How Do I Create an AI-Search Prompt Taxonomy for a Multi-Product Company?
To create an AI-search prompt taxonomy for a multi-product company, give every tracked prompt one primary scope and one buying stage, then add tags. The scope is a specific product, the company as a whole, or a cross-product question. The tags cover persona, segment, region and whether the prompt names a brand. Allocate prompts to products by business priority, with enough in each cell you plan to report to give stable numbers. Keep company-level and suite questions as their own group, because they behave differently from product questions. Watch for brand architecture effects. Assistants may credit a product's features to the parent brand, or fail to connect a separately branded product to the company. A clear taxonomy makes each product's gaps visible without double counting.
Which dimensions should the taxonomy include?
| Dimension | Values | Rule |
|---|---|---|
| Primary scope | Product A, Product B, Product C, Company, Cross-product | Exactly one per prompt |
| Buying stage | Problem, Category, Comparison, Vendor evaluation, Pricing, Implementation | Exactly one per prompt |
| Persona | Such as IT buyer, finance lead, end user | One or more tags |
| Segment | Industry, company size | One or more tags |
| Region and language | Markets you track | One per prompt run |
| Brand scope | Unbranded, company-branded, product-branded, competitor-branded | Exactly one per prompt |
| Source | Sales calls, tickets, search data, interviews | For audit, not reporting |
Single-value dimensions keep counts clean: each prompt adds to one product and one stage. Multi-value tags let you slice results without double counting totals.
How do you decide a prompt's scope?
- Product scope: the question can only be answered by one of your products. "Best help desk software for small teams" belongs to your help desk product.
- Company scope: the question is about you as a vendor. "Is [company] a reliable vendor?" or "Where is [company] based?"
- Cross-product scope: the question is about using several products together, or about suites. "Should I buy CRM and help desk from the same vendor?" or "All-in-one platform for sales and support".
- When in doubt, assign the product the buyer is most likely to purchase first, and tag the others.
Write these rules down with examples, so two people would label the same prompt the same way.
What does each stage look like for one product?
Using the help desk product as an example:
| Stage | Example prompt |
|---|---|
| Problem | "Our support team keeps missing customer emails. How do other small companies handle this?" |
| Category | "Best help desk software for a 10-person support team" |
| Comparison | "Help desk tools that integrate with our CRM, compared" or "Alternatives to [competitor] help desk" |
| Vendor evaluation | "Is [product] easy to set up?" or "What do users complain about with [product]?" |
| Pricing | "How much does [product] cost for 10 agents?" |
| Implementation | "How long does it take to migrate from [competitor] to [product]?" |
Problem-stage prompts rarely name any product, so they show whether assistants connect the problem to your category at all. Evaluation and pricing prompts name your product, so they test accuracy more than visibility. Keep them in the set, but read them for what they are.
What does an example layout look like?
Consider an illustrative software company with a CRM, a marketing automation tool and a help desk:
| Scope | Problem | Category | Comparison | Evaluation | Pricing | Implementation | Total |
|---|---|---|---|---|---|---|---|
| CRM | 6 | 10 | 10 | 8 | 4 | 2 | 40 |
| Marketing automation | 4 | 8 | 8 | 6 | 3 | 1 | 30 |
| Help desk | 4 | 8 | 8 | 6 | 3 | 1 | 30 |
| Company | – | – | – | 8 | 2 | – | 10 |
| Cross-product | 2 | 4 | 2 | 2 | – | – | 10 |
| Total | 16 | 30 | 28 | 30 | 12 | 4 | 120 |
The CRM gets the most prompts because it earns the most revenue. Category and comparison stages get the most prompts because that is where recommendations happen. At 120 prompts, no single product-by-stage cell reaches the 15–20 prompt minimum below, so report this set by product row and by stage column and use individual cells for diagnosis.
How many prompts does each cell need?
Enough for the numbers you report to mean something. A cell with four prompts, run five times each, gives 20 answers. Its appearance rate could easily swing 20 points from month to month by chance. Two practical rules:
- Report detailed cells only when they have enough answers, for example at least 15 to 20 prompts, each run several times.
- Roll small cells up to the product level or the stage level for reporting, and use the detail for diagnosis. Even at 15–20 prompts, treat month-to-month changes under about 20 points as noise until a rerun confirms them.
How to measure the chance variation in your own data is covered in how to measure the noise floor for ChatGPT citation tracking.
How does brand architecture change the results?
- Branded house, where products carry the company name, such as "[Company] CRM". Assistants may merge products, crediting one product's features or reviews to another, or to the company generally. Track product-level accuracy, not just mentions.
- House of brands, where products have their own names. Assistants may not know the products belong to the same company, which weakens cross-product and company-level answers. Test "Who makes [product]?" and "[Company] products".
- Mixed architectures need both checks.
Name confusion across products and with other companies is covered in how to stop AI confusing your brand with a similarly named company. Testing new product names before launch is covered in how to check product names in AI search before a rebrand launch.
How should results be reported?
- A product-by-stage grid of appearance or recommendation rates, with cells below the minimum size greyed out.
- Product-level trends on the headline dashboard.
- Company and cross-product results as a separate panel, since they reflect the brand rather than any one product.
- Accuracy by product, such as features and prices attributed correctly.
- Owners per product, who receive their product's gaps as an action list.
How do you keep the taxonomy stable?
- Version the prompt set, and change it on a schedule, as covered in what a golden prompt set is and why to keep it fixed.
- Add new products as a new scope with their own prompts, rather than reshuffling existing ones.
- Retire products by moving their prompts to an archive group, keeping them for occasional checks of how the retirement is described.
- Review labels quarterly with product owners.
What might an illustrative case look like?
Consider an illustrative company whose overall AI visibility report showed a steady 30% appearance rate, which leadership found reassuring. After it adopted a product-by-stage taxonomy, the grid told a different story. The CRM appeared in 45% of category answers, while the help desk appeared in 9% of comparison answers. Assistants also credited the CRM's reporting features to the help desk. The help desk team built comparison pages and corrected feature descriptions on review profiles. The company added "Who makes [help desk]?" to its company-level prompts to track whether assistants linked the two.
What do people ask about prompt taxonomies?
Should every product get the same number of prompts?
No. Allocate by business priority, but give every product enough prompts to report on at least at the product level.
Where do cross-product prompts go in totals?
In their own group. Adding them to a single product inflates that product's results.
How often should labels change?
Rarely. Stable labels keep trends comparable. Review them quarterly and change them only at version updates.
Can one tool handle a multi-product taxonomy?
Most tracking tools let you tag prompts, which is enough if the tags follow your rules. Export the raw answers so you can rebuild the grid yourself if a tool changes.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources