What you will learn
People increasingly ask full questions instead of typing short phrases. Prompt research helps you see how a buyer might frame a problem, which sources are surfaced, and where your brand lacks a credible answer. The goal is not to reverse-engineer an assistant. It is to improve the evidence a customer can use.
Why this matters
AI outputs are variable. They can change between sessions, models, plans, locations, and product updates. A one-time screenshot is not a strategy. A consistent observation method lets you learn from variation instead of hiding it behind a made-up score.
Prompt research also reveals a broader problem: customers often need evidence that your site does not publish. A missing comparison, unclear pricing condition, thin case study, or absent third-party proof may matter more than the wording of any one prompt.
The prompt evidence loop
A prompt is a test case derived from a real customer scenario. Record what the system says and cites, then compare that evidence with your owned and third-party sources. The useful output is a concrete evidence improvement, not an unsupported rank.
Core concepts
Prompts should represent decisions
Write prompts around a buyer’s situation, constraints, and desired result. ‘What payroll provider is suitable for a 25-person restaurant with staff in Ontario and Quebec?’ is more useful than asking an assistant to rank your brand.
Use it when: Does this prompt reflect a decision that a real customer might make?
Freeze the test conditions
Record the prompt wording, language, market, device or account state where relevant, date, model or product, and any follow-up questions. Use the same core prompt set for comparison over time.
Use it when: Could someone else repeat this observation with the same conditions?
Map source types, not only brands
Classify cited or referenced material: official documents, editorial reviews, forums, product sites, directories, videos, research, or documentation. Source type often reveals what evidence the customer is missing.
Use it when: What type of evidence appears credible for this question, and do you have it?
Separate absence from negative sentiment
A brand not being mentioned is different from being criticized. Capture wording exactly, distinguish facts from inference, and avoid treating a neutral omission as reputational harm.
Use it when: What was actually said, and what are you inferring from it?
The practical method
- 01
Choose customer scenarios
Use sales and support evidence to identify high-stakes decisions, objections, comparisons, and local or technical constraints.
- 02
Write a balanced prompt set
Include discovery, comparison, objection, implementation, and alternative prompts. Avoid wording that forces your brand into the answer.
- 03
Set observation rules
Fix the schedule, markets, accounts where appropriate, recording method, and reviewer. Document platform changes rather than pretending they do not exist.
- 04
Capture answers and source evidence
Record the response, citations or links, source type, brand mention, factual accuracy, caveats, and customer relevance.
- 05
Identify the evidence gap
Ask whether the gap is owned content, missing product information, weak third-party proof, outdated facts, or a problem with the underlying offer.
- 06
Choose a defensible response
Create or improve useful evidence, correct inaccurate public information, seek genuine expert review, or decide the gap is not worth pursuing.
Guided workshop
Run an AI prompt study that produces useful evidence
This section turns the lesson into a bounded working session. It is designed to leave you with a repeatable observation log that separates prompt conditions, cited sources, accuracy checks, gaps, and a responsible next action.
Practice scenario
Practice scenario: A travel insurer sees competitors mentioned in assistant answers about medical cover for a winter sports trip. Someone asks the team to ‘rank number one in AI.’ The team cannot verify a stable ranking, and the answers change when it changes the destination, traveller age, or follow-up question.
Instead of inventing a score, the team creates a small study. It uses real questions from policyholders, captures the date, market, account state when relevant, answer, cited sources, factual accuracy, and the next question a buyer would still need answered. The legal team reviews wording that could be mistaken for coverage advice.
The observations reveal an owned content gap: the policy explanation does not clearly distinguish adventure sports cover, exclusions, and the evidence a traveller should check before purchase. The first action is to improve that source, not to make a claim about assistant rankings.
Build it step by step
Choose customer scenarios, not brand-forcing prompts
Build a small set from real discovery, comparison, and objection questions. Include relevant constraints such as market, product type, budget, or timing so the study resembles a genuine decision.
Make it tangible: Save a prompt set with scenario labels. It helps the team decide which customer questions are worth observing. Check writing prompts that demand your brand be named before moving forward.
Record the conditions of every observation
Capture date, product, language, market, device or account context when relevant, prompt wording, follow-up sequence, answer, and cited sources. Save the exact output or screenshot where permitted.
Make it tangible: Save a reproducible observation record. It helps the team decide whether two observations are reasonably comparable. Check comparing answers without recording the conditions before moving forward.
Check accuracy before visibility
Ask a subject expert to review material claims, policy boundaries, numbers, and advice. Mark what is accurate, incomplete, misleading, or unverified. Customer safety matters more than a mention.
Make it tangible: Save an accuracy review column. It helps the team decide which source needs a correction or clearer caveat. Check celebrating a citation that contains a harmful error before moving forward.
Classify the evidence gap
Group findings into missing owned facts, unclear explanations, weak third-party evidence, inaccessible pages, stale information, or unanswered customer questions. Use the category to route work to the right owner.
Make it tangible: Save a gap taxonomy and owner list. It helps the team decide what small improvement should happen next. Check treating every missing mention as a content problem before moving forward.
Improve the source, then observe again
Make one evidence-led change to the page, documentation, product data, or third-party source. Wait for an appropriate review period and repeat the same bounded scenarios before drawing conclusions.
