For prompt tracking, pick the questions a buyer asks before they know your name. Most should be "best X for Y" and "alternatives to X" questions, plus the problems that lead people to them. Start with about 25, keep your brand's name out of all but one, ask each one several times, and don't reword them once tracking starts.
Prompt tracking only tells you something if the prompts are ones your buyers would ask. Your mention rate is the share of answers to your list that name you, so the list sets the number. Questions you already win will flatter you, and questions no buyer asks tell you little.
Below are eight kinds of prompt, each with an example for four businesses: a help desk app for small online stores, an uptime monitoring service, an accounting firm for startups and a web design agency that builds in Webflow. The nine guides Google ranked or cited for "prompt tracking" and two related searches, when I read them on 29 September 2026 (US, desktop), take their examples mostly from shops, local businesses and software, and I found none for an agency or an accounting firm. Their advice runs from 5 prompts to 150 per topic and up to 100 runs of each, and one says you would need to repeat each prompt "dozens or hundreds of times, depending on the level of accuracy you need." None of the nine works out how precise a rate is for a given number of answers. After the eight kinds: how many to track, where to find the words buyers use, and four checks before a prompt goes on the list.
Find out what the answers say about you.
One measurement, your own crawler data, and the three access findings that matter most. Free, no card, about two minutes.
Start freeThe eight kinds of prompt at a glance
| # | Kind | Example (help desk app) | What it tells you |
|---|---|---|---|
| 1 | Best for a situation | best help desk software for a small Shopify store | Whether you're named when a buyer is choosing |
| 2 | With a constraint | help desk that works with Shopify for under $50 a month | Whether you're named for the things you win on |
| 3 | Alternatives and comparisons | Zendesk alternatives for a three-person support team | Whether you're named beside the incumbents |
| 4 | The problem | how do I stop answering the same order-status emails every day | Whether you're found before the buyer knows the category |
| 5 | How to choose | what should I look for in help desk software for an online store | Whether you're named while the shortlist forms |
| 6 | Your market | help desk software for online stores in the UK | Whether you're named where you sell |
| 7 | What the category is | what is a shared inbox for customer support | Whether your pages are cited as a source |
| 8 | Your brand's name | is your brand good for small Shopify stores | What the answer says about you |
Answers to kinds 1 to 3 tend to name competitors side by side, so most of the list belongs there.
1. Best for a situation
The buyer knows the category and says who they are. This kind is usually the one most worth winning. The situation is what makes it yours: "help desk for a small Shopify store" and "help desk for a call centre" can get different answers. Write the situations your best customers are in, not the whole category.
- Help desk app: best help desk software for a small Shopify store
- Uptime monitoring: best uptime monitoring tool for a small SaaS team
- Accounting firm: best accounting firm for a seed-stage startup
- Web design agency: best Webflow agency for a B2B SaaS website
2. The same question, with a constraint
Buyers add what they care about most: a price ceiling, an integration, a deadline, a requirement. Pick the constraints you win on. If you're cheaper than the incumbent, or you plug into the tool your buyers already use, those are the answers to be in, and the ones your own pages can say the most about.
- Help desk app: help desk that works with Shopify for under $50 a month
- Uptime monitoring: uptime monitor with a public status page and Slack alerts
- Accounting firm: startup accountant who can handle R&D tax credits
- Web design agency: Webflow agency that can launch a site in four weeks
3. Alternatives and comparisons
"Alternatives to X" and "X vs Y" are asked by someone weighing options. Use the names your buyers are leaving or comparing you with. A head-to-head between two competitors is worth one slot too, because an answer may add a third option, and you'll want to know whether it's you.
For a service business the comparison is often with a way of working rather than another firm.
- Help desk app: Zendesk alternatives for a three-person support team
- Uptime monitoring: Pingdom alternatives for a small team
- Accounting firm: should our startup hire a bookkeeper or use an accounting firm
- Web design agency: Webflow agency vs a freelancer for a SaaS website redesign
4. The problem, before the category
Some buyers don't know what the thing they need is called. They describe what's going wrong. These prompts show whether you're found at the earliest point, and they're the easiest kind to take from sales calls, because they're close to what people say out loud.
