One number, 0 to 100, answering a single question: does an AI assistant put this brand in front of a buyer? It combines five measured dimensions, each capturing a distinct way a brand can fail — or succeed — at being seen.
This page publishes the scoring function, the sampling design, the engines and the limitations. No brand can pay to rank. Scores are estimates from a model, revised monthly, and reported with confidence intervals.
A large language model does not return the same answer twice. Ask it the same buyer question three times and you may get three different sets of brands, in three different orders. This is not a bug to be engineered away; it is how the systems work.
The consequence is unforgiving: a single query is not a measurement. Any "AI visibility score" derived from one call to one model is a screenshot of a coin toss. Most of what is currently sold as AI visibility measurement is exactly that.
Everything below follows from taking that problem seriously. We sample repeatedly, across paraphrased prompts and across four independent engines, and we treat the spread of the results as information rather than noise to be hidden. A brand that appears erratically is not the same as a brand that appears reliably, even when their average is identical — and the score says so.
Every edition is a fresh collection run. Nothing is carried over, adjusted or smoothed against the previous month.
| Prompts per category (n) | 12 buyer-intent questions, held fixed across editions |
|---|---|
| Draws per prompt, per engine (k) | 3, in independent sessions |
| Engines (E) | 4 |
| Sampled responses per category (N) | n × k × E = 144 |
| Sampled responses per edition | 2,160 across 15 categories |
Prompts are held constant between editions so that score movement reflects a change in the world, not a change in the question. Prompt-set revisions trigger a methodology version bump and are logged below.
Four assistants are queried, each through an independent lane running in parallel. Each lane resolves the query with live web retrieval enabled, so that what we record is what a buyer would actually see — not what the model recites from memory.
The exact model identifier and the access timestamp are recorded for every individual response and retained with the run. This is what makes a score movement diagnosable after the fact: when an engine changes underneath us, the log tells us when.
Collection is automated. A scheduled orchestration workflow fans the prompt set out across the four engine lanes in parallel, captures each raw response, and passes it to an extraction stage that resolves three things per response: which brands are named, in what order, and which sources the engine cited in support. Entity resolution maps surface forms to canonical brands, so that "Palo Alto", "Palo Alto Networks" and "PANW" collapse to one record. Results are written to a versioned store; every score on this site is traceable to the individual responses that produced it.
An automated extractor pulls every proper noun an engine emits. Left unfiltered, a question about cybersecurity vendors returns AWS, Python and Gartner alongside the actual vendors, and the ranking becomes meaningless. The pipeline therefore excludes, by rule:
Category membership is then reviewed by a human before publication. We consider this an asset, not an embarrassment: an unsupervised taxonomy produces a clean-looking index that is quietly wrong, and no serious benchmark ships one. The review adjusts who belongs in a category. It never adjusts a score.
Five dimensions, each normalised to the interval [0,1], each weighted, summed and scaled to 0 to 100. The weights are fixed for the lifetime of a methodology version.
Score = 100 × ( 0.30·p + 0.20·q + 0.20·e + 0.15·c + 0.15·a ) p Presence share of sampled responses citing the brand q Rank log-discounted position when cited e Engine coverage fraction of engines citing above threshold c Consistency 1 − coefficient of variation across runs a Source authority independence of the sources behind the citation
Is the brand cited at all?
The share of sampled responses, across the full prompt set and all four engines, in which the brand is named. A brand absent from the answer scores zero here. Product quality is irrelevant if the model never surfaces it.
p = citations / N
N is the total number of sampled responses for the category. Bounded [0,1].
How prominently?
Where the brand lands within the answer when it does appear. Being named first is not the same as being buried eighth. Position is discounted logarithmically, the standard decay used in ranked-retrieval evaluation.
q = mean over citing responses of 1 / log2(1 + pos)
pos = 1 yields 1.00; pos = 3 yields 0.50; pos = 8 yields 0.32. Bounded (0,1].
On how many engines?
How many of the four engines cite the brand above a minimum reliability threshold. Broad coverage signals durable, structural visibility rather than a single-model artifact that vanishes at the next update. A single lucky hit does not count as coverage.
e = |{ engines citing the brand in >= 10% of their responses }| / 4 The threshold exists to separate structural visibility from sampling noise. Bounded [0,1].
Is it stable?
How stable citation is across paraphrased prompts and repeated draws. Because LLM outputs are non-deterministic, a brand that appears only intermittently has fragile visibility. We score that fragility explicitly rather than averaging it away.
c = 1 - (sigma / mu), clamped to [0,1]
sigma and mu are the standard deviation and mean of the per-run citation rate. This is the complement of the coefficient of variation.
Why is it cited?
The independence of the sources an engine leans on when it surfaces the brand. A citation grounded in editorially independent third-party coverage is a stronger signal than one grounded in the brand’s own marketing pages, and it is far more durable across model updates.
a = weighted mean of the independence tier of each cited source
Independent third-party editorial = 1.0 · industry directory or aggregator = 0.6 · brand-owned or self-published = 0.2.
