How AI reads this page
ChatGPT, Claude, Gemini and Perplexity don't look at colours, images or animations. They receive text, headings, links and structured data. Below is exactly what they read when they open “Which prompts should you track in AI answers? How to pick the right questions”.
The AI receives 1,914 words in 8 sections, 12 links and 3 blocks of structured data.
Text
The page content in Markdown, without menus or graphics. It is the format closest to how a language model reads text.
# Which prompts should you track in AI answers? How to pick the right questions > How to build the prompt list for monitoring your brand in ChatGPT, Gemini and Google AI Mode: prompt types, how many, how to phrase them and why to repeat them. URL: https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere Language: English Version IT: https://geosnap.ai/blog/quali-prompt-monitorare-risposte-ai-come-scegliere Author: Rinald Sefa, CMO Geosnap · Category: Product Publisher: Geosnap (Maind Group S.r.l.), https://geosnap.ai ## In short The prompts you track are not your keywords: questions put to AI are three times longer and mostly match no keyword at all. You need 10 to 30 stable questions, written the way a customer would say them, mostly without your brand name, spread across category, comparison, evaluation, problem and local intents. And they must be repeated: the same prompt gets different answers, so what counts is frequency over time, not a single answer. - A Google AI Mode search is three times longer than a classic search, and ChatGPT's search prompts grew from 4.7 to 8.7 words in a year. - Between 65% and 85% of ChatGPT prompts match no traditional keyword: pasting your keyword list measures questions nobody asks. - At least two thirds of your prompts should leave out your brand name: that is where you make the shortlist, or don't, for someone who is choosing. - Asked again two days later, Claude kept 30% of its cited sources, Gemini 38% and Perplexity 65%: you have to repeat prompts and look at frequency. - Keep the list stable for at least three or four weeks and review it quarterly, without touching the core you use to compare months. Written by the Geosnap AI agent, reviewed and approved by Rinald Sefa. Translated from the Italian original. [Read the original](https://geosnap.ai/blog/quali-prompt-monitorare-risposte-ai-come-scegliere) An AI visibility tool tells you how often your brand shows up in answers from ChatGPT, Gemini or Google AI Mode. That number is only as good as the questions you chose to ask. Ten badly chosen prompts give you a reassuring, useless report; ten good ones show you where you are missing, who gets cited instead of you and what to fix. This guide explains how to build the list: which types of questions you need, how many are enough, how to phrase them and why they must be repeated over time. ## Why isn't a keyword list enough to pick prompts? Because people don't talk to AI the way they type into Google. Google says it with one number: according to its [first-year review of AI Mode](https://blog.google/products-and-platforms/products/search/ai-mode-us-insights/) (May 2026), the average AI Mode search is triple the length of a traditional search query, and the fastest-growing questions are about planning and choosing: "where to", "ideas for", "where should I". ChatGPT is moving the same way. [Semrush's study](https://www.semrush.com/blog/chatgpt-search-insights/) of more than a billion clickstream rows (April 2026) found that prompts triggering web search grew from 4.7 to 8.7 words in a year, and that for most of the period between 65% and 85% of prompts matched no keyword in a traditional search database. In other words: if you paste your keyword list into a monitoring tool, you are measuring questions almost nobody asks. There is a third layer. When ChatGPT receives a question, it splits it into several internal searches (the so-called *query fan-out*). [Peec's analysis](https://docs.peec.ai/research/chatgpt-fanout-length-trend) of more than 20 million fan-outs (February 2026) counts 2.3 to 2.8 per prompt, with the average length rising from about 6 to about 12 words in four months. Internal searches are getting more specific, so generic pages are harder to retrieve. A good monitoring prompt has to reflect that specificity, not a three-word keyword. ## Which types of prompts should you track? You need at least five types, and most of them without your brand name. Someone typing "what's the best practice management software for a dental clinic" doesn't know you yet: that is where visibility is won or lost. Prompts with your name only measure what the AI thinks of people who already look for you. | Prompt type | Example (made up) | What it measures | | --- | --- | --- | | Category, no brand | "What's the best reservation software for a 40-seat restaurant?" | Whether you make the shortlist the AI gives to someone choosing | | Comparison | "Is an agency or a freelancer better for running a hotel's social media?" | How the AI describes your category against the alternatives | | Evaluation | "Is it worth paying a consultant for e-invoicing compliance?" | The arguments and objections the AI attaches to your offer | | Problem and how-to | "How do I reduce no-shows at my beauty salon?" | Whether you are named as the solution to a concrete problem | | Local and planning | "Where can I find an evening English course in Bologna under 500 euros?" | The "where" and "ideas for" questions Google sees growing fastest | | Brand | "What do you know about company X and who is it for?" | Accuracy and sentiment of what the AI says about you | Comparison questions and "best X" lists deserve a permanent slot: they are the format AI engines cite most on commercial questions, as we saw in our article on [comparison pages](https://geosnap.ai/en/blog/pagine-confronto-migliori-x-perche-ai-le-cita). "Where" and "ideas for" questions are the ones Google flags as growing: in the same AI Mode review, planning searches grew 80% faster than the average over the past six months. ## How many prompts do you need, and how do you write them? For a brand with a clear offer, 10 to 30 stable prompts, written the way a customer would say them, are enough. Coverage matters more than count: every type in the table, every main product line and, if you serve several areas, every geography. A