Prompt tracking (also called prompt monitoring) is the regular checking of a fixed list of questions asked to ChatGPT, Gemini, Perplexity and Google's AI features, to see whether your brand is named, whether your site is cited as a source, and which competitors appear next to you. In our test on 27 October 2026, the same request, "who can install solar panels on our factory roof", written three different ways made ChatGPT name 20 different installers, with not a single name shared between the three answers.
So the right question is not "which tool should I use?". It is which prompts go into your sample.
On Google this problem never existed. Take Sunscape, which develops solar plants across Italy: from March 2026, in six months, organic impressions went from 329 to 5,809 a month and clicks from 46 to 195, as we explain in how long SEO takes.
In AI answers you have to build the list of questions yourself. And the sector of our test is Sunscape's own.
The mistakes we see most often when a company sets up tracking:
- copying two-word SEO keywords and calling them prompts;
- writing one wording per question and treating it as "the" answer;
- filling the sample with questions that contain your brand name, where showing up is a given.
In this guide:
- the test on 12 prompts (4 questions × 3 wordings);
- which prompts to track and how to find them in Search Console;
- how many prompts you need, how often to run them and what to log in the history.
What is prompt tracking for AI visibility? (and why it is not rank tracking)
Prompt tracking is the AI-answer version of rank tracking: instead of a keyword and a position from 1 to 100, you follow a question and record who appears in the answer, in what order and with which sources.
Three differences from rank tracking:
- A prompt has no search volume. "Prompt tracking" gets 40 Google searches a month in the UK and 10 in Italy (Google Ads via DataForSEO, 27 October 2026), and no tool knows how often a question is asked to ChatGPT.
- The answer changes even when nothing changes. The SparkToro study with Gumshoe (January 2026, 2,961 answers, US market) found less than a 1 in 100 chance that ChatGPT or Google's AI repeats the same list of brands, and less than 1 in 1,000 that it repeats them in the same order.
- The answer depends on who asks. OpenAI's help centre confirms that ChatGPT may estimate your location from your IP address and use saved memories to rewrite the search.
In practice: one prompt on its own measures nothing. A set of prompts, repeated over time and read as an average, does.
The metrics (mentions, citations, share of voice, sentiment) are calculated afterwards, on that data: see AI visibility metrics.
The test: the same question in three wordings (12 prompts on commercial solar in Italy)
Wording decides who gets named: of four questions, only the one about panel brands produced three answers with the same core names.
Method. On 27 October 2026 we asked ChatGPT 12 prompts in Italian, with web search on and location set to Italy, through DataForSEO: 4 intents of a business buyer, each written three ways (quoted here in translation).
- short, like a keyword: "fotovoltaico per aziende" (solar for businesses);
- question: "which companies install solar systems on industrial buildings in Italy?";
- question with context (role, sector, place): "I run an engineering firm near Bergamo with a 3,000 m² factory: who can quote for a rooftop solar system?".
The other intents: solar with no upfront investment (PPA or leasing), panel brands, and system cost. One run per prompt, so a snapshot; we counted brands named in the text (not in images) and the domains cited.
| Intent | Brands named (short / question / with context) | Brands in all three | Sources cited in all three |
|---|---|---|---|
| Installers for industrial roofs | 5 / 10 / 5 | None | None |
| Solar with no upfront investment (PPA, leasing) | 6 / 16 / 5 | Edison Next, SENEC | senec.com, edisonnext.it |
| Panel brands | 5 / 7 / 5 | JinkoSolar, Trina Solar, LONGi | None |
| System cost | 0 / 0 / 0 | No brand named | None |
1. On "local" questions, each wording is a different market
The three answers about installers named 20 different companies, none of them twice. Short: installers around Milan. Question: ten companies spread across Italy. With context: five firms near Bergamo, with phone numbers.
So which of the three is "right"?
All three: they are three different customers. Track one and you measure one customer.
2. Strong brands hold their ground
On panel brands, JinkoSolar, Trina Solar and LONGi appear in all three answers, whatever the wording. SparkToro saw the same in the US with 142 headphone questions written by different people: Bose, Sony, Sennheiser and Apple stayed in 55-77% of answers.
