- An AI visibility score is a composite of how often you're mentioned, how favorably, and how accurately across AI answer engines, usually scaled 0–100 or as a percentage. It's an estimate built from a sample, not a fixed property of your brand.
- A 45-day University of St. Gallen study measuring ChatGPT, Perplexity, Gemini, and Google AI Mode found cited sources turned over roughly 65% day to day, and identical prompts re-run simultaneously overlapped only 32–43%. Their own convergence analysis recommends at least 7 runs per prompt per day before trusting a detection rate, and a 2–4 week rolling window before trusting a trend.
- The same brand's citation rate can vary by up to 615x between platforms (Superlines, 62 brands across 10 platforms). A single blended score can hide a brand that dominates one engine and is nearly invisible on another.
- Semrush's AI Visibility Index and Ahrefs' Brand Radar are both real, large-scale tools, and they'll hand you different numbers for the same brand, because they draw from different prompt corpora, update on different schedules, and disclose different amounts of their own methodology.
- Before trusting any AI visibility score (a vendor's or your own), ask how many prompts, how many runs per prompt, over what time window, and whether it's blended across platforms or broken out per engine.
Search "AI visibility score" and you'll find two very different kinds of answers: a definition (what it is, how it's calculated) and a sales pitch (a vendor's dashboard with a number on it, implying that number is a settled fact about your brand). Neither one, on its own, tells you the thing that actually matters: how much to trust the specific number you're looking at.
Below is what an AI visibility score actually measures and how it's typically built, what a rigorous 45-day academic study found about how much these numbers swing (day to day and platform to platform), and a practical way to tell a trustworthy score from a guess dressed up with a decimal point. This builds on the KPI foundation covered in AI visibility KPIs; this piece goes deep specifically on the score itself, as an artifact, rather than the broader set of metrics around it.
What is an AI visibility score, and how is it typically built?
An AI visibility score is a composite metric, usually scaled 0–100 or expressed as a percentage, built from three things: how often a brand is mentioned in AI-generated answers to relevant queries (citation or mention frequency), how favorably it's framed when it is mentioned (sentiment: is the model presenting you as a leader, a budget pick, or a risk), and how accurately the model represents your actual facts (information accuracy: correct pricing, current features, the right category).
The mechanics behind most tools follow the same basic shape: a fixed set of unbranded, buyer-language prompts (the questions a real prospect would type, not your brand name) run repeatedly across the models that matter, with each mention counted and often weighted by where it appears in a ranked or listed answer (a mention leading the response typically counts for more than one buried at the end). The resulting composite gets normalized into the single number the dashboard shows you.
That mechanical description matters because it clarifies what a score is: an estimate built from a sample of runs, not a fixed property of your brand the way a domain age or a review count is. Sample-based estimates have a margin of error. Most AI visibility scores are reported without one.
Why is a single score one frame of a noisy signal, not a fact?
The clearest evidence for this comes from an actual academic study, not a vendor's marketing page. Researchers at the University of St. Gallen (Julius Schulte, Malte Bleeker, and Philipp Kaufmann) ran a 45-day measurement of ChatGPT, Perplexity, Gemini, and Google AI Mode, published as "Don't Measure Once: Measuring Visibility in AI Search (GEO)". Their headline finding: visibility needs to be characterized as a distribution, not a single-point outcome, because the underlying system is probabilistic in a way classical search results never were.
Two numbers from that study are worth knowing cold:
- Cited sources turned over roughly 65% from one day to the next. Two-thirds of what got cited yesterday is different from what gets cited today, for the identical brand and category, with nothing about the brand having changed.
- Identical prompts, re-run at the same moment, overlapped only 32–43%. This isn't day-to-day drift. It's the same question, asked simultaneously, returning substantially different answers.
The paper's own convergence analysis translates this into a practical rule: trust a per-brand detection rate only once you've run at least 7 runs per prompt per day, and trust a trend only across a rolling window of 2–4 weeks. A score built from fewer runs than that, reported as a single daily number, isn't wrong exactly. It's just reporting one sample from a distribution wide enough that the next sample could look meaningfully different.
Why does the same score still vary wildly by platform?
Even a properly-sampled score has a second problem: which platforms it's blended across. Superlines' March 2026 analysis of 62 brands across 10 platforms found citation rate varying by up to 615x for the identical brand depending on the engine. In their tracked data, Grok cited sources at 27.01% of responses, Perplexity at 13.05%, ChatGPT at 0.59%, and Claude at 0%. A single blended "AI visibility score" that averages across platforms this different can make a brand that completely dominates one engine and is nearly invisible on another look like an unremarkable middle-of-the-pack number, hiding the exact information you'd need to act on.