travel agency in Turin with three products needs more prompts than an online shop with a single catalogue. Where to get them. [Orbit Media's analysis](https://www.orbitmedia.com/blog/ai-citation-sources/) of 13,184 citations (September 2026) recommends starting from real customer conversations rather than keyword tools: sales calls, support tickets, questions that arrive through the contact form. Three more practical sources: Google Search Console queries rewritten as full questions, your site's FAQ, and the questions customers ask at the first meeting. How to write them. A formula that works: - **8 to 15 words**, in line with ChatGPT's search prompts and the average length of AI Mode searches. - **One concrete context**: company size, city, budget, a constraint. "Best CRM" is a keyword; "best CRM for a five-person law firm" is a prompt. - **Variants of the same question**: different wording, different specificity, different angle. The AI answers "best" and "cheapest" differently. - **No brand name** in at least two thirds of the list. If you use Geosnap, the starting list is built this way: from your website and country the system works out your sector and offer, generates variants for each level of intent and phrases them conversationally (the method is described in the [prompt simulation logic](https://geosnap.ai/en/doc/prompt-simulation-logic)). You can then add your own custom prompts: the most direct way to include the questions you actually hear from customers. ## Why does the same question get different answers, and how do you measure anyway? Because AI engines are not deterministic: the same prompt, two days later, cites partly different sources. In Orbit Media's analysis, repeating 72 questions two days apart, Claude kept 30% of the URLs it had cited, Gemini 38% of its domains and Perplexity 65%. The four platforms tested cited the same domain for the same question only 1.7% of the time. And a mention is not a citation: on ChatGPT the brand was named in 56% of answers but its page cited in 47%; on Gemini, 72% against 41%. Three practical consequences: 1. **Run every prompt several times** and look at frequency, not at a single answer. One answer is weather; the average of several answers is climate. That is why Geosnap collects several answers per prompt, each with no previous conversation context, and works on aggregated data. 2. **Keep the list stable for at least three or four weeks** before you judge a trend. If you change the prompts every week you are not measuring your visibility, you are measuring your list. 3. **Track mentions and citations separately**, per engine. A well-known brand can be named often without its page being cited: that is not the same problem as a brand that is absent. ## The most common mistakes when choosing prompts The most frequent mistake is tracking only questions that contain your own name: the report will be green and say little. Then, in order of frequency: - **Copying keywords.** "Accounting software pricing" is not a question anyone asks an AI assistant. - **Prompts that are too generic.** "What's the best software" gets a different answer every time and has nothing to do with your customer. - **Changing the list every week.** You lose the time series, the only thing that separates a real drop from noise. - **Looking only at your own brand.** The sources cited instead of you (third-party comparisons, forums, trade publications) are the list of pages you need to get into: our guide on [getting cited by Google AI Mode](https://geosnap.ai/en/blog/google-ai-mode-fonti-come-farsi-citare) explains where to start. - **Reading a single engine.** ChatGPT, Gemini and Google AI Mode draw on different sources: a prompt where you are present on one and absent on another is information, not a measurement error. A well-built prompt list is an afternoon's work and lasts for months. Review it every quarter: add the new questions you hear from customers, drop the ones nobody asks any more, but leave the core alone, because that is what lets you compare one month with the next. ## Sources - [How AI Mode is changing and expanding the way people search](https://blog.google/products-and-platforms/products/search/ai-mode-us-insights/) (Google (The Keyword), 2026-05-19) - [ChatGPT search insights: 1 billion+ rows of clickstream data](https://www.semrush.com/blog/chatgpt-search-insights/) (Semrush, 2026-04-07) - [ChatGPT fan-out queries are getting longer and more specific](https://docs.peec.ai/research/chatgpt-fanout-length-trend) (Peec, 2026-02) - [I tracked 13,184 AI citations across 4 LLMs: here's what I learned](https://www.orbitmedia.com/blog/ai-citation-sources/) (Orbit Media Studios, 2026-09-02) ## Frequently asked questions ### How many prompts should I track? For a brand with a clear offer, 10 to 30 stable prompts are enough. Coverage matters more than count: every question type (category, comparison, evaluation, problem, local, brand), every main product line and every area you serve. A company with several products or locations will need more. ### Should I include my brand name in the prompts? In a minority of the list, to check the accuracy and sentiment of what the AI says about you. At least two thirds of the prompts should be category questions with no brand, because that is where someone who is choosing meets your name, or doesn't. ### Why does the same prompt give a different answer every time? AI engines are not deterministic, and the internal searches they run change with every execution. In a test repeated two days apart, Claude kept 30% of its cited sources, Gemini 38% and Perplexity 65%. That is why each prompt should run several times and you should look at how often you appear, not at a single answer. ### How often should I update the prompt list? Not before three or four weeks, or you won't be able to tell a trend from noise. A quarterly review is a good rhythm: add the new questions you hear from customers and drop outdated ones, but keep a stable core so you can compare months with each other. ### Can I use Google Search Console queries as prompts? Yes, but rewritten as full questions with a concrete context. "dental practice management software" becomes "what's the best practice management software for a dental clinic with two chairs?". Queries with many impressions and few AI citations are an excellent starting point.
