A tracking prompt is a saved question you submit again under recorded conditions to observe how an AI search service answers it. The purpose is to sample a business question consistently. Start with a small panel you can explain and maintain.
Collect questions around customer decisions
Pull candidate questions from sales calls, support tickets, site search, public discussions or keyword and prompt tools. Record where each came from. An exact customer question, a question drafted by your team and a tool-modeled prompt give you different kinds of evidence. Keep customer records private and remove identifying details before reuse.
Sort questions by what the buyer is doing. Discovery asks which options exist. Comparison asks which fits. Purchase asks about constraints such as lead time. Support asks how an existing customer completes a task. Decide which jobs your panel should cover before you add variants.
Screen candidates before they enter the panel
Ask four questions about each candidate.
- Relevance. Does it involve a product, service or market you actually cover?
- Decision. Can you name the choice it helps someone make?
- Specificity. Is there enough context to tell whether an answer fits, without telling the model whom to recommend?
- Action. If the answer is incomplete or wrong, can you identify a page, policy or information owner to check?
Near duplicates can usually share one representative question. Keep two formulations if they test different buying constraints. A dental practice’s booking question and a creator agency’s representation question should have separate topic owners and source pages even if both begin with “How do I choose?”
Build a Prompt Selection Matrix
The matrix below contains authored examples for two fictional businesses. Every selected question tests a different decision. Market and platform are proposed settings. The baseline dates are planned, with no runs or observations implied. Replace the sample date with the first valid run date for your own panel.
| Prompt | Intent or buyer decision | Market | Language | Platform | Priority | Owner | Decision and reason | Baseline Date |
|---|---|---|---|---|---|---|---|---|
| Which dental practices in my area accept new patients? | Discovery: find an eligible provider | US | English | ChatGPT | High | Front desk | Baseline: new-patient eligibility | 2026-12-01 planned |
| Can I book a first appointment online at Cedar Clinic? | Booking: choose the correct route | US | English | ChatGPT | High | Booking lead | Baseline: channel and next step | 2026-12-01 planned |
| What records should I bring to a first visit at Cedar Clinic? | Preparation: know what to bring | US | English | ChatGPT | Medium | Practice manager | Baseline: preparation instructions | 2026-12-01 planned |
| Does Cedar Clinic provide itemized estimates before planned treatment? | Cost planning: assess quote process | US | English | ChatGPT | Medium | Billing lead | Baseline: estimate timing | 2026-12-01 planned |
| Which suppliers sell small batches of plain cartons for online shops? | Sourcing: shortlist suppliers | US | English | ChatGPT | High | Sales | Baseline: small-order fit | 2026-12-01 planned |
| When should a shop choose stock cartons over custom-printed ones? | Comparison: choose product type | US | English | ChatGPT | Medium | Product | Baseline: offer trade-offs | 2026-12-01 planned |
| What artwork does Harbor Carton need before it can quote? | Quote preparation: supply required files | US | English | ChatGPT | High | Prepress | Baseline: intake requirements | 2026-12-01 planned |
| Can Harbor Carton deliver custom packaging before a seasonal launch? | Scheduling: test lead-time fit | US | English | ChatGPT | Medium | Operations | Exploratory: time-sensitive constraint | Pending promotion |
For the clinic, keep the booking and preparation questions separate. One determines how to get an appointment and the other affects what happens after booking. Reject “What’s the best dentist?” unless you can give it a buyer task and useful context. It does not specify the decision the clinic can investigate.
For the supplier, keep the artwork question even though it is branded. It tests a verifiable quote requirement. Exclude “Where can I order a small batch of plain cartons?” from the baseline if it duplicates the first supplier question under the same conditions. Keep it in the candidate log with the reason for exclusion.
Keep the literal wording, origin, platform setting, active status and person responsible for each row. Record a date and version if wording changes. Use the platform named in the matrix only if you can actually run the intended test there. FoundUp AI prompt tracker
Resolve duplicates before freezing the panel
Compare “Which suppliers offer small orders of stock cartons?” with “Where can an online shop order a small batch of plain stock cartons?” If both ask for the same offer and buying decision, choose one as the baseline and save the other in the candidate list. A question about printed cartons earns a separate slot if printing changes the minimum, price or lead time.
Use the customer wording that best preserves the decision. If a candidate contains a brand name, decide whether it is a factual check about that company or a discovery question that gives the assistant an answer in advance. Keep the branded version when the fact check matters and write a separate unbranded question for discovery.
Before activating the panel, ask another teammate to identify the buyer task, the information owner and the page to inspect for every row. Set aside a question if the teammate cannot tell what the answer would help the business check.
Review the panel without losing its history
Maintain a baseline panel for decisions you want to compare over time and an exploratory panel for new questions. Promote an exploratory question when it adds a distinct, ongoing buyer task. Pause questions if the offer ends or the task becomes irrelevant, and keep their history. If you replace a question or change a material setting, start a new series for the affected comparison. Keep old and new results identifiable.
At the next review, record the reason for each change. For the seasonal packaging question, a log might say “pause after the holiday window, retain prior observations, revisit before the next launch cycle.” If you reword it into a year-round question about urgent orders, give it a new version and compare it within its own series.
Turn a prompt finding into a website content task
Illustrative clinic finding. A sampled answer sends a first-time patient to the existing-patient booking route. Compare the saved answer with the booking page and the actual intake instructions. If the two routes are hard to distinguish, assign the booking lead to confirm their labels and the editor to update the text next to each link. Accept the change when a colleague can find the first-visit route without guessing. Keep the original question for a later observation.
Illustrative supplier finding. An answer says an artwork file is optional even though the quote team requires it for custom printing. Check the quote policy and the visible sources. If the product page leaves the requirement unclear, give the prepress owner one task to approve the file requirements and place them beside the quote instructions. Ask someone to use the page to prepare a sample request. A later answer can be checked against the approved wording.
Keep these tasks in a content change log. Store the prompt ID, observed passage, source or missing-source status, page owner and acceptance test so the next reviewer can reconstruct the decision.