uterank.
AI VISIBILITY

An LLM visibility tracker built around prompt-level evidence

An LLM visibility tracker records whether specific prompts return specific brands as named mentions or source citations, and plots that result over time per prompt per engine. The useful ones expose the full generated answer for every run so a human can read the reasoning, not just the aggregate percentage.

Live product dataNo credit cardSetup in minutes
Outerank
Live data
AI visibility pulse
Citation coverage68%
+18.4%
ChatGPTTracked
PerplexityGrowing
GeminiVerified
3 opportunities ready to act on
FREE CITATION CHECK · ONE REAL AI RUNSee whether ChatGPT actually cites you — one real prompt, one engine, no card.Create a free account and we'll run one live citation preview on the engine of your choice. No credit card.
Start the free citation check
OR SCAN YOUR SITE FIRST — NO SIGNUP
Gratuit · sans inscription · résultats en secondes
Les reçus · en direct, datés, de première main
THE HOOK

Why a single visibility percentage lies to you

The pitch for most LLM visibility trackers is a dashboard with one number — a “visibility score” between zero and a hundred. The number goes up and down. It is a convenient thing to put on a slide. It is also almost always wrong, because it compresses prompt-by-prompt behaviour into a single average, and the average hides which prompts moved.

What a real tracker needs to show is: for each tracked prompt, on each engine, was your brand mentioned, was a URL on your domain cited, and what did the answer actually say? That is a chart per prompt, not a chart for the whole account. The composite score sits on top of that as a convenience, not as the signal.

This page describes how we built that in Outerank — specifically for the person who runs an SEO or marketing function alone and does not have time to track thirty prompts across three engines in a spreadsheet.

01

What a prompt-level tracker actually captures per run

For every tracked prompt, every engine, every run, the tracker records seven things: the exact prompt string, the full engine answer, whether your brand name appeared in the answer prose, whether a URL on your domain appeared as a cited source, the full list of cited source URLs with their domains, the detected sentiment of any brand mention, and the timestamp.

The “mentioned but not cited” distinction is where most dashboards fail. A brand can be named inside a comparison paragraph without the engine attaching a source URL to it. That is a mention; it is not a citation; it does not necessarily send traffic. A dashboard that treats both as a single number conflates the two most important states in the data.

The tracker should expose both as separate time series. Mention rate is a reach signal; citation rate is a traffic signal. The two usually correlate but not always, and the divergence is diagnostic.

02

Why historical delta per prompt is the chart that matters

A single run is a snapshot in a very noisy medium. The engines genuinely change their answers between calls — temperature, retrieval variance, context window differences. One good run is not evidence; one bad run is not evidence. The signal lives in the trend over at least four runs.

Prompt-level delta charts turn the noise into actionable signal. If prompt A shows your brand cited six runs in a row and then drops for three consecutive runs, that is a change worth investigating. If prompt B oscillates between cited and uncited every run, that is the ambient noise floor, and a tool that reports it as alarming is wasting your attention.

Outerank exposes a per-prompt chart with four buckets: stable positive, stable negative, improving, and degrading. Stable positive is background health; stable negative is the current content gap list; improving validates recent fixes; degrading is the alert queue.

03

What a “good” tracked prompt looks like

The most common first-run mistake is tracking generic head terms — “SEO tool”, “ecommerce platform”. These produce bland answers that mention everyone, and the signal is almost zero. A head term is a keyword, not a prompt. Buyers do not type head terms into ChatGPT; they type questions.

A good tracked prompt is specific, buyer-shaped, and realistic. “Best AI visibility tool under $100/month” is a good prompt. “What SEO tool is cheapest for solo founders tracking three competitors” is a good prompt. “ChatGPT vs Perplexity for citation research” is a good prompt. Each of these produces answers specific enough that your presence or absence is a signal.

The second mistake is tracking only your own brand. Half the value is tracking the prompts that are category questions where your brand does not get mentioned yet. A prompt is an opportunity surface; the ones you are already winning tell you less than the ones you are losing.

04

How much it costs to run a prompt-level tracker honestly

Live LLM calls cost real money per request. A prompt run against three engines costs roughly nine credits in our system (three credits per engine per prompt), and on a weekly cadence that is thirty-six credits per month per prompt. Twenty-five tracked prompts on a weekly cadence is 900 credits per month, which is why the Pro plan’s 500-credit allowance maps to roughly bi-weekly cadence on a 25-prompt set.

