Tracking LLM visibility means running the same set of buyer prompts on each AI platform on a fixed schedule and recording, for every run, whether the brand was mentioned, whether its pages were cited, how it was described, and which competitors appeared. That locked prompt set is what lets a team measure AI search visibility as a trend over weeks and months.
Two KPIs usually sit on top of it. Brand mention rate is the share of prompt runs in which an answer names the brand. AI share of voice is the brand's share of all brand mentions across the same prompts, competitors included. Both need the prompt set underneath before they mean anything.
What should an LLM visibility tracker record?
An LLM visibility tracker should record six fields for every prompt run: mention, citation, cited source, sentiment, competitor presence, and prompt coverage. Teams that track LLM mentions without the other five fields can see that a number moved but not why.
| Field | What it records |
|---|---|
| Mention | The brand name appears anywhere in the answer, including a wrong or incomplete description |
| Citation | The answer links to a page on the brand's domain |
| Cited source | The exact URLs and domains the answer links to, the brand's or anyone else's |
| Sentiment | The description of the brand reads positive, neutral, negative, or inaccurate |
| Competitor presence | Competing brands named or cited in the same answer |
| Prompt coverage | The share of prompts in the set where the brand appears at all |
A mention and a citation are separate events. An answer can recommend a brand by name without linking to it, and it can link to a brand's guide while recommending a competitor. AI citation tracking therefore needs its own column. The differences between AI citations and mentions also shape how a brand earns each one.
Log cited sources even when they belong to someone else. A third-party review page that keeps appearing in category answers shows where the answer engine gets its information, and it can explain a competitor's lead.
How does a KPI become a prompt set?
A KPI becomes a prompt set by writing the outcome in one plain sentence and listing the questions a real buyer would type to get an answer that reveals it. For "brand mention rate in AI answers for the category," that means the ways a buyer phrases the need: problem-first questions, comparisons, "best X for Y," and questions tied to a use case.
The translation follows a repeatable sequence:
- Write the KPI in one sentence, for example "increase the share of relevant AI answers that cite our documentation over a named competitor's."
- Draft the questions a buyer would ask that would surface that outcome, in the buyer's own phrasing.
- Sort the questions into a few intent buckets.
- Cut duplicates, and cut prompts so broad that every brand in the category gets mentioned.
- Lock the list before the first run, because changing prompts mid-series breaks the comparison between weeks.
Every prompt should trace back to a question a real buyer asked. That traceability is what answers a client or manager who wants to know why a number moved.
Take a company selling API monitoring software with the KPI "increase brand mention rate on outage-related prompts." Its prompts read like "why did our API go down and how do I find out faster next time," "best way to get alerted before customers notice an outage," and "X vs Y for uptime monitoring."
How should the prompt set be segmented?
An LLM visibility prompt set should be segmented by funnel stage, intent type, and topic, so one share-of-voice figure can be broken down into where a brand gains and where it loses. A brand can lead on awareness prompts while losing every comparison against one competitor, and a blended average hides that split.
Funnel stage separates awareness prompts ("what is X") from consideration ("best X for Y") and decision prompts ("X vs Y"). Intent type splits informational questions from comparisons and transactional ones, and topic groups prompts around the categories the business competes in. Decision and comparison prompts usually say the most about competitive standing, because they are where a buyer weighs one brand against another. Weight them accordingly: a comparison against the closest competitor matters more than a generic category question.
Topic segmentation ties the plan back to content work. Closing a gap the plan reveals is a generative engine optimization job: structuring pages so each model's retrieval layer can find and trust them. Topic buckets show which content cluster a drop belongs to.
How many prompts and runs does a tracking plan need?
An LLM visibility tracking plan for one well-defined category commonly starts with 20 to 30 prompts and grows as segments are added. That range is a starting point. The right size is whatever covers the phrasing variety buyers use while staying small enough that someone maintains it every cycle.
A single-product company in one category can run a tight set. A platform that competes in several categories needs a locked set for each, tracked on its own, so signals from unrelated markets never blend.
