet: base prompt in column A, then one column per dimension (size, industry, geography, stack, budget), changing exactly one per variant row.
- Script the follow-ups as their own rows chained to the opener ('which of those is easiest for non-technical account managers, and why?').
- Label every variant with its base prompt ID so results can be grouped later.
Each variant differs from its base by exactly one dimension, and every Tier 1 prompt has at least one follow-up chained to it.
- PRO TIP
Three variants per base prompt is plenty to start. Depth of repetition beats breadth of variation in the first quarter.
- WATCH OUT
Change two dimensions at once and you can no longer say which one moved the answer. One knob per variant, always.
OPTIONAL / AI CO-PILOTRun this lesson with AI, the prompt, sources and expected output+
Choose the business priority, interpret exceptions, protect confidential data, challenge weak evidence, and approve the final decision.
Clean, classify, compare, calculate, and draft rows from supplied evidence. AI may surface patterns; it does not own strategy.
ChatGPT / Gemini / Perplexity, plus prompt expansion from People Also Ask / forums.
| Base prompt | Role | Company type | Constraint | Stack | Geography | Variant |
|---|---|---|---|---|---|---|
| Best reporting software | Ops lead | Agency | 20 clients | HubSpot | UK | Best HubSpot reporting tool for UK agencies |
CLAUDE / CHATGPT PROMPTAnalyze the evidence without outsourcing the decision+
Remove personal or confidential data. Attach the named exports, explain every column, and tell the model when the dataset was collected.
You are assisting a human B2B SaaS SEO and AEO operator with Module 05: Map the prompts that shape the shortlist. Current lesson: Add meaningful variations Objective: Small wording changes can produce different shortlists and sources. Required artifact: Prompt universe BUSINESS CONTEXT I WILL PROVIDE - Product, category, target market, pricing model, sales motion, and geography - The priority customer segment and the commercial outcome for this 90-day cycle - Captured answer set exported from ChatGPT / Gemini / Perplexity - Prompt expansion exported from People Also Ask / forums - Definitions for any internal fields, stages, scores, and abbreviations TASK 1. Vary one dimension at a time: company size, industry, role, budget, integration, geography, or risk. 2. Create follow-ups that ask why, what evidence supports the choice, and which option does not fit. 3. Separate branded validation prompts from unbranded discovery prompts. REQUIRED OUTPUT Return a table using these exact columns: Base prompt | Role | Company type | Constraint | Stack | Geography | Variant. For every recommendation, cite the source row, URL, call note, or data point that supports it. Add a confidence column in your analysis: High, Medium, or Low. List missing evidence separately instead of guessing. Finish with a section named HUMAN DECISIONS REQUIRED. RULES - Do not invent search volume, revenue, customer statements, product capabilities, or competitor facts. - Do not treat correlation as causation. - Preserve contradictory evidence and explain why it conflicts. - Do not make the final priority or publishing decision. Prepare the evidence for a human owner. - Use this example only as a format reference, not as evidence: Follow-up: “Which of those tools is easiest for non-technical account managers, and what evidence supports that?”
Context variations change one dimension at a time. A named human owner must verify this before the lesson is complete.
Follow-up: “Which of those tools is easiest for non-technical account managers, and what evidence supports that?”
Tag commercial importance
Not every prompt deserves equal attention.
A bathroom scale only tells you something if you weigh in under the same conditions each time. Same scale, same morning, same lack of shoes. Your prompt set is that scale, and this lesson is where you calibrate and then stop touching it.
Score each prompt for fit, revenue proximity, volatility, and current visibility, then freeze a tracking set of 30 to 50 with a version number and a date. Thirty prompts run every month beat a hundred run once. And when this month's answer annoys you, resist editing the prompt. Edits go in v2 next quarter; v1 is busy building your trendline.
- 01
Tag stage: explore, shortlist, validate, or implement.
- 02
Score product fit, revenue proximity, answer volatility, and current visibility from 1-5.
- 03
Choose a stable tracking set of 30-50 prompts and a larger research set for discovery.
- Add the four scoring columns from the steps and compute a tier per prompt. Aim for roughly 10 Tier 1, 10 Tier 2, and 10 research prompts to start.
- Copy the chosen set to a tab named 'Tracking set v1' with today's date, and treat that tab as read-only.
- Put a quarterly reminder in the calendar to review and version the set (v2, v3) rather than editing v1.
A stable tracking set of 30-50 prompts is written down and versioned, and every prompt carries a tier you could justify.
