1. Start with the decisions customers make
The objective is to understand whether an AI assistant connects the right brand, product or service to a real decision. In this sector, start with product relevance, price and inventory accuracy, discoverability and fulfillment trust.
Who is asking?
A shopper comparing products within a budget and delivery constraint.
A realistic situation
A customer needs an affordable practical housewarming gift, wants delivery in four days and prefers easy returns.
Choose a manageable set of buyer problems rather than tracking hundreds of near-identical queries. Include questions for initial education, supplier discovery, evaluation and verification. Separate unbranded recommendations from prompts that already contain your product or company name.
Use prompts to locate content and visibility gaps, not to manufacture flattering answers. If an assistant cannot support its recommendation with clear evidence, record that uncertainty as part of the result.
2. Copy and adapt the core prompt library
This starter library covers the most important research and purchase contexts for retail & e-commerce. Replace terms in square brackets with real locations, products, use cases or current dates. Treat the prompts as examples for your research program, not as observed user-search-volume data.
12 distinct questions · Search and answer engines may respond differently.
| Topic | Natural-language prompt | Use | Copy |
|---|---|---|---|
| Category discovery | Which compact home-office accessories improve a desk setup without taking much space? | Discovery | |
| Budget filtering | What are good gifts for a new apartment under $70? | Discovery | |
| Feature comparison | What matters when comparing cordless vacuums for pet hair on hard floors? | Evaluation | |
| Private label | How should shoppers compare a retailer-owned kitchen appliance with established brands? | Comparison | |
| Availability | How can I check whether [product] is available for pickup near [city]? | Validation | |
| Delivery reliability | What shipping and delivery details should I verify before a last-minute gift purchase? | Education | |
| Returns | Where can I confirm the return window for [retailer] [category]? | Fact check | |
| New releases | Which features have changed in the latest generation of [product type]? | Discovery | |
| Seasonal gifting | What useful gifts work for a remote-working graduate under $100? | Discovery | |
| Promotions | What is a reliable way to compare a sale price with recent everyday prices? | Education | |
| Material claims | How can I verify that [product] uses [material] as described? | Fact check | |
| Buying channels | What are the trade-offs between ordering [category] directly from a brand and from a marketplace? | Evaluation |
3. Make a broad prompt decision-ready
A useful prompt contains enough context to make a recommendation meaningful. Instead of stacking keywords, describe the job to be done, the important constraints and the kind of answer a buyer actually needs.
This is a synthetic example created for monitoring design; it is not evidence that people use precisely this wording.
What to vary, and what to keep stable
Run an initial prompt as written and keep that wording fixed when comparing historical results. Then create separate named variants for audience, region, price range, urgency or other constraints. Document the new variant rather than overwriting the original question.
4. Use different prompts for different questions
Mixing prompt types into a single visibility number can make performance look stronger than it is. Define a role for each question before you run it.
| Prompt role | How to write it | What to measure |
|---|---|---|
| Unbranded discovery | Describe the need without naming your organization. | Spontaneous brand inclusion and recommendation context. |
| Head-to-head comparison | Name options and give a fair, specific buying situation. | Trade-offs, preferred options and competitor framing. |
| Branded validation | Ask whether a named product, provider or publisher meets a requirement. | Completeness and correctness of brand-specific answers. |
| Factual verification | Test a precise, externally verifiable claim. | Whether facts match dated authoritative records. |
Keep the platform, country or city, language and run date alongside each answer. An assistant's output can vary by context and repeated runs; one result should not be treated as an established ranking.
6. Evaluate answers, citations and factual quality
Use a consistent scorecard that distinguishes whether a brand appears from whether it is recommended and whether its claims are trustworthy. Interpret results by topic and platform; avoid implying that a citation is guaranteed merely because a page is crawlable.
| Measure | Working definition | Review note |
|---|---|---|
| Unaided mention rate | Unbranded runs that name your brand ÷ eligible unbranded runs. | Exclude branded validation questions. |
| Recommendation rate | Eligible runs that explicitly suggest your brand ÷ eligible runs. | A passing mention does not count as a recommendation. |
| Owned-site citation rate | Runs with a linked owned-domain source ÷ runs where sources are available. | Preserve exact URLs where shown. |
| Claim accuracy | Verified claims divided by claims selected for independent review. | Report uncertain and unverified claims separately. |
| Competitor presence | Frequency of other brands in comparable prompt groups. | Record use case and context, not just name counts. |
Prices, stock, delivery promises and returns may vary by location and date; verify before presenting them as live facts.
