Flag a false citation
If the extractor claims a mention that never happened, or misses one that did, flagging it from the prompt matrix excludes it from the score and atomically recomputes the affected weekly snapshot.
The instrument
Stop losing deals you never knew about
One clear read on where you stand with every major AI assistant, how warmly you're recommended, and exactly what to fix — refreshed every week, on autopilot.
- 6 AI assistants
- 60 checks a week
- One score, every Monday
- Nothing hidden
Short answer
What does the report actually tell me?
What is AI visibility tracking?
It is the answer to the question keeping you up at night: when a buyer asks AI what to buy, does it name you — or your competitor? You get three plain answers: were you mentioned, how strongly were you recommended, and which websites did the AI rely on to decide?
Why not just count mentions?
Because being named fourth as a fallback looks identical to being named first as the top pick if all you count is mentions — but only one of them wins the deal. That is why every result is weighted by both how high you appeared and how favourably you were described.
How often do you check?
Once a week, every week. Each round produces sixty checks — ten buyer questions across six assistants — and they are all scored automatically. It is a weekly check rather than a live feed, which is what makes week-to-week comparison meaningful.
Six assistants, six opinions
They do not agree with each other
Each assistant reads different sources and answers in its own way. You can be the top pick in one and completely absent in another on the same afternoon — which is why checking a single assistant tells you almost nothing.
- ChatGPTOpenAI · conversational shortlist answers
- ClaudeAnthropic · long-form reasoned recommendations
- GeminiGoogle · assistant answers
- PerplexityCitation-first answer engine
- Google AI OverviewsGoogle Search · synthesised overview
- Google AI ModeGoogle Search · conversational mode (udm=28)
Google AI Overviews and AI Mode run on the same search index. We report them separately and never pretend they are two different companies. Per-surface playbooks.
Illustrative example data. Scores, stances and citations are computed from each brand's own weekly cycle — see the methodology.
- 10
- Buyer questions
- 6
- AI assistants
- 60
- Checks a week
- 0–100
- Your score
Written from your real pages
Scored the same way
Stored before anything is scored
Comparable week to week
How it works
From your website to a number you can act on
Four stages, run the same way every week. Each one keeps its evidence, so any score we show you traces back to the exact answer it came from.
- 01
Your market, mapped from real buyer language
Onboarding reads your homepage, pricing page, features or docs, plus the equivalent pages on each competitor. From that evidence it builds ten prompts across awareness, consideration and decision stages — written the way buyers actually ask, with real budgets, team sizes and switching concerns.
- 02
Every question checked on every assistant
Ten prompts across six AI surfaces produces sixty checks per weekly cycle. Each raw answer is preserved before any interpretation happens, so the evidence behind every score is always one click away.
- 03
Each answer graded on warmth and placement
Every answer is graded on how strongly you were recommended, the sentiment of what was said, where you placed, and which sources decided it — including the winning competitor, the stated reason, your gap and a verbatim quote. Unusable responses are flagged rather than counted against you.
- 04
One score you can defend, updated weekly
Results are combined into a 0–100 visibility score alongside a simple citation share, stored per week so you compare like with like. Correct anything we misread and the score updates immediately — and every point traces back to the exact answer it came from.
Your weekly record
One grid, and the week a placement moved
Sixty scored cells a week. The grid shows where you stand today; the chart below shows the week a position slipped and whether it came back.
Citation matrix · current cycle
Stance by prompt by surface
4 of 10 rows shownstance · sentiment · position
| Tracked buyer query | ChatGPT | Claude | Gemini | Perplexity | AI Overviews | AI Mode |
|---|---|---|---|---|---|---|
| "We are a 20-person team moving off seat-based pricing. What should we shortlist?"Consideration | First choice | Recommended | Recommended | First choice | Alternative | Mentioned only |
| "Which tools integrate natively with Slack and Postgres under $2k a year?"Decision | Recommended | Not mentioned | Recommended | Alternative | Mentioned only | Not mentioned |
| "What are the real drawbacks of the incumbent before we buy?"Decision | Mentioned only | Cautioned against | Alternative | Not mentioned | Mentioned only | Not mentioned |
| "If you were building this stack from scratch today, what would you pick?"Awareness | Alternative | Mentioned only | Recommended | First choice | Not mentioned | Mentioned only |
- First choice
- Recommended
- Alternative
- Mentioned only
- Cautioned against
Illustrative example data. Scores, stances and citations are computed from each brand's own weekly cycle — see the methodology.