- PRO TIP
Thirty prompts run monthly beats one hundred run once. Consistency is the entire value of the set.
- WATCH OUT
Editing a frozen prompt because this month's answer annoyed you destroys the trendline you spent months building.
OPTIONAL / AI CO-PILOTRun this lesson with AI, the prompt, sources and expected output+
Choose the business priority, interpret exceptions, protect confidential data, challenge weak evidence, and approve the final decision.
Clean, classify, compare, calculate, and draft rows from supplied evidence. AI may surface patterns; it does not own strategy.
People Also Ask / forums, plus prompt baseline from Citation tracker.
| Opening prompt | Answer need | Likely follow-up | Decision stage | Expected evidence | Target asset |
|---|---|---|---|---|---|
| How to automate reports? | Workflow | What breaks? | Validate | QA process | /reporting-qa/ |
CLAUDE / CHATGPT PROMPTAnalyze the evidence without outsourcing the decision+
Remove personal or confidential data. Attach the named exports, explain every column, and tell the model when the dataset was collected.
You are assisting a human B2B SaaS SEO and AEO operator with Module 05: Map the prompts that shape the shortlist. Current lesson: Tag commercial importance Objective: Not every prompt deserves equal attention. Required artifact: Prompt universe BUSINESS CONTEXT I WILL PROVIDE - Product, category, target market, pricing model, sales motion, and geography - The priority customer segment and the commercial outcome for this 90-day cycle - Prompt expansion exported from People Also Ask / forums - Prompt baseline exported from Citation tracker - Definitions for any internal fields, stages, scores, and abbreviations TASK 1. Tag stage: explore, shortlist, validate, or implement. 2. Score product fit, revenue proximity, answer volatility, and current visibility from 1-5. 3. Choose a stable tracking set of 30-50 prompts and a larger research set for discovery. REQUIRED OUTPUT Return a table using these exact columns: Opening prompt | Answer need | Likely follow-up | Decision stage | Expected evidence | Target asset. For every recommendation, cite the source row, URL, call note, or data point that supports it. Add a confidence column in your analysis: High, Medium, or Low. List missing evidence separately instead of guessing. Finish with a section named HUMAN DECISIONS REQUIRED. RULES - Do not invent search volume, revenue, customer statements, product capabilities, or competitor facts. - Do not treat correlation as causation. - Preserve contradictory evidence and explain why it conflicts. - Do not make the final priority or publishing decision. Prepare the evidence for a human owner. - Use this example only as a format reference, not as evidence: A shortlist prompt for your core segment is Tier 1; a broad definition prompt is usually Tier 3.
30-50 prompts form a stable tracking set. A named human owner must verify this before the lesson is complete.
A shortlist prompt for your core segment is Tier 1; a broad definition prompt is usually Tier 3.
Map prompts to assets
Every priority prompt needs an answer plan.
A prompt universe that ends as a spreadsheet was a research project. The version that earns its keep reads like a work order: every priority prompt names the page that should answer it and the third-party evidence an engine would want before trusting you.
A comparison prompt might need your comparison page plus a G2 presence plus two credible reviews. Write both halves down, then route the gaps: missing pages to Module 03, missing sources to Module 08, each with an owner and a date in the same row. When this sheet is done, three other modules already have their to-do lists.
- 01
Name the page that should answer the prompt and the third-party evidence an engine may need.
- 02
Identify missing comparisons, definitions, proof, and source coverage.
- 03
Send on-site gaps to Modules 03/07 and off-site gaps to Module 08.
- Add columns to every Tier 1 prompt: target asset, relevant third-party evidence, observed gap, plausible explanations, verification, and confidence.
- Where the target asset does not exist, file it into the Module 03 page map. Where the evidence is missing, file it into the Module 08 source plan.
- Give each routed gap an owner and a date in the same row. The prompt universe should read like a work order, not a report.
Every Tier 1 prompt names a target asset, relevant third-party evidence, hypotheses, and a verification plan; each approved action has an owner.
- PRO TIP
Map prompt families, not individual prompts. Ten phrasings of 'best agency reporting tool' share one target asset and one evidence set.
OPTIONAL / AI CO-PILOTRun this lesson with AI, the prompt, sources and expected output+
Choose the business priority, interpret exceptions, protect confidential data, challenge weak evidence, and approve the final decision.
Clean, classify, compare, calculate, and draft rows from supplied evidence. AI may surface patterns; it does not own strategy.