3. Price questions name nobody, but cite sources
The three answers about cost did not name a single supplier. They did cite 11 different domains, including installers' sites such as tgreen.it, darmaenergia.com and leris.it. And for a 100 kW system they gave different estimates: €70,000-120,000 in the short version (and €90,000-140,000 further down the same answer), €75,000-100,000 in the one with context.
These questions do not measure mentions: they show whose pages the AI takes its numbers from.
4. Location creeps in even when you leave it out
Of 8 prompts with no place in them, 4 answers localised themselves: one to Garbagnate Milanese, two to Campania, one to Piedmont (that last one while a prompt naming Piedmont was running in parallel, so we treat it with caution). And all 12 answers ended by asking for region, consumption or roof type.
For anyone tracking: the tool's location is a variable, to be fixed and written into the history.
One wording shows you one market.
You have just seen three versions of the same question produce 20 different installers. If your report follows one of them, you can look absent exactly where your customers are asking. We build the sample from your customers' questions and from Search Console, and run it every month under the same rules.
Book a strategy callWhich prompts to track (four types and the three-wording rule)
Track the questions a customer would ask before contacting you, without your brand name, grouped by intent and written in three versions: short, question, and question with the customer's context. Prompts with your brand name are useful, but they are a minority.
The four types we use in B2B:
- Supplier: "who can do X for us", "which companies do Y in the North West";
- Comparison and alternatives: "X or Y for a 50-person company?", "alternatives to [competitor]";
- Problem: the question asked before the buyer knows the solution exists;
- Cost and decision: "how much does it cost", "what is the payback". Here, as in the test, the cited source matters more than the name.
For the brand split, the most concrete ratio we found is in HubSpot's guide (October 2026): 50-60% unbranded prompts, 20-30% competitor prompts, 10-20% with your own name.
The three-wording rule comes from the test: one question written three ways gives you three samples of the same need, to be read together. If you appear in one out of three, your presence depends on how the customer asks.
How to find prompts in Search Console (and which to throw out)
Long, question-style queries are prompts your customers have already written. On our own site, in the 90 days to 27 October 2026, the query-page rows containing a question mark and landing on Italian pages number 129, all with zero clicks.
One ready-made example (in Italian): "software per misurare la visibilità su chatgpt?" (software to measure visibility on ChatGPT), 13 impressions at an average position of 5.8.
The quickest filter, in the Performance report: a query regex such as \?|^(how|what|which|who|why|best).
So they are all good?
No. Some were not written by a person. In our export, 4 questions end with "my location is italy" or "my location is lugano district, ticino, switzerland": the suffix some tools add to a prompt. And our article on query fan-out receives 15 variants of the same long sentence, identical except for the name of the tool being compared (13 different names), for a total of 132 impressions and zero clicks.
Our reading: prompts run in batches, probably by monitoring tools. Keep them out of your sample.
How many prompts should you track? (25 to 50, then it multiplies)
For a company in one market, 25-50 prompts are enough to start: that is HubSpot's range and close to the one we use in our AI visibility guide (30-50). The number that matters, though, is prompts × engines × locations.
An example:
- 10 intents × 3 wordings = 30 prompts;
- 30 prompts × 4 engines (ChatGPT, Gemini, Perplexity, AI Overviews) = 120 answers per run;
- with 3 sites measured separately, 360 answers.
According to its documentation, Semrush includes 25 prompts in the AI Visibility Toolkit; SISTRIX, in its prompt monitoring changelog, goes from 100 to 1,500 prompts depending on the plan (pages checked on 27 October 2026). How to compare them on everything else is in how to choose an AI visibility tool.
With several sites, "supplier" questions need repeating town by town: we cover this in local AI visibility.
How often should you track brand mentions in AI? (less than you think)
For most B2B companies one run a month is enough, with several wordings per question; weekly or daily runs are worth it during a launch, a campaign or a reputation issue.
