Best GEO Software
All posts
By Best GEO Software Teamai prompt trackinghow-toprompt researchgeo

How to do AI prompt tracking: build a prompt set from a template, then keep it healthy

A step-by-step template for AI prompt tracking: five prompt categories, sourcing from Prompt Explorer and Prompt Volumes, tagging, personas, and a maintenance routine.

AI prompt tracking means running a fixed set of buyer questions through ChatGPT, Gemini, Perplexity, Google's AI answers, and other engines on a schedule, and recording which brands each answer names and cites. The tool does the running. The part that decides whether the data means anything is the set itself: which questions, how they're worded, how they're grouped, and how you change them over time without breaking your trend lines.

This guide is about that set. It gives you a template with five prompt categories, shows where to source real wording, how to tag what you add, when personas and locations are worth the extra prompts, and a maintenance routine for the months after. Each step maps to the Promptwatch feature that does it. Once the set is running, how to do AI rank tracking covers reading the results in the first 30 days.

Step 1: fix the frame before you write a prompt

Three decisions shape everything else, and two of them are hard to undo.

The first is market. In Promptwatch, prompts live in monitors, and a monitor's language and location are fixed once you create it. Decide now which country (and, if it matters, which state or city) and which language you're measuring. In Prompt Explorer, the language picker sets what the suggested prompts are written in, and the country sets the market, so a Dutch-language set aimed at Belgium is a real option rather than a workaround.

The second is engines. Pick the ones your buyers use. Promptwatch tracks ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Copilot, Amazon Alexa, and Meta AI through their real interfaces, and models such as Claude, DeepSeek, Mistral, and Grok through their APIs, which shows what they answer from training data. Each paid plan lets you pick from the full list of supported engines, and every engine you add multiplies the responses you use.

The third is a size budget. Your plan's prompt cap is the ceiling: 50 on Essential, 150 on Professional, 350 on Business, unlimited on agency plans (where the response allowance is the limit). Write to about 80% of the cap so you have room for new prompts later. Best AI prompt tracker 2026 matches set sizes to plans, and how many prompts should you track has a sizing method.

Step 2: fill the template

Here's the template we'd start from. The shares are our starting split for a typical B2B or consumer brand, not a rule. Move them once you see the data.

CategoryShare of setPrompt patternPrompt typeIntent
Category questions35%"best [category] for [use case]"OrganicCommercial
Problem and solution25%"how do I [job] without [pain]"OrganicInformational
Comparisons20%"[brand] vs [rival]", "alternatives to [rival]"Competitor ComparisonCommercial
Brand10%"is [brand] good for [use case]"Brand SpecificBranded
Local10% (0 if you have no locations)"best [service] in [city]"OrganicTransactional

For a 50-prompt pilot that works out to roughly 18 category questions, 12 problem prompts, 10 comparisons, 5 brand prompts, and 5 local ones.

Category questions are the core of the set because they measure whether you're named when the buyer hasn't named anyone. Never put your own brand in them, or you're measuring nothing. Vary the use case rather than the wording: "best payroll software for contractors" and "best payroll software for a restaurant" tell you more than five rewordings of one question.

Problem-and-solution prompts catch buyers earlier, before they know the category exists. Read a few of these answers in full: whether you show up in the advice itself or only in a list of tools at the end is worth knowing.

Comparisons name you or a rival directly, so they'll flatter your numbers. Keep them, because they show how AI describes you head to head, but keep them typed separately so they don't inflate the headline. Brand prompts are your reputation check: pricing, fit, and "is it any good" questions. Local prompts only belong if you sell somewhere physical; how to write local prompts goes deeper on wording.

Step 3: source real wording from Prompt Explorer and Prompt Volumes

A template gives you slots. Real buyer language fills them better than brainstorming does.

Prompt Explorer suggests prompts from five places: your Google Search Console queries, the keywords you already rank for, manual search, People Also Ask, and any page on your own site. Search Console is the best starting point for problem-and-solution slots, because those queries are already questions people typed. A product page works well for category slots: point Prompt Explorer at it and it proposes prompts based on that page. New monitors can also begin from a set of suggested starting prompts. If you'd rather not click through suggestions, the Agent Chat side panel works on Prompt Explorer, and asking it to track the ticked suggestions adds them, including into a new monitor.

Prompt Volumes then helps you choose between candidates. It gives each prompt a monthly volume, a difficulty score, and citation rates across AI search, Google, and Bing. Use volume to pick between two phrasings of the same slot and to cut prompts nobody asks. Use difficulty to balance the set so it isn't all unwinnable head terms or all easy wins. For the background on how those numbers are estimated, see how prompt volumes and difficulty scores are calculated.