Repeat runs matter as much as prompt count. The same prompt can return a different answer on the next run. Running each prompt more than once per cycle and reporting the share of runs where the brand appeared gives a rate that can be compared week to week.
Who should own the prompt set?
The team that owns the KPI should own the prompt set behind it, because whoever writes the prompts decides what the number measures. An agency should review the list with the client, since the client's team knows how its buyers phrase questions. In-house teams get better phrasing from sales and support call notes than marketing would invent alone.
Ownership also decides who explains a flat quarter. A prompt set built once and never reviewed can produce a flat line as easily as a real lack of visibility, so name one owner for the prompts, separate from whoever owns the content fixes.
Why does each AI platform need its own prompt set?
Each AI platform needs its own prompt set because the platforms retrieve and cite sources differently, so results from one do not stand in for another. Search Atlas research on 5,504,399 responses across 748,425 queries compared Perplexity Sonar, which had web search on, with Gemini and GPT models, which ran with web search off. Perplexity showed the highest citation density and the broadest domain diversity, while Gemini wrote long answers with fewer citations. Overlap between the models was limited but meaningful: about 60 to 65% of shared queries included at least one common cited domain (A Comparative Evaluation of LLM Responses).
Traditional rankings are no shortcut either. A Search Atlas study of 18,377 matched query pairs found a median 25 to 30% domain overlap between Perplexity answers and search results, and below 15% for GPT and Gemini (How GPT Results Differ from Search Engine Results).
How should Google AI Mode be checked?
Checking Google AI Mode starts with confirming that each tracked prompt returns an AI Mode answer. A list copied from keyword research may hold short head terms that do not read like questions, so test the phrasing before reading a zero as a visibility problem. Once the prompts return answers, check whether the brand's pages are among the linked sources.
Google's AI features guidance says traffic from AI Mode and AI Overviews is included in Search Console's Web search data. Search Console reports clicks and impressions for the page. What an answer said about the brand comes from the prompt set.
How does Gemini change the citation count?
Gemini produced fewer citations than Perplexity in the Search Atlas comparison, though Gemini ran there with web search off. When a Gemini answer lists no sources, no citation can be counted for that run. Record those runs apart from runs that list sources, so a low citation rate is not read as a brand being absent from the answer.
Which pages does ChatGPT tend to cite?
In Search Atlas research on 90,000 citations (URL Freshness in LLM-Generated Answers), OpenAI models cited the widest range of page ages, including the largest share of older sources, when web search was on. With search off, they still cited moderately recent material. OpenAI models power ChatGPT, so long-lived reference pages may keep earning ChatGPT citations. Track which of the brand's older URLs appear alongside new ones.
Why does a brand go missing on Perplexity?
Perplexity searches the web on every prompt, so a brand that goes missing there most likely lost the citation at the retrieval step, possibly to a competitor's page. The same freshness study found a strong recency bias on every platform with web search on, with Gemini citing the freshest material and Perplexity in the middle. When a Perplexity citation is lost, compare the publish and update dates of the page that replaced it.
What goes into each run's record?
Each run's record is one row in a tracking log, with enough detail that another person could repeat the run and get a comparable result. Overwriting last week's answer with this week's destroys the time series. The example row below uses the API monitoring company from earlier.
| Column | What to enter | Example row |
|---|---|---|
| Prompt ID and version | A fixed ID for the prompt and a version number that changes only at the quarterly audit | P-07, v1 ("best way to get alerted before customers notice an outage") |
| AI platform | The platform the prompt ran on | Perplexity |
| Run date and time | When the prompt ran | 2026-10-12, 09:30 |
| Model or mode | The model or mode name, if the interface shows one | Default mode |
| Location or account | The location and logged-in state used | United States, logged out |
| Mention | Whether the answer names the brand | Yes |
| Cited URLs | Every URL the answer links to | The brand's alerting guide, two competitor pages |
| Sentiment | How the answer describes the brand | Neutral |
| Competitors | Competing brands named or cited | Competitor A (cited), Competitor B (named) |
| Notes | Anything inaccurate or worth a follow-up | Describes a feature the brand retired |
The model or mode name matters because AI platforms update without notice. A jump that lines up with a platform update means something different from one that lines up with a content publish date. A spreadsheet works for a small set. Larger sets usually move to a tracking platform, and the guide on how to evaluate AI visibility tracking platforms compares the options.