- Help desk app: how do I stop answering the same order-status emails every day
- Uptime monitoring: how do I find out my website is down before customers tell me
- Accounting firm: our books are a mess and we're about to raise money, what should we do
- Web design agency: how can our marketing team change the website without waiting for developers
5. How to choose
"What should I look for in X" comes from someone building a shortlist. The answer may list what to compare and name products as examples. Two of these are usually enough, since they sit further from a decision than kinds 1 to 3.
- Help desk app: what should I look for in help desk software for an online store
- Uptime monitoring: how do I choose an uptime monitoring tool
- Accounting firm: what should I ask a startup accountant before hiring one
- Web design agency: how do I choose a Webflow agency
6. Your market: place, industry or stack
If you sell in one country, to one industry or on one technology, put it in the prompt. For a firm with an office, that's the city. For a developer tool it's often the stack.
- Help desk app: help desk software for online stores in the UK
- Uptime monitoring: uptime monitoring for Laravel apps
- Accounting firm: startup accountant in Austin
- Web design agency: Webflow agency in London
7. What the category is
"What is a shared inbox" comes from someone learning. Answers to these can explain the idea without naming any product, so a low mention rate on them may say little about you. What they can show is whether your pages are among the sources, so track one or two if you publish guides.
- Help desk app: what is a shared inbox for customer support
- Uptime monitoring: what is synthetic monitoring
- Accounting firm: what is revenue recognition for a SaaS startup
- Web design agency: what does a Webflow agency do
8. Your brand's name, tracked apart
A prompt with your name in it is answered about you, so it will very likely name you. That makes it a poor measure of whether buyers find you, and a good way to read what the answer says: your price, who you're for, a feature you dropped last year. "Your brand vs Zendesk" counts here too.
Keep it to one, and read its answers on their own. With 25 prompts, one that names you every time can lift the mention rate by up to 4 points (1 in 25), though every buyer who asks it already knows your name.
- Help desk app: is your brand good for small Shopify stores
- Uptime monitoring: your brand pricing
- Accounting firm: your firm reviews
- Web design agency: has your agency built sites for SaaS companies
How many prompts to track, and how often
Start with coverage. For the same number of answers, more prompts beat more runs of a few: ten runs of five prompts measure you on five questions, and your buyers ask more than five. Runs still matter, because single answers vary, so the plan needs both.
Then check what a plan buys you. Here is the 95% range (a Wilson score interval) on one engine, for a rate where one answer in five names you:
| Answers | One way to get there | Range when 20% name you |
|---|---|---|
| 10 | 10 prompts, asked once | 6% to 51% |
| 30 | 10 prompts, 3 runs each | 10% to 37% |
| 75 | 25 prompts, 3 runs each | 13% to 30% |
| 150 | 25 prompts, 3 runs a week for two weeks | 14% to 27% |
With ten answers, a 10% rate and a 50% rate have overlapping ranges. At 75 a week you have a working number, and it tightens as the weeks add up. The AI visibility guide covers when a move in the rate is more than noise.
A split to start from for 25 prompts. It's the one I'd use, not one I've seen tested:
| Kind | Prompts |
|---|---|
| 1. Best for a situation | 6 |
| 2. With a constraint | 4 |
| 3. Alternatives and comparisons | 5 |
| 4. The problem | 4 |
| 5. How to choose | 2 |
| 6. Your market | 2 |
| 7. What the category is | 1 |
| 8. Your brand's name | 1 |
Then keep the wording fixed. A reworded prompt is a new question, and its count starts again from zero.
Where to find the questions buyers ask
Sales calls and pre-sale emails
This is usually the best source, and the one only you have. Take the first thing a prospect says about why they're looking, drop your name, and keep their situation. "We're three people and drowning in where's-my-order emails" gives you a problem prompt and a best-for prompt. If you ask new customers how they found you, the ones who say ChatGPT may be able to tell you what they asked.
Search Console
Google counts AI Overview and AI Mode traffic in the Performance report, under the Web search type, and says a user who asks a follow-up question in AI Mode is "essentially performing a new query." So long, question-shaped queries in your Queries table are worth reading, though I haven't found a filter that shows which came from AI Mode. Search Console's Generative AI performance report breaks impressions down by page, country and device, but not by query.