Make it tangible: Save a before-and-after change note. It helps the team decide whether the change improved the source or exposed another gap. Check changing prompts, pages, and market conditions at once before moving forward.
Report observations with clear limits
Share patterns, not a fictional universal rank. Explain the sample, conditions, sources, accuracy review, and what the team will do next. Keep the language useful for non-specialists.
Make it tangible: Save a one-page prompt-study summary. It helps the team decide where to invest in source quality or customer clarity. Check claiming that an assistant result guarantees future discovery before moving forward.
Working template
Use these fields in a document, task, or spreadsheet. Keep the evidence close to the decision.
- Customer scenario: Describe the buyer, need, market, constraints, and decision they are trying to make. A customer-facing owner should recognise the scenario as realistic.
- Prompt and follow-up: Record the exact wording and any follow-up questions in order. Another researcher should be able to repeat the same study.
- Conditions: Note date, product, language, market, and account or device context when material. The reviewer should understand what may explain a different answer.
- Sources and accuracy: List cited sources and mark factual statements as accurate, incomplete, misleading, or unverified. A subject expert should review claims that could affect customers.
- Evidence gap: Choose the missing fact, unclear explanation, stale source, access issue, or third-party proof gap. The work owner should know which system they can change.
- Next action: Specify one bounded improvement and the observation you will repeat after it is live. The team should be able to learn without inventing causation.
Quality review before you ship
Use these checks while the evidence, owners, and customer context are still easy to correct.
- Run each prompt as a research probe, not a scorecard. Save the date, wording, locale, account state, and cited sources so another person can understand the observation without treating it as a universal result.
- Follow any named source before changing a page. Ask whether it is accurate, current, relevant to the buyer's question, and available to a normal reader; an answer system's wording is not evidence by itself.
- Compare repeated checks for stable gaps rather than copying one response into a report. A recurring missing fact can justify work; a single odd answer usually deserves a note and a later recheck.
Decision rules for the real world
The answer mentions a competitor
Do: Inspect the cited evidence and unanswered customer question before deciding what to improve.
Avoid: Do not copy the competitor's wording or attempt to force a mention.
The answer is useful but inaccurate
Do: Correct or clarify the most authoritative available source and document the risk.
Avoid: Do not treat visibility as more important than customer safety.
Results vary widely
Do: Narrow the scenario, preserve conditions, and report variation as a finding.
Avoid: Do not average unlike observations into a false score.
No cited source is visible
Do: Focus on the customer task and the quality of your own explanation, then note the observation limit.
Avoid: Do not assume the system used or ignored a particular page.
Coach notes
- A prompt study is a research method, not a leaderboard. Its value comes from repeatable conditions and better sources.
- When a question touches health, finance, legal matters, or safety, add the right expert review before publishing changes.
- Keep the sample small enough that someone can read every answer carefully.
Worked example: a cybersecurity consultancy
The consultancy asks several assistants for the best incident-response firm. Its own brand rarely appears. A rushed response would be to create a page titled ‘Best incident-response firm’ and declare itself the winner.
Prompt research shows that answers repeatedly cite public incident case studies, certifications, response times, sector experience, and independent reports. The consultancy’s site has vague service claims but no disclosed methodology, anonymized outcomes, or guidance for a buyer during the first 24 hours of an incident.
Make it stronger
Sample variation rather than hiding it
Run important prompts several times over a defined period. Summarize consistent patterns and note unstable answers. Do not average incompatible observations into a false precise number.
Audit factual claims
When an answer makes a claim about your product, verify it against primary sources. Correct your own public materials first; then use the platform’s feedback route where one exists and the issue is material.
Protect privacy and terms
Do not submit confidential customer information, sensitive internal data, or personal data to public tools. Follow the product’s usage terms and your organization’s data policies.
Current field note
Study AI answers like changing observations
AI answers vary by product, date, location, account state, and follow-up. A prompt study is useful when it records those conditions and helps the team find an evidence gap—not when it creates a fictional rank.
- Use customer decisions and constraints, not prompts that force your brand into the answer.
- Record the prompt, date, market, product, cited sources, and whether the answer was useful and accurate.
- Compare recurring evidence gaps over time; do not treat one response as a verdict.
Official reference: Google: Optimizing for generative AI features ↗
Lesson artifact
AI prompt observation log
Build a small prompt map for one high-consideration customer decision.
Scenario: Write the customer context, constraints, and consequence of choosing poorly.
Prompt set: Create five prompts: discovery, comparison, objection, implementation, and alternative. Keep them neutral.
Observation sheet: Define the fields you will record: date, product, market, response, sources, mention status, accuracy, and caveats.
Gap decision: Choose one evidence gap that your business can honestly improve, and explain why it would help a customer.
Before you move on
- My prompts describe real customer decisions and constraints.
- I record the conditions of each observation.
- I distinguish a mention, a citation, an omission, and a negative claim.
- I map the type and quality of evidence behind answers.
- My planned response improves customer information even without an AI citation.
Module checkpoint
Turn customer language into a useful site plan
By now, you should have: A source-backed intent log, topic map, and repeatable prompt study.
- Can a real customer question support each proposed page?
- Does each page have one clear job in the journey?
- What evidence is missing before you publish?
Put the lesson into practice.
Create a free Spacebrain account and use the SEO suite with your own data providers.