Citation tracker, plus seed prompt library from Gong / CRM / support.
| Prompt family | Commercial value | Role coverage | Current visibility | Asset match | Tier | Owner |
|---|---|---|---|---|---|---|
| Agency reporting tools | High | 3 / 4 | Low | Category page | 1 | AEO lead |
CLAUDE / CHATGPT PROMPTAnalyze the evidence without outsourcing the decision+
Remove personal or confidential data. Attach the named exports, explain every column, and tell the model when the dataset was collected.
You are assisting a human B2B SaaS SEO and AEO operator with Module 05: Map the prompts that shape the shortlist. Current lesson: Map prompts to assets Objective: Every priority prompt needs an answer plan. Required artifact: Prompt universe BUSINESS CONTEXT I WILL PROVIDE - Product, category, target market, pricing model, sales motion, and geography - The priority customer segment and the commercial outcome for this 90-day cycle - Prompt baseline exported from Citation tracker - Seed prompt library exported from Gong / CRM / support - Definitions for any internal fields, stages, scores, and abbreviations TASK 1. Name the page that should answer the prompt and the third-party evidence an engine may need. 2. Identify missing comparisons, definitions, proof, and source coverage. 3. Send on-site gaps to Modules 03/07 and off-site gaps to Module 08. REQUIRED OUTPUT Return a table using these exact columns: Prompt family | Commercial value | Role coverage | Current visibility | Asset match | Tier | Owner. For every recommendation, cite the source row, URL, call note, or data point that supports it. Add a confidence column in your analysis: High, Medium, or Low. List missing evidence separately instead of guessing. Finish with a section named HUMAN DECISIONS REQUIRED. RULES - Do not invent search volume, revenue, customer statements, product capabilities, or competitor facts. - Do not treat correlation as causation. - Preserve contradictory evidence and explain why it conflicts. - Do not make the final priority or publishing decision. Prepare the evidence for a human owner. - Use this example only as a format reference, not as evidence: Prompt -> comparison page, G2 category presence, two customer reviews, and a clear fit table.
Every Tier 1 prompt maps to an on-site and source action. A named human owner must verify this before the lesson is complete.
Prompt -> comparison page, G2 category presence, two customer reviews, and a clear fit table.
Add diagnostic prompts and test explanations
Discovery, validation, accuracy, sentiment, and perception prompts reveal patterns. They do not identify a root cause without verification.
Two patients walk in with the same cough. One has allergies, one has pneumonia. Same symptom, opposite treatments, and guessing wrong wastes months. Being missing from AI answers is that cough: either the model does not know what you do, or it knows and does not trust you enough to say so.
The cross-read is the diagnostic. Ask a discovery prompt with no brand in it, then a validation prompt that names you and asks whether you do the thing. Named in both: healthy. Validation only: a trust problem, cured off-page with third-party proof. Neither: a comprehension problem, and no amount of listicle outreach helps until your own pages say plainly what you are.
While you are in there, add sentiment and market-perception prompts. The second kind tells you whether the criteria AI hands undecided buyers are criteria you can win on, which is worth knowing before you spend a quarter competing on the wrong ones.
Verify accuracy, citations, product fit, and repeatability before acting.
Check factual accuracy, prompt interpretation, sources, and run variance.
Possible explanations include source coverage, fit, wording, and sampling variance.
Check crawl/index state, page clarity, product fit, sources, and prompt design.
- 01
Add three prompt types to every Tier 1 attribute: accuracy ("What does X cost and integrate with?"), sentiment ("What do people think of X?"), and market perception ("What should I look for in a [category] tool?").
- 02
Run validation prompts ("Does X offer [capability]?") beside discovery prompts for the same attribute, then record the observed pattern: named in both, validation only, neither, or discovery only.
- 03
For each pattern, list plausible explanations, inspect crawl/index status, owned pages, cited sources, product fit, prompt interpretation, and run variance; assign confidence and route only the smallest testable action.
- For each Tier 1 attribute, run its discovery prompt and its validation prompt 3-5 times each, in fresh temporary chats, and tally: named or absent, confirmed or denied.
- Place each attribute in the grid as an observed pattern only: present in both, validation only, neither, or discovery only. Do not name a root cause yet.
- For each pattern, inspect crawl/index state, owned-page clarity, cited sources, product fit, prompt interpretation, and repeated-run variance. Record confidence, then route the smallest testable action to the relevant module with a retest date.
Every Tier 1 attribute records the observed signal, at least two plausible explanations, verification evidence, confidence, an owner, and a retest date.
- PRO TIP
Run the diagnostic the same week each month, right after the Module 06 baseline re-run, so the two datasets always line up.
