Definition
What is prompt tracking?
Prompt tracking is the repeated measurement of how AI engines answer a fixed list of questions about your market. Each prompt is run on each engine and country you follow, and every answer is stored with the brands it names, their position and the sources it cites. Over time, the set shows whether your brand is gaining or losing ground in AI answers.
The prompt became the unit because the question became the input. People type short queries into a search box and full questions into a chat window, often with the details that matter to them: team size, budget, the tools they already use.
AI engines work from that full question. OpenAI says ChatGPT usually rewrites a question into one or more targeted searches before answering, and that saved memories and approximate location can shape those searches[1]. Google says AI Overviews and AI Mode may run several related searches across subtopics to build a single response[2]. Two buyers can reach the same category through different questions, and each gets an answer written for the question as asked.
A tracked prompt records the question the way a buyer would ask it, for example “best project management tool for creative agencies”, run in the United States, in English. Each prompt also carries a type, such as commercial, comparison or branded, and the tags that group it with similar prompts, so results can be read by theme as well as one by one.
Comparison
Prompt tracking and keyword rank tracking differ on 5 points, starting with the unit
Prompt tracking follows full questions inside AI answers, while keyword rank tracking follows short queries on a results page. A rank tracker records where a URL sits in a list. A prompt tracker records whether a brand is named, how early, and which sources back the answer. The two measure different surfaces, so a search team can run both on the same topics.
| Criterion | Prompt tracking | Keyword rank tracking |
|---|---|---|
| Unit tracked | A full question, in a buyer's words | A short query |
| What is recorded | The brands named, their position, the sources cited | The position of a URL |
| Main metrics | Visibility, share of voice, average position | Rank, estimated clicks |
| Same input, same output | No: answers vary between runs | Largely, for a given place and time |
| Runs per check | Several per prompt | One per query |
The gap shows up in the data. A 2025 study compared Google's organic results with five generative search systems from Google, OpenAI and Perplexity. On average, 53% of the domains AI Overviews drew on sat outside the organic top 10[3]. A page that ranks well for a keyword has no guarantee of feeding the answer, or of getting its brand named in it. Rankings remain worth tracking, since engines still search the web, but they are a weak guide to which pages an answer will use.
Picture a rank tracker that shows Loopwork's pricing page third for “project management software pricing”. A buyer asks an AI engine how much project management software costs per seat for a 20-person agency and gets an answer that names three tools and their price ranges, with or without Loopwork. Third place on the results page says nothing about that answer. The prompt tracker does, and it keeps the pages the engine cited, which leads back to pages the team can work on.
Sizing
A useful prompt set holds 80 to 300 prompts per market
A prompt set should hold enough prompts that every commercial theme has several phrasings, typically 80 to 300 per market. Below that range, a single answer moves the averages too much to read a trend. Above it, near-duplicates add runs without adding information, and the team spends its time tagging instead of reading results.
Size the set from the themes, and let the count follow. Loopwork sells to creative agencies in the United States and the United Kingdom. Its team lists 12 themes, from client approvals to pricing per seat, and writes eight phrasings for each: 96 prompts per market, 192 for both.
The phrasings matter more than the count. Five ways of asking about client approvals show how sensitive the answers are to wording, where one phrasing run five times only shows how much a single answer varies. Google's help page for AI Overviews tells users to ask several versions of a question[4]. Buyers rephrase, and the set should cover the phrasings they use.
Research points the same way. A 2026 study that followed 32 prompts on four AI engines for about 45 days found that results from one or two prompts mostly reflect those prompts, and it recommends a large, varied set. Its own set started from high-volume search keywords, expanded with the questions Google lists under People Also Ask[5]. A short list can be tested before the full set is written: the AI prompt tracker reports, for each prompt, how often your brand appears and which brands the answers name.
Repetition
Each prompt needs several runs, because 2 runs of the same prompt can name different brands
Prompt tracking runs each prompt several times because AI engines can give different answers to identical questions. Model providers say so in their documentation, and studies of AI search find answers and sources changing within minutes. Averaging several runs per prompt separates a real change in the answers from the noise of a single draw.
OpenAI's API documentation states that its models are non-deterministic by default, so the same request can return different outputs, and that fixing a seed makes them only mostly deterministic[6]. Anthropic's documentation says that even at a temperature of 0, results are not fully deterministic[7].
AI search adds its own movement, because the web searches behind each answer change too. In the 2025 study, the same queries run two months apart shared 18% of their pages for AI Overviews, against 45% for organic results, and short yes-or-no answers changed in up to 27% of cases within five minutes[3]. The 2026 study found only 34% to 42% of cited sources unchanged from one day to the next[5].
Treat each prompt as a sample. Run “best project management tool for creative agencies” five times on one engine and Loopwork may appear in three of the answers: a visibility of 60 on that prompt, where a single run would have shown 100 or 0. Anthropic gives the same advice for model evaluations: sample each answer several times and report the score with its standard error[8].
Method
A prompt set is built in 4 steps, from themes to repeated runs
A prompt set is built from the themes buyers ask about, written in several phrasings, tagged so results can be grouped, and run several times on each engine. The first three steps need a spreadsheet and some customer research. The fourth needs a way to run hundreds of prompts on several engines without retyping them.
-
Step 1
Choose the themes
List the topics buyers raise before they choose, using sales calls, support tickets, your top search queries and the questions Google shows under People Also Ask. Keep the themes that tie to revenue, such as pricing, integrations or a specific use case.
-
Step 2
Write several phrasings per theme
Write five to ten prompts per theme, the way different buyers would ask: short, long, with a constraint, as a comparison. Keep most of them unbranded, since branded prompts measure reputation more than discovery.
-
Step 3
Tag every prompt
Tag each prompt by theme, market and funnel stage so the results roll up into groups you can act on. Name the tags for the people who will read the report.
-
Step 4
Run each prompt several times
Run the set on each engine and country you follow, several times per prompt, and keep it unchanged between checks. The Prompt Tracking API takes the whole set in one request and runs it on the models you choose. See the Prompt Tracking API.