The short answer
- Choose AI tracking prompts from buyer decisions and supporting question evidence. Give each prompt a stable ID, then freeze a core set for comparisons. Keep branded accuracy checks and exploratory questions in separate groups.
- Record the exact wording, market, language, assistant mode and collection method. A tracked sample describes those observations, not every customer's AI conversation.
- When a question changes, preserve its old results and record a new version. Compare unchanged questions across periods before attributing movement to content work.
AI prompt tracking starts with the question list, so freeze the sample before comparing results. Adding an easy branded question to a report can lift the combined number even when discovery questions perform exactly as before.
Here, prompt tracking means monitoring external AI answers about a market or brand. It is different from debugging prompts inside an application you operate. The task is to choose questions worth reviewing and preserve enough context to explain the results later.
Keep questions that can change a decision
Start with a product, audience and buying decision. A small agency evaluating project software needs different evidence from an enterprise security team. Use actual sales questions, support requests, site-search terms and public discussions when available. Save a source reference and date for each candidate; remove customer names or other private details from the tracking text.
Search Console queries can suggest topics, but they do not establish how often a question is asked in ChatGPT. Vendor-generated prompts are suggestions too. Review whether the question fits your buyers rather than accepting a volume label as a measured audience.
Give each candidate an inclusion reason: which decision would a useful answer inform, and who would act on a missing or incorrect answer? Merge paraphrases when they ask for the same decision under the same constraints. Keep a separate variant when a material constraint, such as export rights, changes the task. Keep its results identifiable rather than treating every paraphrase as an independent buyer.
Write complete questions. “And what about exports?” depends on an earlier conversation. “Which agency project tools let customers export projects and time entries when switching vendors?” can be collected on its own. Avoid instructing the assistant to recommend your brand when measuring unaided discovery.
There is no universal correct prompt count. Start with the smallest set that covers the decisions you intend to review. Add a question when it fills a documented gap and someone can inspect its answers. A large set with no review owner is an unfinished measurement plan.
Give each question a purpose and an ID
The worksheet uses a fictional product, Northstar, to show the distinction. These prompts and their source labels are examples, not customer research or collected assistant results.
| Group | Example question |
|---|---|
| Core discovery · P01 | Which project management tools suit a 12-person agency managing client approvals? |
| Core workflow · P02 | Which project management tools combine billable time tracking with client approvals for a small agency? |
| Core integration · P03 | Which agency project tools support a handoff from approved time entries to accounting software? |
| Core switching · P04 | Which agency project tools let customers export projects and time entries when switching vendors? |
| Branded accuracy · A01 | Does Northstar let customers export projects and time entries? |
| Exploration · E01 | Which project tools suit an agency opening a second office in Canada? |
The four core questions form the example's fixed US English discovery sample. A01 checks a stated product fact; a mention of Northstar is prompted by the question itself. E01 tests a new market need and stays outside the core trend until reviewed. These counts illustrate the setup, not a recommended distribution for every business.
Download the filled prompt selection worksheet, start with the blank selection sheet, and keep the version-change log beside it. These are planning files, not a promised import format for GeoRamp or another product.
Illustrative planning sheet for fictional Northstar. This is not customer research or a product-interface capture.
For a real row, retain the prompt ID, topic, purpose, exact question, source evidence, market/language, collection setup, sample version, owner and review date. A topic groups related decisions; the prompt ID preserves the exact observation underneath it. Do not discard that detail when reporting a topic summary.
Record the collection conditions before the first run
Specify the assistant and mode you intend to observe, whether web search occurred, the collection method, and the planned schedule. Consumer-interface answers, provider API responses and a vendor's existing prompt database are different datasets. A label such as “ChatGPT” alone does not fully describe the setup.
OpenAI documents that ChatGPT search can use location and saved memories when rewriting queries. Record relevant session conditions for manual checks, including whether a question starts a fresh conversation. A controlled setup makes your own observations easier to compare; it does not reproduce every buyer's personalized session.
Ahrefs' custom-prompt documentation likewise distinguishes a focused list from its broader dataset and specifies assistant, location and refresh frequency. Check the actual collection definition in whichever tool you use.
Keep failed runs visible and separate them from successful answers with no brand mention. Inspect raw responses and linked sources. The AI search visibility guide explains mention and citation denominators. Once the question set is defined, the workload pricing guide helps cost the collection plan.
Change the set without rewriting the trend
Suppose the four core prompts produce 20 successful observations, five of which mention Northstar. The appearance rate is 25%. Now add five successful observations from the branded accuracy question, all mentioning Northstar. The combined rate becomes 10 of 25, or 40%, while the original core remains five of 20, or 25%.
Those are fictional counts chosen to demonstrate a composition effect. They are not a product test, forecast or estimate of how many people saw an answer. Reporting the combined increase as improved discovery would misdescribe what changed.
At a scheduled review, add new candidates to exploration first. Keep unchanged IDs when wording and setup stay the same. If P02 gains a new budget restriction, create P02b, link it to P02, and record the reason and effective date. Preserve the old wording and observations. If you run both versions during a transition, label that overlap instead of splicing their histories together.
Compare the unchanged cohort over both periods with the same reporting definitions and disclosed completion counts. If no comparable cohort remains, start a new baseline and mark the break. Retire irrelevant questions with a reason; do not quietly remove a valid buyer question because your brand is absent.
When an answer reveals a factual or source gap, assign the evidence and next page decision through the content gap template. A movement after a page edit is an observation to investigate, not proof that the edit caused it.
Leave with a list someone can maintain
Before the first review, check that every active question has a purpose, evidence reference and owner. Keep accuracy and exploration groups separate, preserve collection failures, and make the sample version visible on the report. A spreadsheet is enough if a small manual review remains manageable.
We make GeoRamp. If recurring collection is the bottleneck, bring the completed sheet to a demonstration and check how its questions, collection method and review process map to your requirements. Keep the version register yourself unless the chosen workflow demonstrates that history.
Bring a defined prompt sample to your demo
Review your questions, collection requirements and evidence handoff against the working GeoRamp workflow.
Review your prompt sample in a demo →