Vendors who advertise “unlimited” tracking are hiding the meter. They either rate-limit internally once you push past expected usage, run a cached answer for repeat prompts (which defeats the point of tracking), or absorb the margin into a $400+/mo plan where the budget is wide enough to cover the real cost.

The honest cadence at $89/mo is bi-weekly. The honest cadence at $159/mo (Studio) is weekly to daily depending on prompt count. Any price point below $89/mo on real-engine tracking is either loss-leading or lying about the engine call.

UNIQUE DATA · FROM OUR OWN INDEX

Prompt-level deltas on fifty tracked brands — anonymised

The Index runs the same thirty-prompt set across fifty brands on a bi-weekly schedule. Here is the distribution of per-prompt delta states on the last run, aggregated across all fifty brands and three engines (ChatGPT, Perplexity, Gemini).

Delta stateShare of tracked promptsWhat it means
Stable positive34%Brand cited consistently for the prompt. Background health.
Stable negative47%Brand not cited for the prompt across runs. The opportunity queue.
Improving9%Rising from negative to positive over the last four runs.
Degrading4%Falling from positive to negative. The alert queue.
Oscillating / noise6%Flips between runs with no trend. Reduce weight; do not act on it.

The useful observation here is that the opportunity queue (47%) is nearly five times larger than the alert queue (4%). Most teams spend their attention on the small set of degrading prompts; the actual leverage is in the much larger set of stable-negative prompts the brand has never been cited for and could be if the content existed. See how this fits into the broader AI visibility tool, and the paid-ads context on AI search visibility.

The anonymised distribution above comes from the public AI Visibility Index — fifty brands tracked bi-weekly across three engines, with the per-brand breakdown open for inspection.

WHAT THIS DOES NOT DO

Honest limitations

  • No reasoning trace. The engines return the answer; they do not return the retrieval chain. We cannot tell you which specific documents the model considered and discarded.
  • Three engines, not five. ChatGPT, Perplexity, Gemini. Copilot and Grok are not supported today.
  • Prompt wording matters. “Best X for Y” and “top X for Y” can produce different answers on the same engine. The tracker runs the exact string you give it; small rewrites are a separate run.
  • Regional variance. Answers can differ by the IP region the call is made from. We call from US-East by default; EU/APAC regions are a Studio feature.
  • Credits are a real resource. Twenty-five prompts × three engines × four runs/month = 300 credits. Pro’s 500-credit allowance absorbs that comfortably; weekly cadence (600 credits) nudges you to Studio.

If any of these are deal-breakers, we would rather tell you now than at the end of a billing cycle.

FREQUENTLY ASKED

Questions worth asking before you start

What is an LLM visibility tracker?

A tool that records whether specific prompts return your brand as a mention or source citation, across ChatGPT, Perplexity and Gemini, and plots that result over time per prompt per engine. It differs from a rank tracker because it samples the generated answer, not a results page.

How many prompts should I track?

Twelve to twenty-five is the useful range. Fewer than twelve and the composite score moves on noise; more than thirty costs enough credits to push most teams to Studio without a clear return. Pro includes twenty-five.

How often does the tracker refresh?

Bi-weekly on Pro, weekly or daily on Studio. The engines’ answers do shift between runs; the signal needs at least four runs to be readable, which is why bi-weekly is a reasonable floor.

Does the tracker call Claude or Copilot?

No, not today. ChatGPT, Perplexity and Gemini only. Those three cover the bulk of consumer AI referral traffic in US English.

Can I export the per-prompt answers?

Yes. CSV export of every run’s full answer, cited URLs, and mention state is included in Pro. PDF reports with the chart deltas are a Studio feature intended for client-facing use.

What’s the smallest useful tracked set to start with?

Six prompts: three brand prompts (“is X trustworthy”, “how much does X cost”, “X vs Y”) and three category prompts (where you want to be cited and are not sure you are). That covers the two main uses — brand monitoring and opportunity discovery — with the smallest credit footprint.

YOUR NEXT MOVE

Ready to track AI citations continuously?

Outerank Pro runs your prompts across every AI engine on a schedule, stores every answer, and alerts you the moment a citation disappears.