Presence is a binomial proportion measured over N sampled responses, so its uncertainty is calculable rather than a matter of opinion. We compute the 95 percent Wilson score interval on p and propagate it through the weighting to produce an interval on the published score.
A benchmark that hides its mathematics is asking to be trusted rather than checked. The formula, the sampling design and the engine manifest are all above. What stays closed is the material that would let the index be gamed rather than verified.
The prompt strings stay closed for one reason: publishing them would let a brand optimise against the twelve questions rather than against the category, and the index would stop measuring anything. This is the same reason a held-out test set is never released.
What this index cannot tell you. We publish this section because the alternative — implying a precision we do not have — is how measurement gets a bad name.
A high score means AI assistants put a brand in front of buyers. It does not mean those buyers convert. Roughly 82 to 88 percent of AI citations produce no click, and no link inside an AI answer is taggable, so click-tracking captures under 20 percent of reality — for us and for everyone else. Anyone claiming a direct citation-to-revenue number today is guessing. Modelling that link is our v4.0 research objective, not a current capability.
The Presence term is a binomial proportion over N sampled responses. We publish its 95 percent Wilson score interval alongside every score. Two brands whose intervals overlap should be read as tied, not ranked. Small gaps between adjacent brands are frequently not statistically meaningful, and we say so rather than manufacturing false precision.
Model versions, retrieval indexes and system prompts change without notice and without changelog. A score movement between editions can reflect a genuine change in a brand’s visibility, or a change in the engine. We log the model identifier and access date for every run so that shifts can be attributed after the fact — but we cannot always separate the two in real time, and we do not pretend otherwise.
We sample buyer-intent questions per category, not the full space of things a human might ask. A brand strong on queries outside our set will be under-measured. Widening prompt coverage is the single largest known source of improvement in the index.
Source ecosystems differ by language and market. Scores are not portable across languages, and the index currently says nothing about consumer categories.
The LLM Visibility Index is built and operated by ZivRank, and ZivRank is measured in it. It appears in the AI visibility agencies category alongside its direct competitors. This is a conflict of interest and we state it plainly rather than leaving it to be discovered.
Here is how it is contained:
We considered excluding ZivRank from its own category. We rejected it: removing the data point does not remove the interest, it only removes the reader's ability to audit it. Disclosure plus reproducibility is the stronger safeguard, and it is the standard applied by every credible index whose operator competes in a measured market.
What the index measures today is what engines cite. That is a description of the past. The open problems in this field are the two that follow — predicting citation before publication, and connecting citation to revenue. Neither is solved, by us or by anyone. Here is what we are building, and where it currently stands.
Replace the heuristic independence tiers of dimension a with an authority score propagated over the citation graph itself — which sources co-occur, which sources are cited by other cited sources. Authority becomes measured rather than assigned. Target: a source-level authority model with published rank correlation against the current heuristic.
A classifier that predicts P(cited) for a given page against a given buyer query, trained on the labelled citation record the index has been accumulating since 2024. Retrieval simulation plus a supervised head. Target metric: held-out AUC, published. This moves the index from describing the past to estimating the future — the open problem in this field, and the one no vendor has solved.
Causal estimation of the relationship between measured AI visibility and lagged business signals — branded search volume, demo requests, pipeline — using matched control groups and holdout experiments. Not click attribution, which is not technically possible inside an AI answer. Lift inference, with intervals. This is the hard problem the industry is currently papering over.
Each of these will ship with the same disclosure standard as the scoring function above: the method published, the evaluation metric published, the failure modes published. A capability we cannot evidence is a capability we will not claim.
Full scoring function published. Sampling parameters, engine manifest and confidence intervals disclosed. Limitations and conflict-of-interest sections added. Research programme published.
Engine coverage weighting refined. Vertical AI added as a tracked category.
Source authority introduced as a fifth dimension. Prompt-variant sampling expanded.
First public edition: presence, rank, engine coverage and consistency.
The LLM Visibility Index research programme has been running since 2024. The index was first published in January 2026 and is operated by ZivRank SAS, incorporated in 2026.
No. Scores are derived only from observed engine outputs. There is no paid placement, no sponsored tier, and no mechanism by which a brand can purchase a position. Brands are added to a category on the basis of category relevance, not commercial relationship.
The score is a weighted sum of five measured dimensions on a 0 to 100 scale: Presence (30 percent), Rank (20 percent), Engine coverage (20 percent), Consistency (15 percent) and Source authority (15 percent). Each dimension is normalised to the interval [0,1] before weighting. The full formula is published on this page.
Four: ChatGPT, Perplexity, Gemini and Claude. Each is queried through its own lane in an automated collection pipeline, with the model identifier and access date logged for every run.
ZivRank is scored by the same automated pipeline as every other brand, on the same prompt set, with no manual adjustment. Its category is disclosed on this page as a conflict of interest, and its underlying measurements are published in full so that any reader can check them. We consider disclosure plus reproducibility a stronger safeguard than exclusion, which would simply remove the data point without removing the interest.
Monthly. Each edition is a fresh collection run, not an adjustment of the previous one. The methodology version in force is recorded against every edition.