Prompt tracking is the repeated monitoring of how AI platforms answer a defined set of questions, recording brand mentions, source citations, and changes in those answers over time. If you have ever typed your category into ChatGPT, seen a competitor named, and wondered how often that happens, prompt tracking is the way to turn that one-off look into something you can measure and repeat. This guide explains what is being tracked, how a list of questions becomes a report, and where the results can and can't be trusted.
What is prompt tracking?
Prompt tracking means submitting the same questions to AI platforms on a schedule, saving each answer, and counting what shows up in them. A prompt is the question or instruction you send, such as "What should a small support team consider when choosing a help desk?" A prompt tracker is the tool that sends those questions, captures the responses, and turns what it finds into visibility reports.
What gets measured is the answer. The tracker looks at which brands the answer names, which pages it cites, how it describes the options, and how all of that changes the next time the question is asked.
A prompt run is one prompt sent to one AI model or experience, with the response saved. If you repeat the same prompt tomorrow, that is a second run, and it sits next to yesterday's answer instead of replacing it. Over time you end up with a pile of dated observations, and the pile is the real asset. The platform matters too, since Google AI Mode, Google AI Overviews, and ChatGPT Search are different experiences and can give different answers and links.
The tracker asks its own questions and watches the answers it gets back, so the results describe the prompts, platforms, and period it covered and nothing wider.
The signals a tracker looks for
Most of the reporting comes down to a few separate signals, and it's worth keeping them apart because each one answers a different question.
| Signal | What it means here | What it does not mean |
|---|---|---|
| Brand mention | The answer names your brand in its text | It may have no link |
| Owned-site citation | The answer cites a page on your own website | The answer may not name your brand in its prose |
| Third-party citation | The answer cites an outside page, such as a review | A page that talks about you is not your page |
| Recommendation | The answer actively suggests an option | Being named is not the same as being endorsed |
An answer can name your brand without linking to your site. It can also cite one of your educational pages without recommending your product at all, or cite a third-party article that happens to mention you.
The difference between linked and unlinked mentions is covered in AI brand mentions explained.
How does prompt tracking work?
Prompt tracking works as a loop of six steps: set the conditions, send the prompts, save the answers, pull out the signals, calculate rates, and repeat.
Set the conditions
The system first needs the prompt text and a list of AI platforms to ask. Depending on the product, other settings can apply too, such as language, location, how often to collect, which model to use, and whether web search is switched on.
Semrush documents daily monitoring across ChatGPT Search, Google AI Mode, and Gemini, and Ahrefs lets you check custom prompts daily, weekly, or monthly. No tracker offers every setting, and the point to hold onto is that the results belong to the conditions actually used.
Send the prompts on a schedule
At each scheduled time the system sends the prompts to the chosen platforms and captures what comes back. How it does that varies a lot. The IAB, the advertising industry body, describes several approaches: actively simulating queries, passively watching opted-in panels of real users, using data the platforms provide themselves, or a mix.
So it isn't safe to assume every prompt tracker works the same way.
There's a simple piece of arithmetic that helps with planning. The number of planned checks is prompts times platforms times collection occasions times repeats per occasion. Ten prompts on three platforms, checked weekly for four weeks with one run each, is 120 planned checks. Some of those can fail, and some platforms may not be covered for every prompt, so the number of answers you actually collect can end up lower.
Those 120 checks are measurements made by the tracker. They are not 120 people or 120 buyer conversations.
Save the answer and its context
Semrush provides snapshots of the AI results for tracked prompts, along with a report of cited domains and URLs. Gauge stores the full response, the brands mentioned, the sentiment, and the cited sources, and says it keeps a daily history of answers.
If you were designing your own record, the useful fields are the exact prompt, the platform and model, the date, the language and location settings, whether search was on, the full answer text, the brands found, the cited URLs, and the collection status. That's a recommended list and not a claim that every product shows all of them.
Pull out the signals
Next the system finds brand and product names in the answer and extracts the source references. Some tools also classify sentiment, how strongly a brand is recommended, or where it sits in a list.