How often should a tracking plan be re-run?
A locked prompt set should be re-run weekly in competitive or fast-moving categories and monthly elsewhere, using the same prompts every time. Irregular runs break the trend line, because a long gap can hide a model update or a competitor launch that would otherwise explain a shift.
Once a quarter, audit the prompt list itself. Buyer language changes and new competitors enter comparison prompts. Check that each prompt is still one a buyer would type today, and add new prompts as separate, dated additions so the historical set stays comparable.
When does a change in mention or citation rate count?
A change in mention or citation rate counts only after the team has checked whether the model, the content, or a competitor changed first. A jump the week after a platform update is more likely the update's doing than anything the content team shipped that week.
Before the first report, the team sets its rules. The baseline comes from several runs, so normal answer-to-answer variation is already visible, and every rate is reported with the prompts and runs behind it, because a move from 2 of 10 to 3 of 10 is one answer. The team also decides in advance how large a move, held across how many runs, counts as a change worth acting on. There is no universal threshold, so the team sets one and keeps it for the quarter.
A number that moves in one segment tells a team more than a blended score. If share of voice rises on informational prompts and stays flat on comparison prompts, awareness improved and the competitive comparison did not, so the next fix belongs on the comparison content.
Does LLM visibility show AI traffic?
LLM visibility does not show AI traffic, because it measures exposure inside the answer and a reader can act on an answer without clicking. Referral visits and conversions belong to a separate measurement: tracking LLM traffic in analytics. Report the two side by side and keep each to what it measures.
What belongs in an LLM visibility report?
An LLM visibility report should show each platform on its own line, with the rate, the prompt and run counts behind it, and the change against the baseline. A combined score that averages platforms hides which one has the gap, and a fix aimed at the wrong platform may not move it.
A useful row reads: platform, prompt bucket, mention rate, citation rate, runs, baseline, change, and a note on what changed in the period. Agencies need the same rows plus a client-facing explanation of uncertainty and cadence, which the guide to AI search visibility reporting for agencies lays out.
How does a measured gap become a shipped fix?
A measured gap becomes a shipped fix through a loop: the locked prompt set surfaces a gap on one platform and segment, a fix ships against that gap, and the identical prompts run again. The fix is finished only when that re-run shows the number moved.
The fix might be a rewritten comparison page or a correction to a page that answers keep citing with wrong details. Skipping the re-run is where the loop most often breaks. Three months later nobody can say whether the fix worked, because there is no before-and-after number tied to the same prompts.
LLM visibility tracking FAQ
What is the difference between an LLM mention and an LLM citation?
An LLM mention is the brand name appearing in an AI answer, while an LLM citation is a link from the answer to a page on the brand's domain. A brand can be mentioned without being cited, and a page can be cited in an answer that recommends a competitor, so a tracking plan records the two as separate fields.
How many prompts does an LLM visibility tracking plan need?
An LLM visibility tracking plan for one well-defined category commonly starts with 20 to 30 prompts and grows once funnel stage, intent, and competitor comparisons are tracked separately. A business that competes in several categories needs a locked prompt set for each one.
Does a high LLM mention rate mean AI platforms send traffic?
A high LLM mention rate does not show that AI platforms send traffic. Mentions and citations measure exposure inside the answer, and a reader can act on an answer without clicking. Referral visits and conversions come from the site's own analytics.
How often should LLM visibility be re-checked?
LLM visibility should be re-checked weekly in competitive or fast-moving categories and monthly elsewhere, using the same locked prompts on every run. The prompt list itself should be audited once a quarter.
What the free AI Visibility snapshot and LLM Visibility show
The free AI Visibility snapshot gives a first baseline: 10 prompts across ChatGPT and Gemini, showing where AI mentions the brand and which competitors appear in its place. From there, Search Atlas LLM Visibility keeps tracking those queries across platforms and shows the trend for each one, so a drop on one platform is visible next to the baseline it moved from.
