Prompt Tracking for AI Search: Which Questions to Monitor

Start from what customers say, not what you want to rank for

Keyword lists make a poor starting point. People ask AI engines full questions with context attached: the problem, the place, the budget, sometimes the brand they already distrust. The strongest prompt sets are built from the words customers already use with you, collected from places you already have access to.

• Call recordings and intake notes. The first sentence a caller says is usually the prompt they would type. "My water heater is leaking and I'm in Alamo Heights" beats "water heater repair" as a tracked question.
• Search Console queries. Filter the Performance report for queries containing who, what, how, best or near me. Long, question-shaped queries are prompts already.
• Bing's grounding queries. The AI Performance report in Bing Webmaster Tools lists the phrases Copilot used to retrieve your pages, which shows how a machine rephrases a customer.
• Your reviews. Customers describe the job in plain words ("came out on a Sunday", "found the leak two other companies missed"). Those phrases turn into buying questions.
• Forums and Reddit threads in your trade. The way strangers ask each other for recommendations is the way they ask a chatbot. Our page on Reddit, forums and directories explains why those threads also feed the answers.
• Your sales team's objections. "Is this company more expensive than that one" and "are they licensed" are now asked of a machine before anyone calls you.

Six question types every set needs

A balanced set covers the whole path from a vague problem to a signed estimate. Leave out a type and you lose sight of a stage where customers are choosing. These six cover most local and online businesses.

How big the set should be

Small enough to read, big enough to trust. Every prompt multiplies by engines, runs and locations, and every answer needs a human to check it. Thirty questions on four engines with three runs each is 360 answers a month. That is a morning of reading with a decent spreadsheet, and enough to see a real trend across a quarter.

Tool plans are priced on exactly this arithmetic. As of October 2026, Otterly.AI's Lite plan covers 15 prompts and its Standard plan 100, Semrush's AI Visibility Toolkit includes 25 custom prompts, and Scrunch's Starter plan 350. The GEO tools comparison lists the rest with their published prices.

Engines, locations and a schedule you can keep

Track the engines your customers use, not every engine in existence. For most US local businesses that means Google's AI Overviews and AI Mode, ChatGPT, Gemini and Perplexity. Add Copilot if your buyers sit at office desks on Windows, and Claude for software and professional services. Siri counts too: with the ChatGPT extension on, it can hand some questions to ChatGPT after asking the user.

Location matters more than most dashboards admit. Local questions return local answers, and Google says AI Mode personalizes some results using places a person searched for in Search and Maps. Choose the city or suburb each prompt is asked from, record it, and keep it fixed. A plumber serving Boerne and Helotes should track both rather than one blended San Antonio answer. The generative engine optimization guide explains where each engine finds its local data.

• Weekly for the 10 to 15 buying questions that bring in the most money, especially while fixes are underway.
• Monthly for the rest of the set: problems, comparisons and reputation questions.
• After a fix, rerun the affected questions once the engines have had time to catch up. OpenAI says robots.txt changes take about 24 hours to reach its search systems; a corrected listing or a new directory profile can take weeks to show in answers.
• Quarterly, review the set itself. Drop questions nobody asks, add new services and phrasings, and note every change in the report.

Building the first set in an afternoon

Before the first run, decide what goes in the log, because a log you cannot sort is a scrapbook. For every answer, keep the question exactly as asked, the engine and mode (signed in or not, web search on or off), the date, the location, the full answer text, every business named in order, and every cited URL. Add two columns of judgment: whether you were recommended, listed or warned against, and whether any fact about you was wrong. That is enough to compute every figure a report needs, by hand or in a tool.

1. 1

   Collect raw questions

   Pull 50 to 100 candidates from calls, emails, Search Console, reviews and forum threads. Do not tidy them yet; the rough wording is the useful part.

2. 2

   Sort by type and value

   Tag each one as buying, problem, comparison, price, brand fact or reputation. Star the questions tied to your most profitable services.

3. 3

   Trim to the starting size

   Cut to the size in the table above. Drop near-duplicates: engines treat them alike, and they swell the sample without adding information.

4. 4

   Write them the way people talk

   Full sentences, the customer's words, and a place name wherever a local customer would add one. Leave your brand name out except in comparison, fact and reputation questions.

5. 5

   Set engines, location and schedule

   For each question, record which engines, which location and how often. That record becomes the method section of every future report.

6. 6

   Run a baseline before any work

   Ask every question several times on each engine, log every answer in full, and calculate your starting AI share of voice. Without a baseline, later gains are anecdotes.

Mistakes that quietly ruin the data

• Tracking only your brand name. Engines nearly always describe a business when asked about it by name. It feels like visibility and measures almost nothing.
• Rewording questions every month. A new wording is a new question. Keep the old one and add the new one beside it.
• Testing in your own logged-in account. Memory and personalization shape what you see. Use clean sessions or an API-based tool.
• Treating one run as the truth. One answer is one draw from a hat. Three runs per question per engine is a sensible floor for a small set.
• Ignoring the sources. The cited links show which pages, directories and reviews an engine trusts. That list is your to-do list.
• Chasing informational vanity. Being cited for "what is a heat pump" is pleasant. Being named for "heat pump installer in Converse" pays the rent.

A prompt set is a promise to ask the same questions the same way, month after month. Break it and the chart becomes fiction.

A spreadsheet, a subscription or a program

DataCram builds and runs prompt sets inside its monthly SEO programs, where GEO is included: the questions customers ask go through the engines on a schedule, the answers are tracked over time, and the fixes happen in the same program. Programs start at $625 a month, most local businesses pay between $625 and $2,250, and the tier is set after a free GEO audit. In-house teams that want only the prompt set and a method can ask about GEO consulting. Before any of that, try one question in the free AI visibility check.