The detection method matters. Gauge documents that it uses a language model to read each saved response and identify the brands in it, while the academic study discussed later in this guide used fixed-list text matching, which can miss synonyms and wrongly count a generic word as a brand. A dashboard number built on either method isn't automatically right, which is why being able to open the underlying answer is so useful.
Calculate rates and compare
The extracted observations then become counts, rates, platform breakdowns, cited-page reports, competitor comparisons, and trends. If a rate drops, you should be able to open the answers that changed rather than only seeing a lower score.
Repeat and compare
Later collections show whether a pattern holds, shifts, or swings from run to run. One answer can show that something happened once, but it can't tell you how often it happens.
What do prompt-tracking metrics tell you?
Prompt-tracking metrics tell you how often something appeared in the answers that were observed, which is a frequency inside a sample and not a ranking of your brand across all of AI search. The number only means something once you know what was counted and what it was divided by.
The calculations below are conventions that explain how the mechanics work. They aren't a standard that every vendor follows. Let N be the number of eligible answers that were successfully collected in the period you are looking at.
| Metric | A transparent way to calculate it | What it answers |
|---|---|---|
| Brand mention rate | Answers naming the brand at least once, divided by N, times 100 | In what share of the observed answers did the brand appear? |
| Owned-domain citation rate | Answers citing at least one owned page, divided by N, times 100 | In what share of answers was your website cited? |
| Mention-based share of voice | The brand's mention count, divided by all counted mentions for every brand in the comparison set, times 100 | How much of the selected brands' observed presence belonged to this brand? |
| Change in rate | Current rate minus previous rate | How many percentage points did it move? |
Count once per answer
For a mention rate, a brand named five times in one answer is still one positive answer. Ahrefs documents this directly, with an example where three mentions in the text count as one. Ahrefs also separates page-level citations from domain-level ones. An answer that cites two pages from your site has two page-level citation events but counts once as your domain being present. For that reason you shouldn't add up page citation rates and call the total your domain citation rate.
Decide the denominator
The denominator is the "N" you divide by, and it needs a stated policy. A few habits keep it honest. These are reporting recommendations, not facts about how every tracker is built.
- Keep collection failures apart from answers that came back fine but didn't mention you.
- Don't quietly drop answers just because they had no mention or no citation.
- Label a restricted denominator, such as search-enabled answers only.
A rate isn't fully readable without its period, sample size, platform coverage, counting unit, and denominator.
A worked example
Everything in this example is made up. It is not DeepSmith data, a customer result, or a benchmark.
Say you track five prompts on two platforms, collected on two dates, with one run each, and every run returns an eligible answer. That's 5 prompts times 2 platforms times 2 dates, or 20 answers. A mention means the brand appears at least once in the answer text, and an owned-domain citation means at least one page from the brand's site is cited.
| Observed sample | Eligible answers | Answers mentioning the brand | Answers citing the owned domain |
|---|---|---|---|
| Platform A | 10 | 7 | 2 |
| Platform B | 10 | 5 | 3 |
| Total | 20 | 12 | 5 |
Platform A's mention rate is 7 out of 10, or 70%. Platform B's is 5 out of 10, or 50%. Overall it's 12 out of 20, which is 60%, and the owned-domain citation rate is 5 out of 20, or 25%.
The fair way to say this is that the brand appeared in 60% of these 20 observed answers and its website was cited in 25% of them. It would be wrong to say that 60% of AI users saw the brand, because nothing in this example observes users at all.
A third-party review that discusses ExampleBrand needs its own reading, since the answer may mention the brand while citing someone else's website. For how to calculate and interpret the citation number, see how to measure citation rate.
Share of voice is not mention rate
Mention rate asks how often your brand appeared. Share of voice asks what portion of the presence among a chosen group of brands belonged to you. Several brands can show up in one answer, so their mention rates can add up to more than 100%. Share of voice divides by a chosen total instead.