This is the same instability the University of St. Gallen study documents from a different angle, and the practical implication is the same: a trustworthy score is broken out per platform, not collapsed into one blended figure, because the blend erases the variance that's often the most actionable part of the picture.
Why do two vendor tools hand you two different scores for the same brand?
Semrush and Ahrefs both publish large-scale AI visibility tools, and they'll disagree with each other on the identical brand, not because one is wrong, but because they're not measuring from the same inputs. Semrush's 2026 AI Visibility Index scaled to an analysis of roughly 126 million U.S. AI search prompts collected January through April 2026, combining keywords from its own search module with AI-generated questions fed into major LLMs, a large sample, though the platform has drawn criticism for not disclosing its underlying data sources and methodology in much detail. Ahrefs' Brand Radar draws on a differently-built corpus of more than 421 million organic prompts, with disclosed monthly volumes per engine (ChatGPT and Copilot at roughly 13.3 million each, Perplexity 13.3 million, Gemini 12.4 million, AI Overviews 143 million, AI Mode 41 million), storing raw responses in a corpus you can search directly, updated monthly, which means it captures a trend rather than anything closer to real time.
Different prompt sets, different platforms weighted differently, different update cadences, and a different amount of disclosed methodology. Handed two different scores for the same brand from these two tools, the right response isn't to pick whichever number is more flattering. It's to ask what each one actually sampled before treating either as ground truth.
How do you tell a trustworthy AI visibility score from a guess with a decimal point?
| Ask this before trusting a score | Why it matters |
|---|---|
| How many prompts, and are they unbranded/buyer-language? | A score built from branded searches ("is [company] good") measures something different from one built on the queries an actual prospect would type. |
| How many runs per prompt, and over what window? | Per the University of St. Gallen convergence analysis: fewer than roughly 7 runs per prompt per day, treat the detection rate as unreliable; trust a trend only across 2–4 weeks. |
| Is it blended across platforms, or broken out per engine? | Citation rate can vary up to 615x between platforms for the same brand (Superlines). A blended score can average out the exact signal you need. |
| Does the vendor disclose its prompt corpus and methodology? | Two major tools (Semrush, Ahrefs) draw from different, unevenly-disclosed inputs and will disagree on the same brand. Treat an undisclosed methodology the same way you'd treat any other unsourced claim. |
Frequently asked questions
What's a good AI visibility score?
There's no universal "good" number. Raw scores vary too much by category, brand size, and which vendor's prompt corpus and methodology produced them to be comparable across brands or industries. Benchmark your own score against your own prior period and against named competitors measured the same way, not against an absolute threshold.
How often does an AI visibility score actually change?
A lot, even with nothing about your brand changing. A 45-day academic study (University of St. Gallen) measuring ChatGPT, Perplexity, Gemini, and Google AI Mode found cited sources turned over roughly 65% from one day to the next, and identical prompts re-run at the same moment overlapped only 32–43%. A score pulled from a single day's runs is closer to one frame of a noisy signal than a stable reading.
Why do different tools give me different AI visibility scores for the same brand?
Because they're not measuring the same inputs. Semrush's AI Visibility Index draws on its own combination of search-module keywords and AI-generated questions across roughly 126 million prompts; Ahrefs' Brand Radar draws on a 421-million-prompt corpus with disclosed per-engine monthly volumes and updates monthly. Different prompt sets, different update cadences, and different amounts of disclosed methodology will produce different numbers for the identical brand. Neither is necessarily wrong, they're answering slightly different questions.
Is a low AI visibility score bad?
Only if it was measured with enough runs to be trustworthy in the first place. Given how much a single day's measurement can swing (per the University of St. Gallen study, roughly 65% source turnover day to day), a low score built from a handful of runs might just be measurement noise. Check how many prompts, how many runs per prompt, and over what window before treating any single score as a verdict.
Sources referenced
- Schulte, Bleeker & Kaufmann, "Don't Measure Once: Measuring Visibility in AI Search (GEO)" (University of St. Gallen, 45-day study across ChatGPT, Perplexity, Gemini, Google AI Mode)
- Superlines: AI Search Statistics 2026 (March 2026 citation-variance analysis, 62 brands × 10 platforms)
- Semrush: 2026 AI Visibility Index (126M US AI search prompts, Jan–Apr 2026)
- Ahrefs: Brand Radar methodology (421M+ prompt corpus, disclosed per-engine volumes)
- SparkToro / Gumshoe.ai AI recommendation consistency study (Jan 2026, ~2,961 prompt runs)
This piece assumes the KPI foundation in AI visibility KPIs. For how these measurements translate into an ROI calculation, see AI visibility ROI; for the underlying definitions, see AI visibility strategy. For the vendor-coined metric that often gets confused with this score, see what is "share of model"?.