Step 4: tag every prompt on the way in

Tagging after the fact is miserable, so do it as prompts enter the set.

Topics come first. Make one topic per product line or buyer job, not per category in the template. The Topics page then reports Visibility, Citation Share, Average Position, Share of Voice, and Sentiment per topic, with By Competitor and By Rank views. The Manage tab is where you build them: create a topic, name it, pick its color, and tick its prompts in one place, searching or filtering by monitor. A Suggested tab proposes prompts that share words with the topic and shows the matched words. Nothing saves until you click Save changes, so you can rearrange freely.

Then set prompt type and intent straight from the template columns. Prompt types are Organic, Brand Specific, and Competitor Comparison; intent is Branded, Informational, Navigational, Commercial, or Transactional. This is what lets you report category visibility without comparison prompts padding it. Our post on segmenting branded, organic, and competitor prompts explains why that split matters in reporting.

Use tags for everything else. We'd add two from day one: a "core" tag for the prompts you'll never edit (more on that below) and a batch tag with the month a prompt was added. Bulk editing makes both quick.

Step 5: add personas and places only where answers differ

Personas and locations multiply your set. Forty prompts as three personas is 120. So test before you multiply.

Take five category questions and run them as two personas in Promptwatch, for example a CFO and an IT lead. With persona stacking turned on for a monitor, Promptwatch wraps the persona's context around each prompt before sending it, so the engine answers as if that buyer had asked. After a few days, open Persona Insights: it draws a radar per persona against the general audience (answers run without a persona) and shows the biggest gap in a card. If both personas see the same shortlist as the general audience, skip personas for now. If one sees a different set of brands, that persona earns its own copy of the category slots.

Locations work the same way. Country-level tracking is on every paid plan. State and city start on Business and are on every agency plan, and coordinate targeting (a pin with a 20 m to 5 km radius) is on Business, Agency Scale, and Enterprise. For a second country, Copy Prompts copies a set to a monitor in another country if the language matches, and it can create that monitor for you. For a second language, write the prompts natively instead of translating the first set word for word.

Step 6: read the fan-outs after the first runs

Once a few days of answers are in, open the query fan-outs. Promptwatch shows, per model, the exact sub-queries an engine ran behind each response. This tells you whether a prompt triggers the searches you assumed. A category question you meant as "software for contractors" might fan out into pricing and review searches, which is useful to know when you decide what content to publish.

Treat fan-outs as content targets, not as new prompts. Adding every sub-query to the set bloats it with phrasings no buyer types. And re-check them after engine changes: Promptwatch's ChatGPT site: operator report measured the average number of fan-out queries per ChatGPT response going from about 1.08 to about 1.83 on 8 August 2026 (chart window 26 July to 10 August 2026), and the report advises re-running fan-out analyses for tracked prompts.

Step 7: run a maintenance routine

A prompt set drifts. Products launch, competitors appear, and prompts that seemed right in month one turn out to measure nothing. A light routine keeps it honest.

WhenTaskWhere in Promptwatch
WeeklyRead full responses for prompts that movedResponse view per prompt
WeeklyCheck usage against your quotaUsage Overview, alert at 85%
MonthlyAdd new prompts in one batchPrompt Explorer, Prompt Volumes
MonthlyRetire prompts that never name any brandResponses export
QuarterlyRe-check fan-outs, personas, and locationsQuery fan-outs, Persona Insights
QuarterlyCheck the plan still fits the setUsage Overview

A few rules make the routine work. Freeze the core set: never reword a prompt tagged "core", because a reworded prompt is a different measurement and its history no longer compares. If the wording is wrong, add the new version and retire the old one. Add prompts in monthly batches with a batch tag, so you can read trend lines with and without the newcomers. Use the Responses export to audit the set; it lists each brand mention with its position, like Acme (1), Globex (2), so prompts where answers name no brands at all stand out quickly. And remember that topic prompt counts leave out prompts on paused monitors, so pausing a monitor quietly shrinks your topics.

When you do move plans, a downgrade lets the owner pick which prompts keep running. Pick the core set first.

Where to start

The template, the tagging, and the routine work in any tool. What makes Promptwatch the practical choice is that each step has a feature behind it: Prompt Explorer and Prompt Volumes to source the set, the Topics Manage tab to organize it, personas and Persona Insights to test segments, fan-outs to check what each prompt triggers, and usage alerts to warn you before the quota runs out. If you're still comparing tools, AI prompt tracker compared shows how each one counts prompts and responses, and what are the best AI rank trackers has the overall ranking.

Start the 7-day free trial on monthly billing, build the first 50 prompts from the template on day one, and tag them as you go. By the end of the week you'll have enough runs to see which slots earn their place.