Continuing the example, suppose your brand has 12 counted presences, Competitor A has 18, and Competitor B has 10. Your share of voice is 12 divided by 40, which is 30%. That's 30% of the counted brand presence among these three brands. It isn't an audience share or a market share. And if you swap in a different set of competitors, your share of voice changes even when your own answers stay exactly the same. Ahrefs warns about this in its documentation, noting that the result depends on the brand and competitor entities you have configured. A longer walk-through is in calculate share of voice.
Why does prompt tracking matter in an AI search program?
Prompt tracking matters because it gives you evidence to base decisions on, and it doesn't create visibility by itself. Its best use is linking a pattern in the answers to a next step you can defend.
It gives you a baseline
With it, you can say which platforms you checked, on which dates, how many eligible answers there were, and which calculation you used. "We're invisible in AI" is usually too broad for the evidence in hand, and a specific baseline is easier to act on. If you're deciding how many questions to include, how many prompts to track goes through that sizing question, and the research behind this guide doesn't support any single number as right for everyone.
It helps you see what kind of gap you have
Different patterns point to different questions.
| What you observe | What it tells you | A reasonable next look |
|---|---|---|
| Brand mentioned, website rarely cited | There's name presence but little owned-source presence in the sample | See which sources are being cited and whether a suitable page of yours exists |
| Your pages cited, brand not clearly recommended | Your site is showing up as a source | Read the recommendation language and don't assume preference |
| A competitor cited again and again for the same questions | That competitor has a recurring citation pattern in this sample | Open the exact cited pages and read the context |
| An answer gets your product wrong | There's an accuracy problem visible in the saved answer | Compare the claim with your current product information |
| A rate changes sharply | The measured observations changed | Check the raw answers, the sample size, and the collection conditions before naming a cause |
These are starting points for investigation. They don't prove why a platform picked a particular source.
It connects measurement to content decisions
The loop is simple. Observe the answers, inspect the evidence, decide what to improve, publish or update, and observe again. The decision might be to correct unclear product information, to fill a real information gap, or to look into why a competitor's page keeps being used as a source. If you want to see how a team turns tracked data into a plan, tracked prompts into a content plan covers that step.
Tracking points at an opportunity. It doesn't promise that one new article will win a citation.
Business results sit on a different layer
Prompt tracking measures visibility. Referral traffic, qualified leads, and revenue need their own analytics, and the IAB notes that measuring clicks after a citation depends on what data platforms make available, so it stays directional in many setups. More mentions don't automatically mean more revenue.
For a wider view of other measures, AI visibility metrics lays out the KPIs side by side, and Best AI visibility tools is our roundup if you're weighing software. Our hub page on prompt tracking tools focuses on prompt-level tracking.
What can prompt tracking not tell you?
Prompt tracking can't tell you what every buyer sees, how large the market is, or why a platform picked a source. It tells you what happened in the tracked sample under the conditions you set. These are the main places where it can mislead.
The sample is not the whole market
Semrush states that prompt tracking checks the prompts in the campaigns you configured, not every question buyers ask. The right framing is visibility within the tracked sample. Turning that into visibility across all AI search needs evidence about coverage and representativeness that most teams don't have.
One response is not a stable estimate
The IAB describes visibility for a query as a distribution and not a fixed value. The same question can produce different brands, citations, and framing on repeated runs. Repeating checks helps you see that spread.
A research paper called Don't Measure Once shows this in practice. It's an arXiv preprint dated April 8, 2026, and the material reviewed doesn't establish that it has been peer reviewed. One author discloses an affiliation with Aurora Intelligence, so it shouldn't be read as a neutral industry benchmark. The study tracked four engines (ChatGPT, Gemini, Google AI Mode, and Perplexity) across four industries under Swiss collection conditions, and found that average day-to-day overlap in the sets of cited sources was roughly 34 to 42%.
That number is a similarity measure between two sets of sources. It isn't a brand's citation rate or its chance of winning a citation. The study's limits travel with it: four industries, one country's collection conditions, and fixed-list brand matching. It supports a modest point, that repeated checks tell you more than a single screenshot, and it doesn't show that any particular number of checks is enough.