Source of truth
Current product detail pages, catalog feeds, stock status, delivery estimates, return policies and product labeling.
| Claim type | Authoritative source | What to log |
|---|---|---|
| Product details | Current product detail pages | URL or document version + verification date |
| Catalog data | Catalog feeds | URL or document version + verification date |
| Stock | Stock status | URL or document version + verification date |
| Delivery times | Delivery estimates | URL or document version + verification date |
| Returns | Return policies | URL or document version + verification date |
| Labeling | Product labeling | URL or document version + verification date |
Record the source URL or document version and verification date when an answer makes a material factual claim.
7. Turn findings into SEO and GEO improvements
After baseline testing, connect the missing answers to real website improvements. Echorank's technical SEO and GEO lens prioritizes content that is accessible to crawlers, easy for visitors to understand and supported by reliable evidence.
- Structure category and product pages around selection criteria and use cases.
- Publish current shipping, returns, availability and product details as crawlable text.
- Add buying guides that help shoppers understand trade-offs, not only rankings.
- Clearly mark promotion dates and product variants to prevent stale AI summaries.
Prioritize the work
| Priority | Issue | Why it matters |
|---|---|---|
| Fix now | Unsupported high-risk facts in AI answers | High-risk claims need evidence before anyone relies on them. |
| Fix now | Wrong live policy, availability or specification data | Buyers act on live details; stale ones mislead decisions. |
| Fix now | Core information hidden behind inaccessible scripts | Crawlers and assistants cannot use what they cannot read. |
| Build next | Helpful answer-focused pages for missed customer needs | Covers the customer needs that answers currently miss. |
| Build next | Clear comparison tables and product or service details | Gives assistants explicit, checkable facts to compare. |
| Build next | Stronger authority, authorship and evidence signals | Makes it clear who stands behind each claim. |
8. Ask better questions during your visibility review
Use these prompts in a research meeting or with an assistant that has access to your actual collected answers. They are analysis prompts—not evidence that Echorank has automatically performed the analyses.
| Question for the research team | Copy |
|---|---|
| Which competitors appear for high-intent gift and budget prompts? | |
| Does AI cite category guides, product pages or third-party marketplaces? | |
| Which fulfillment facts about the store are inaccurate? | |
| How does recommendation visibility change across budgets and buyer needs? |
When requesting analysis, provide the underlying response exports, prompt definitions, dates and geographic scope. Without those inputs, a model should not invent visibility rates or competitor findings.
9. A practical 30-day rollout
| When | Focus | Deliverable |
|---|---|---|
| Week 1 | Choose priority products, customer decisions and prompt types. Agree on segment tags. | Reviewed prompt inventory and baseline definition. |
| Week 2 | Run comparable queries, capture full responses and URLs, and flag factual claims. | Segmented observation log. |
| Week 3 | Review facts against primary sources and identify page-level information gaps. | Risk-ranked content and technical backlog. |
| Week 4 | Implement initial fixes; rerun stable prompts and document differences. | First monitoring memo and next-quarter plan. |
Cadence for this sector
Launch temporary tracking sets before major promotions, and remove them after the sale or season ends. Keep one evergreen panel of prompts stable so short-term campaigns do not erase your longer-term trend.
10. Common questions
How many prompts should we begin with?
Start with the 12 examples above as inspiration, but monitor only the topics your organization can act on. Add variants once you know which segments drive useful decisions. A small, well-labeled library is easier to interpret than hundreds of repetitive questions.
Should our brand name appear in the prompt?
Not when your goal is to measure unaided discovery. Use your name for product-specific verification, factual checks and fair comparisons. Report those prompt types separately.
Can stronger website content guarantee an AI recommendation?
No. Clear, accurate and crawlable pages help assistants and search engines interpret information, but provider behavior, retrieval, grounding and answer generation differ. Track observations over time and treat every recommendation as a measured outcome, not a promise.