Structured data
Schema.org data (JSON-LD) embedded in the page code. It tells the AI explicitly what the page is about and who published it.
BreadcrumbList describes the path through the site
{
"@context": "https://schema.org",
"@type": "BreadcrumbList",
"itemListElement": [
{
"@type": "ListItem",
"position": 1,
"name": "Home",
"item": "https://geosnap.ai/en/"
},
{
"@type": "ListItem",
"position": 2,
"name": "Blog",
"item": "https://geosnap.ai/en/blog/"
},
{
"@type": "ListItem",
"position": 3,
"name": "Product",
"item": "https://geosnap.ai/en/blog/#prodotto"
},
{
"@type": "ListItem",
"position": 4,
"name": "Which prompts should you track in AI answers? How to pick the right questions",
"item": "https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere"
}
]
}
BlogPosting describes the article: author, dates and sources
{
"@context": "https://schema.org",
"@type": "BlogPosting",
"@id": "https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere#article",
"headline": "Which prompts should you track in AI answers? How to pick the right questions",
"description": "How to build the prompt list for monitoring your brand in ChatGPT, Gemini and Google AI Mode: prompt types, how many, how to phrase them and why to repeat them.",
"inLanguage": "en",
"url": "https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere",
"mainEntityOfPage": "https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere",
"image": "https://geosnap.ai/img/blog/p2026-10-15-prompt-monitoraggio-en.webp",
"articleSection": "Product",
"wordCount": 1342,
"author": {
"@type": "Person",
"name": "Rinald Sefa",
"jobTitle": "CMO",
"url": "https://geosnap.ai/en/about",
"worksFor": {
"@id": "https://geosnap.ai/#organization"
}
},
"publisher": {
"@id": "https://geosnap.ai/#organization"
},
"isAccessibleForFree": true,
"abstract": "The prompts you track are not your keywords: questions put to AI are three times longer and mostly match no keyword at all. You need 10 to 30 stable questions, written the way a customer would say them, mostly without your brand name, spread across category, comparison, evaluation, problem and local intents. And they must be repeated: the same prompt gets different answers, so what counts is frequency over time, not a single answer.",
"datePublished": "2026-10-15",
"dateModified": "2026-10-15",
"citation": [
"https://blog.google/products-and-platforms/products/search/ai-mode-us-insights/",
"https://www.semrush.com/blog/chatgpt-search-insights/",
"https://docs.peec.ai/research/chatgpt-fanout-length-trend",
"https://www.orbitmedia.com/blog/ai-citation-sources/"
],
"translationOfWork": {
"@id": "https://geosnap.ai/blog/quali-prompt-monitorare-risposte-ai-come-scegliere#article"
}
}
FAQPage describes the frequently asked questions
{
"@context": "https://schema.org",
"@type": "FAQPage",
"@id": "https://geosnap.ai/en/blog/quali-prompt-monitorare-risposte-ai-come-scegliere#faq",
"mainEntity": [
{
"@type": "Question",
"name": "How many prompts should I track?",
"acceptedAnswer": {
"@type": "Answer",
"text": "For a brand with a clear offer, 10 to 30 stable prompts are enough. Coverage matters more than count: every question type (category, comparison, evaluation, problem, local, brand), every main product line and every area you serve. A company with several products or locations will need more."
}
},
{
"@type": "Question",
"name": "Should I include my brand name in the prompts?",
"acceptedAnswer": {
"@type": "Answer",
"text": "In a minority of the list, to check the accuracy and sentiment of what the AI says about you. At least two thirds of the prompts should be category questions with no brand, because that is where someone who is choosing meets your name, or doesn't."
}
},
{
"@type": "Question",
"name": "Why does the same prompt give a different answer every time?",
"acceptedAnswer": {
"@type": "Answer",
"text": "AI engines are not deterministic, and the internal searches they run change with every execution. In a test repeated two days apart, Claude kept 30% of its cited sources, Gemini 38% and Perplexity 65%. That is why each prompt should run several times and you should look at how often you appear, not at a single answer."
}
},
{
"@type": "Question",
"name": "How often should I update the prompt list?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Not before three or four weeks, or you won't be able to tell a trend from noise. A quarterly review is a good rhythm: add the new questions you hear from customers and drop outdated ones, but keep a stable core so you can compare months with each other."
}
},
{
"@type": "Question",
"name": "Can I use Google Search Console queries as prompts?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Yes, but rewritten as full questions with a concrete context. \"dental practice management software\" becomes \"what's the best practice management software for a dental clinic with two chairs?\". Queries with many impressions and few AI citations are an excellent starting point."
}
}
]
}
Metadata
The information in the page header, which crawlers and engines read before the content.