Drift monitor · eight cycles
A five-point fall, caught and recovered
A drop of five points or more between two complete weeks raises a warning. The threshold is deliberately conservative: a partial week can never trigger a false alarm, and a one-point wobble never pages anybody.
Stance distribution
Where the sixty checks landed
First choice9
Recommended14
Alternative11
Mentioned only7
Cautioned against2
Not mentioned17
The largest block is usually not-mentioned. That block is the opportunity, and it is also the number a mention-count dashboard hides completely.
The arithmetic
Two numbers, reported side by side
One number tells you how well you were recommended. The other tells you how often you showed up at all. The gap between them is usually the most useful thing in the report.
Position weight
1 / log₂(position + 1)
Position 1 scores 1.00, position 2 scores 0.63, position 3 scores 0.50. Logarithmic decay stops a long tail of low mentions from outranking one genuine recommendation.
Citation share
min(100, citing answers ÷ (tracked prompts × 6) × 100)
The unweighted presence rate. Useful for spotting breadth, dangerous on its own because it counts a caution as generously as a recommendation.
| Stance | What it means | Weight |
|---|---|---|
| First choice | Named as the primary recommendation | ×1.25 |
| Recommended | Listed as a core recommended option | ×1.00 |
| Alternative | Presented as a fallback or also-ran | ×0.75 |
| Mentioned only | Named in passing, no endorsement | ×0.50 |
| Cautioned against | Engine warns buyers away from you | ×0.00 |
Why cautioning scores zero, not negative
A zero multiplier means an engine warning buyers away from you cannot contribute to your score at any position. It does not subtract from other surfaces — instead it raises a critical incident with a correction kit, because a caution needs a response rather than an accounting adjustment.
When we get it wrong
You can correct us in two clicks
Reading an AI answer is a judgement call, and judgement calls are occasionally wrong. If we misread one, you fix it from the question list and the score updates straight away.
Correct the stance
Stance and sentiment are editable per citation. A corrected stance flows through the same weighting pipeline, so the historical snapshot reflects the corrected reading rather than appending a footnote.
See the raw evidence
Every score is traceable to the preserved raw answer, the engine quote, the extracted sources and the comparison to the previous cycle. Nothing in the interface asks you to trust an unexplained number.
Refusals are not zeros
When an engine refuses to answer, returns under twenty characters, or declares it cannot browse, the check is marked unusable rather than scored as a miss. Absence of evidence is never silently converted into negative evidence.
Full-cycle denominators
Per-surface scores are normalised against the full tracked prompt count, not against the subset processed so far. A partially completed cycle reads conservatively rather than flattering you with a small denominator.
Product walkthrough
Watch one full cycle, start to finish
A narrated pass through the tracker: how the matrix reads, how a competitor citation is traced back to the source that caused it, and how a corrected stance immediately moves the score.
Recording in production. The player below activates the moment the Loom URL is set — no third-party script or iframe is loaded before that.
Walkthrough coming soon
Common questions
The details, answered
Also published in a machine-readable format, so an AI answering the same question quotes exactly what you read here.
How is the visibility score calculated?
Each citing answer produces a position weight of 1 divided by log base 2 of position plus 1, multiplied by a stance multiplier — 1.25 for first choice, 1.0 for recommended, 0.75 for alternative, 0.5 for mentioned only and 0 for cautioned against. The sum is normalised against tracked prompts times six surfaces times 1.25, and capped at 100.
What is the difference between visibility score and citation share?
Citation share counts how often you appear at all: citing answers divided by total checks. Visibility score additionally weights where you appeared and how favourably. A brand can hold high citation share with low visibility if it is only ever mentioned in passing, which is why the platform reports both.
How quickly is a citation change reflected?
Immediately. Correcting or flagging a citation from the prompt matrix atomically recomputes the affected weekly snapshot, and the next processed response for that prompt and surface updates the score in the same pass. Dashboard reads invalidate on that write rather than on a timer.
Which engines are covered, and what about the two Google surfaces?
Six surfaces are tracked: ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and Google AI Mode. AI Overviews and AI Mode are two distinct Google Search experiences that share Googlebot, so the platform reports them separately while making clear they are not independent model providers.
Can we track prompts we choose ourselves?
Yes. Onboarding generates ten grounded prompts from a crawl of your site and your competitors, and you can add, edit, deactivate or restore prompts at any time. Newly tracked prompts are backfilled into the current cycle so they begin reporting within the same week.
See your own market
Ten prompts. Sixty checks. Zero more blind weeks.
Connect your site and competitors and see where you actually stand with every major AI assistant — before next week's recommendations go out without you. Fourteen days free, no card required.