There's no defensible universal rule in the sources reviewed for how many prompts, repeats, or days are enough, so be wary of anyone offering a round number. The IAB suggests setting a baseline for normal variation before reading period-over-period changes. For more on day-to-day noise, see are AI visibility scores reliable.
A fresh session is a controlled condition
Gauge documents starting each prompt run without chat history, memory, or context from other prompts. OpenAI documents that ChatGPT search can use approximate location and, when memory is on, relevant saved memories when it rewrites a query. A fresh-session tracker gives you consistent conditions, but it doesn't reproduce a personalized, logged-in, multi-turn customer journey. It can't show exactly what every buyer sees.
One prompt can trigger several searches
OpenAI says ChatGPT Search typically rewrites a query into one or more targeted queries, and Google says AI Overviews and AI Mode may use query fan-out, issuing several related searches. So the question you submit isn't the only search behind the answer, as query fan-out explained describes.
Retrieved is not cited
OpenAI's web-search documentation separates the inline citations in an answer from a longer list of sources the model consulted, and it says the consulted list is often larger. Ahrefs likewise separates pages found in the background from pages actually cited. That leaves a few cautions. A page in a retrieved list isn't automatically a visible citation. A citation doesn't prove that the page was the only influence on the answer. And a citation doesn't mean the page was used to train the model.
The IAB's definition of a publisher citation rate is broader than the linked-citation rule in the worked example above, which is one more reason to define your measurement and not assume the word "citation" means one thing.
Before and after is not proof of cause
Gauge cautions that a change after an edit shows correlation and doesn't prove the edit caused it, since models and retrieval settings can shift in the same window.
Tracking does not guarantee inclusion
Google states that a page must be indexed and eligible to appear in Search with a snippet to be a supporting link in AI Overviews or AI Mode, and that meeting the requirements doesn't guarantee serving, as its AI features documentation explains. A tracker doesn't change that.
Google now reports some of this itself
Visibility data isn't only available through third-party trackers. As of November 12, 2026, Google's Search Console documentation describes a dedicated Generative AI performance report for Search, with impressions for AI Overviews and AI Mode, which Google says rolled out to all websites worldwide by August 31, 2026.
That impression measure counts how many times links to your site were shown to a user in a generative AI feature. It's a platform-recorded count of real exposure, which is a different thing from the share of answers a tracker returned during its scheduled checks. The report has limits, such as needing enough impressions to appear and preliminary data for the newest days, and it describes impressions and site performance, not a log of every conversational prompt or sales attribution.
Use platform reporting where it exists, and keep in mind that controlled prompt tracking answers a different question.
What does a useful result look like?
A useful result is one you can read and connect to a decision. "Our AI visibility is 80" is weak. A stronger statement names the platforms, the period, the number of eligible answers, the counting rule, and the denominator, and it points to the underlying answers. Don't turn the 60% from the made-up example into a target, since nothing in the sources says it's good, bad, or typical.
If you want a way to start, pick one answer pattern from your most recent collection, connect it to one content decision, and measure again at the next reporting cycle.
Related guides
- How to Organize Tracked Prompts by Topic, Funnel Stage and Persona: Structure a prompt list into groups by topic, funnel stage and persona so reports answer real questions.
- How to Read Prompt-Level Win and Loss Data Against Competitors: Read per-prompt results to see where competitors are named instead of you and which losses are worth fixing first.
- How to Track the Same Prompts Across ChatGPT, Gemini and Perplexity: Run one prompt set across three engines and compare results fairly, adjusting for each engine's quirks.
- Why Prompt Tracking Data Goes Unused and How to Fix It: Find out why prompt data sits in dashboards and fix it with owners, a review cadence and a link to content tasks.
- Self-Serve vs Managed AEO: Which Fits Your Team?: Compare running AEO yourself with software against hiring a managed service on control, speed, in-house skill needed and workload, then pick the model that fits your team's capacity.