To pull out the long queries, add a query filter, choose Custom (regex), and use:
^(\S+\s+){7,}\S+$
That keeps queries of eight words or more, counting words by the spaces between them. Google leaves rare queries out of the table to protect searchers' privacy, and doesn't list every other query either, so expect gaps.
Bing Webmaster Tools
Bing's AI Performance report, launched in public preview in February 2026, lists grounding queries. In Bing's words, they are "the key phrases the AI used when retrieving content that was referenced in AI-generated answers." The report covers Microsoft Copilot, Bing's AI summaries and what Bing calls "select partner integrations", and Bing says the grounding-query data is a sample.
Two limits. These are the phrases the AI searched with, which may not be what the person typed. And they come from answers that referenced your pages, so they show where you already appear, not where you're missing. Use them to check your wording, not as the list.
Forums and communities
Recommendation threads on Reddit and in industry communities read a lot like prompts: a situation, a constraint, "what do you use?". A thread title can often go on the list with light editing.
A drafted starting list
When you add a brand in Shruwd, the tool we are building, it reads your homepage and writes a first draft of ten prompts across buying, comparison, problem and learning questions, plus the competitors to track, for you to edit before anything is saved. It's set up to leave your name out of all but one of them. The docs on prompts explain the intents and what an edit does to a prompt's history. Whatever writes your first draft, rewrite it in the words your buyers used on the calls above.
Four checks before a prompt goes on the list
- Would a buyer ask it before they know you? If it only makes sense to someone who has heard of you, it belongs with kind 8.
- Does the question call for a product or a firm? If the honest answer is a definition, the prompt will mostly measure citations, not mentions. If the answer could name products and names none you track, keep it: the question may still be open.
- Does Google show an AI Overview for it? Search it. Google says AI Overviews "often don't trigger". A prompt with no AI answer gives you nothing to count on that engine, and Shruwd's prompt list shows a "with no AI answer" count for each prompt, per engine, so you can see which ones to swap.
- Would you do anything if it moved? If a drop on a prompt wouldn't change what you write or fix, it's using a slot that another prompt could fill.
For check 2, ask ChatGPT from a logged-out browser window, so your own memories and past chats don't shape the answer. One answer is usually enough to judge the prompt. It isn't enough to judge where you stand, and neither are ten: ten answers at a 20% rate leave a range of 6% to 51%. Why ChatGPT recommends your competitor covers counting runs by hand.
What prompt tracking can and can't tell you
It can tell you how often you're named across your list, beside your competitors, and whether that's changing. It won't show you the answer one buyer got. Their ChatGPT can draw on their saved memories and past chats, while a tracker asks from a clean start. Shruwd, for example, asks OpenAI's model through its API with web search available, not through the ChatGPT app.
Frequently asked questions
What is AI prompt tracking?
Prompt tracking means asking ChatGPT, Google's AI Overviews or another AI engine the same list of questions on a schedule and logging which brands and sources each answer names. The share of answers that name you is your mention rate, and set beside your competitors' on the same list, it shows who the answers favour, week by week.
How many prompts should I track?
Start with about 25, most of them "best X for Y" and "alternatives" questions, and ask each three times a week. On one engine, that puts a 20% mention rate between about 13% and 30%. When you want a tighter number, add prompts before runs, because each new prompt covers another question your buyers ask.
Should I track prompts that include my brand name?
Keep it to one. A question with your name in it is answered about you, so it says little about whether buyers find you. Use it to check what the answer says about your price and your product, and read it on its own.
Is there search volume for prompts?
I haven't found one from OpenAI. Search Console counts AI Mode questions with ordinary searches in its Queries table, and I haven't found a way to separate them. Tools that show prompt volume estimate it, for example by turning search keywords into questions. Use it as a guide to wording, and rank prompts by how close they are to a sale.
How do I track brand mentions in AI answers?
By hand, with a spreadsheet; past a handful of prompts, with a tool (the tools list compares nine). Log one row per answer: the prompt, the engine, the date, which brands it names and which sites it links. Ask ChatGPT from a logged-out browser window, and read Google's AI Overview for the same prompt. Your mention rate is the rows that name you over all the rows, and it needs a range: 30 rows at 20% still span about 10% to 37%.