- WATCH OUT
A single run is an anecdote. Answers vary; only the tally across repetitions is a diagnosis.
OPTIONAL / AI CO-PILOTRun this lesson with AI, the prompt, sources and expected output+
Choose the business priority, interpret exceptions, protect confidential data, challenge weak evidence, and approve the final decision.
Clean, classify, compare, calculate, and draft rows from supplied evidence. AI may surface patterns; it does not own strategy.
Gong / CRM / support, plus captured answer set from ChatGPT / Gemini / Perplexity.
| Attribute | Observed signal | Plausible explanations | Verification | Confidence | Testable action | Retest |
|---|---|---|---|---|---|---|
| White-label reporting | Discovery 0/10; validation 9/10 | Source coverage; prompt fit; sampling variance | Inspect citations + repeat controlled run | Medium | Correct two inaccurate profiles | 2026-09-15 |
CLAUDE / CHATGPT PROMPTAnalyze the evidence without outsourcing the decision+
Remove personal or confidential data. Attach the named exports, explain every column, and tell the model when the dataset was collected.
You are assisting a human B2B SaaS SEO and AEO operator with Module 05: Map the prompts that shape the shortlist.
Current lesson: Add diagnostic prompts and test explanations
Objective: Discovery, validation, accuracy, sentiment, and perception prompts reveal patterns. They do not identify a root cause without verification.
Required artifact: Prompt universe
BUSINESS CONTEXT I WILL PROVIDE
- Product, category, target market, pricing model, sales motion, and geography
- The priority customer segment and the commercial outcome for this 90-day cycle
- Seed prompt library exported from Gong / CRM / support
- Captured answer set exported from ChatGPT / Gemini / Perplexity
- Definitions for any internal fields, stages, scores, and abbreviations
TASK
1. Add three prompt types to every Tier 1 attribute: accuracy ("What does X cost and integrate with?"), sentiment ("What do people think of X?"), and market perception ("What should I look for in a [category] tool?").
2. Run validation prompts ("Does X offer [capability]?") beside discovery prompts for the same attribute, then record the observed pattern: named in both, validation only, neither, or discovery only.
3. For each pattern, list plausible explanations, inspect crawl/index status, owned pages, cited sources, product fit, prompt interpretation, and run variance; assign confidence and route only the smallest testable action.
REQUIRED OUTPUT
Return a table using these exact columns: Attribute | Observed signal | Plausible explanations | Verification | Confidence | Testable action | Retest.
For every recommendation, cite the source row, URL, call note, or data point that supports it.
Add a confidence column in your analysis: High, Medium, or Low.
List missing evidence separately instead of guessing.
Finish with a section named HUMAN DECISIONS REQUIRED.
RULES
- Do not invent search volume, revenue, customer statements, product capabilities, or competitor facts.
- Do not treat correlation as causation.
- Preserve contradictory evidence and explain why it conflicts.
- Do not make the final priority or publishing decision. Prepare the evidence for a human owner.
- Use this example only as a format reference, not as evidence: Discovery absence with validation presence may reflect source coverage, prompt interpretation, fit, or sampling variance. Inspect the evidence before choosing page work or third-party work.Every Tier 1 prompt maps to an on-site and source action. A named human owner must verify this before the lesson is complete.
Discovery absence with validation presence may reflect source coverage, prompt interpretation, fit, or sampling variance. Inspect the evidence before choosing page work or third-party work.
The screen below shows one sampled answer where the brand is absent. Save the cited sources as observations, inspect their relevance and accuracy, repeat the run, and compare other plausible explanations before choosing an action.
Recommend reporting platforms for a 15-person marketing agency that needs white-label dashboards and HubSpot data.
For a 15-person agency on HubSpot, three platforms come up most consistently:
- Competitor A, strongest white-label options and agency templates.
- Competitor B, deep HubSpot sync, from $49/mo.
- Competitor C, better for larger teams.
Your product isn’t named at all, and the buyer will never know it existed.
The workspace below shows a built prompt universe: seed questions from real calls, the parameter matrix that turns each seed into realistic variants, and the tier scoring. Note the discipline: variants change one dimension at a time, so when visibility differs between two prompts, you know exactly which word caused it.
Buyer prompt universe
Model realistic prompt chains across roles, stages, constraints, and follow-ups instead of tracking a list of head terms.
Prompt coverage by buying stage
Track the full conversation from problem framing to vendor validation.
Check each item only after the artifact meets the standard. Progress saves on this device; use the academy-home backup to move it elsewhere.