The reason is in SparkToro's data: if two answers to the same question match less than once in 100, a daily run mostly records noise. The study recommends running each prompt 60-100 times and averaging: for 30 prompts on 4 engines that is at least 7,200 answers per run, out of reach for a small business.
Tools take one of two routes:
- run every day and show you the average (Semrush, SISTRIX); HubSpot itself advises collecting daily but reading the data monthly;
- run once a month on more prompts: that is the choice of ScryRank, the monitoring tool developed by Visilay, which covers ChatGPT, Claude, Gemini, Perplexity and AI Overviews with plans of 50, 100 and 200 prompts a month (€39, €69 and €139 a month, price list checked on 27 October 2026).
A monthly run has a limit we will state plainly: if you change a page today, you will not see the effect next week. How stable the data is between runs is what we measure in how accurate AI visibility data is.
A maintenance rule, again from HubSpot: change no more than 20% of the sample each quarter.
Why you need a history for each prompt (and what to log every time)
A per-prompt history matters because the average hides the cause: if your mention rate drops from 40% to 30%, only the prompt-by-prompt detail tells you which questions you lost.
In the test, the three cost questions never named a supplier: put them in the same average as the "supplier" questions and they drag everyone's mention rate down without telling you anything about yourself. They belong in a separate column.
What we log for every answer:
| Field | Example | Why it matters |
|---|---|---|
| Date and time | 27/10/2026 16:05 | Comparing months |
| Engine and mode | ChatGPT with web search | Sources change with and without web search |
| Location set | Italy, no town | In the test 4 answers out of 8 localised themselves |
| Intent and wording | Installers, short version | Reading the three versions together |
| Exact prompt | "fotovoltaico per aziende" | One changed word is another prompt |
| Brands named, in order | 5 names, yours 3rd | Mention rate and position |
| Sources cited | 4 domains | Citation rate and pages to work on |
| Full answer text | Complete copy | Rechecking sentiment and quoted figures |
Which tools provide historical AI visibility data at the level of a single prompt? SISTRIX states it keeps mention history for each prompt from the moment the project is created; ScryRank includes the history of its runs on every plan. The full comparison on this point is in our comparison of 8 AI visibility tools. Before you buy, always ask whether the tool stores the full answer text: without it, the number cannot be checked.
The history is what feeds the monthly AI visibility report.
What tracking tells you about your chances of being cited (sources beat names)
Prompt tracking does not increase citations, but it tells you where to work: where you appear, where you do not, and which pages the AI reads instead of yours. In the test, the cost answers took their numbers from installers' sites and from public bodies such as Invitalia and ARERA; the panel answers from Kiwa PVEL, Altroconsumo and pv magazine.
That is where citability is decided. AI answers cite pages that answer the exact question with a verifiable number. For Google it is written in the documentation on AI features: AI Overviews and AI Mode may run several related searches (query fan-out) to build the answer, and the pages shown vary. No tracking guarantees a citation.
Working out which page the AI uses instead of yours, and how to take its place, is what we do in our GEO service, and what we describe for brands in AI brand mentions. If you would rather hand the tracking over, in our AI monitoring service we choose prompts, engines and competitors and give you the monthly reading.
Frequently asked questions
Yes. They are two names for the same job: a fixed list of questions run at regular intervals on ChatGPT, Gemini, Perplexity and Google's AI features. Only the label changes from tool to tool.
Yes, but only a few: 10-20% of the sample, following the ratio suggested by HubSpot. On questions with your name in them, appearing is almost a given; your visibility is measured by the unbranded questions.
No. Search Console only shows searches made on Google, including AI Overviews and AI Mode. Long, question-style queries are still a good starting point for writing the prompts to track.
Summary
Prompt tracking follows questions, not keywords.
One question in three wordings: 20 installers, none in common.
Strong brands survive rephrasing, local suppliers do not.
25-50 prompts, mostly unbranded, some taken from Search Console.
One run a month, location fixed, a history for every prompt.
One last thing. All 12 answers ended by asking for more context. That means the real conversation carries on after the first prompt, and the final list of suppliers arrives at the second or third message, which no tool measures yet. The closest you can get is the third wording: write it the way your best customer would.
