# How to Change Tracked Prompts Without Breaking Your AI Visibility Trend

> How to add, reword or retire tracked AI prompts without a misleading visibility trend: a stable core set, overlap windows and like-for-like reporting.

- Source: https://www.citlyze.com/blog/change-tracked-prompts
- Author: Citlyze Research
- Published: 2026-11-20
- Tags: Measurement, Strategy

You can change the prompts you track without breaking your AI visibility trend if you keep a stable core set that never changes, run any reworded prompt beside the old one for at least two windows before retiring the old one, and report the core set's trend separately from the total. The trend breaks when the instrument changes and the report doesn't say so: a mention rate that jumps because you swapped five hard prompts for five easy ones looks exactly like progress.

Prompt sets have to change. Buyers ask new questions, products launch, markets open. This guide covers which changes distort a trend, a simple system for making them safely, and how to report through the change.

## Why prompt changes distort the trend

A mention rate is an average over your prompts. Change the prompts and you change what is being averaged, even if nothing about your brand's visibility moved. Four kinds of change do this:

- **Rewording.** "Best CRM for small teams" and "which CRM should a 10-person startup use" are different prompts to an engine. The answers name different brands, so a reworded prompt is a new series, not a continuation.
- **Adding prompts.** New prompts arrive with their own rates. Add five branded questions and the average rises; add five broad category questions and it falls.
- **Removing prompts.** Retiring the prompts you lose makes the average look better on the day you do it.
- **Changing engines or markets.** Adding an engine or a country changes the mix the same way adding prompts does.

Volatility makes this worse. Rates already move between windows for reasons that have nothing to do with you, as covered in [why AI answers change](https://www.citlyze.com/blog/why-ai-answers-change), so a composition change hides easily inside normal noise.

## A system for changing prompts safely

### Keep a core set that never changes

Pick 15 to 30 prompts that represent your most important buying questions and freeze them for at least a year. This core set is your trend line. Everything else can change around it. When someone asks "did our visibility improve this quarter?", answer from the core set first.

### Run old and new side by side

When a prompt needs rewording, add the new version and keep the old one running for at least two measurement windows. Then compare them. If the rates are close, retire the old one and note the switch date. If they differ a lot, the wording matters to the engines, which is worth knowing in itself, and you should report the two as separate series.

### Give every prompt an ID and a status

Track each prompt with a stable ID, its exact text, the date it was added, its group (category, comparison, branded, problem-led) and whether it is active, in overlap, or retired. When you [export the data](https://www.citlyze.com/blog/export-ai-visibility-data), the ID is what joins months together; the text alone breaks the moment someone fixes a typo.

### Change on a schedule, not on a whim

Review the prompt set quarterly, change it once, and write down why. Ad hoc changes spread over many weeks make it impossible to separate their effect from real movement.

## Where new prompts should come from

Add prompts from evidence of demand, not from guesses:

- **Questions your sales team hears.** The phrasing buyers actually use on calls is the closest thing to what they type into an assistant.
- **First-party reports.** ChatGPT's [Business Hub](https://www.citlyze.com/blog/chatgpt-business-hub) shows the themes of answers that referenced your site, and Bing's AI Performance report shows the grounding queries that retrieved your pages for Copilot. Both point at questions you may not track yet.
- **Search Console.** Long, question-shaped queries that already show your pages are often close to how people phrase prompts.
- **Competitor gaps.** Prompts where competitors appear and you don't, found by running a broad category set once.

The method for building the original set is in [finding buyer prompts for AI search](https://www.citlyze.com/blog/find-buyer-prompts-for-ai-search), and it applies to every addition.

## Reporting through a change

Report two numbers for any period that includes a prompt change:

1. **The like-for-like rate:** prompts that were active in both periods, compared with each other. This is the only number that says whether visibility moved.
2. **The total rate:** every active prompt. This shows where you stand on the full set today.

Then add one line on what changed: "Added 8 comparison prompts on March 1; like-for-like rate up 4 points, total rate down 3 points because the new prompts start low." That sentence prevents most of the bad conversations a changed trend causes. The [monthly report template](https://www.citlyze.com/blog/ai-search-tracking-report-roi) has a slot for it.

## A quick check before any change

Before you add, reword or remove a prompt, ask three questions:

1. Will it change the average of the core set? If yes, don't make the change in the core set.
2. Can the old and new versions overlap for two windows? If not, report the new one as a separate series.
3. Is the change written down with its date and reason? If not, the next person reading the trend can't account for it.

Citlyze's [prompt tracking](https://www.citlyze.com/features/prompt-tracking) records each prompt's answers per window with the run count beside every rate, so comparing an old prompt with its new wording means reading two rows side by side.

---

Canonical: https://www.citlyze.com/blog/change-tracked-prompts
