Glossary
LLM Visibility
LLM visibility is the degree to which a brand appears — and appears accurately — in the outputs of large language models when users ask relevant questions. It spans mention frequency, recommendation position, sentiment, and citation of the brand's own pages. Unlike search visibility, it cannot be read from a rankings page; it must be measured by querying the models directly.
LLM visibility has several distinct layers, and conflating them leads to bad decisions. Being mentioned is not the same as being recommended; being recommended is not the same as being described accurately; and none of those guarantee the model cites your pages rather than a third party's summary of you.
Visibility also varies by engine and by phrasing. A brand can be the default recommendation on Perplexity (which retrieves live sources) while being invisible on a base model whose training data predates the product's launch. Measuring one engine with one prompt gives a data point, not a picture.
The operational answer is sampling: a fixed set of buyer-intent prompts, run on a schedule across the engines that matter for your market, with results tracked as a trend. That converts visibility from an anecdote ('a prospect said ChatGPT recommended us') into a metric that can go up or down in response to your work.
What it is made of
LLM Visibility is an umbrella over several distinct measurements. Conflating them is what makes the term mean different things to different people, so each one is defined below with the arithmetic to compute it.
- Mention rate
-
How often the brand appears anywhere in the answer, regardless of whether it is endorsed. This is the floor: a brand that is never mentioned cannot be recommended, but a high mention rate on its own says very little.
How to compute: Prompts where the brand name appears ÷ total prompts run, per engine. Count an appearance once per answer, not once per occurrence.
- Recommendation rate
-
How often the brand is actually put forward as an answer rather than listed in passing. This is the metric that tracks closest to pipeline, and it is usually far lower than mention rate.
How to compute: Prompts where the brand is named among the recommended options ÷ total prompts. Decide in advance whether "also worth considering" counts, and apply that rule consistently across runs.
- Answer position
-
Where the brand falls when an answer names several options. Unlike a search results page there is no tenth position — answers typically name two or three vendors, so the difference between first and fourth is often the difference between being considered and not.
How to compute: Mean ordinal position across the prompts where the brand is recommended. Record the number of options each answer offered too; being third of three is not the same as third of eight.
- Claim accuracy
-
The share of checkable factual statements about the brand that are currently true — pricing, integrations, limits, certifications, ownership. High visibility with low accuracy is worse than invisibility, because the answer is confidently wrong.
How to compute: Correct claims ÷ total checkable claims extracted from the answers. Classify each as accurate, outdated or false; "outdated" is the bucket that usually dominates and the easiest to fix.
- Citation share
-
The proportion of the sources an engine cites that you control. It determines whether the engine is describing you from your own pages or from someone else's summary of you, which in turn determines how quickly a correction propagates.
How to compute: Citations pointing to your domain ÷ all citations across the answer set. Only retrieval-backed engines expose this, so compute it per engine rather than as a single figure.
- Share of model
-
Your recommendation rate relative to named competitors on the same prompt set. The absolute number matters less than the gap and its direction over time.
How to compute: Your recommendation count ÷ combined recommendation count of you and your tracked rivals, on an identical prompt set and run.
How to measure LLM visibility
Visibility is not readable from any dashboard an engine provides — there is no rankings page to check. The only way to get a number is to query the models yourself, on a fixed prompt set, on a schedule. The procedure below produces all six metrics above and needs nothing beyond a spreadsheet and an afternoon.
-
Fix a prompt set and then stop changing it
Write 20–30 questions a buyer in your category would actually type, spread across category, comparison, pricing and trust intent. Then freeze the list. Every metric here is a ratio measured against that denominator, so a prompt set that drifts between runs makes the trend meaningless — which is the most common way teams waste the first three months of measurement.
-
Decide which engines count, and be consistent
Pick the engines your buyers actually use rather than the longest list you can assemble. Coverage breadth is worth less than measuring the same engines every time, because visibility genuinely differs between a retrieval-backed engine and a base model whose training data predates your launch. Record which engine produced every answer; never pool them into one number.
-
Run each prompt in a clean session and keep the verbatim answer
Use a fresh session or incognito window per prompt so conversation history and personalisation do not contaminate the result. Save the full answer text and every cited URL, not a summary — the claim accuracy and citation share metrics are computed from the raw text, and you cannot go back and re-derive them from a score.
-
Score each answer on the six metrics before moving on
Work answer by answer: was the brand mentioned, was it recommended, at what position, which factual claims did it make, which sources did it cite. Scoring while the answer is in front of you is far faster than a second pass, and it forces the judgement calls — what counts as a recommendation — to be made once and applied uniformly.
-
Repeat on a schedule and read the delta, not the level
A single run tells you where you stand; it cannot tell you whether anything you did worked. Monthly is enough for most categories. The useful output is the change between runs on the same prompts and engines, with an eye on which specific questions flipped — a two-point move in an aggregate score is noise, whereas losing the top comparison question to a named rival is a concrete event you can act on.
All six metrics are arithmetic on a spreadsheet, and the first pass is worth doing by hand so you see what the answers actually say. Perciva automates the collection and the diffing; the judgement calls in steps one, two and four are yours either way.
Frequently asked
What does LLM visibility mean?
It means the degree to which a brand appears, and appears accurately, in the answers large language models give to relevant questions. It is an umbrella for six measurable things — mention rate, recommendation rate, answer position, claim accuracy, citation share and share of model — which is why two people using the term often mean different numbers.
How is LLM visibility different from SEO visibility?
SEO visibility is a position on a public results page you can check at any time. LLM visibility has no public page: answers are generated per user and per session, there is no referrer in your analytics when an engine tells someone to look elsewhere, and the result varies with phrasing. The consequence is practical — SEO visibility can be monitored, LLM visibility has to be sampled.
Can LLM visibility be tracked without a tool?
Yes. Fix a prompt set, run it in a clean session in each engine you care about, and score the answers in a spreadsheet. A first pass takes an afternoon. Tooling matters for the repetitive parts — running the set on a schedule and diffing answers between runs — rather than for the analysis itself.
Why does visibility differ so much between engines?
Because they source answers differently. A retrieval-backed engine reads live pages, so recent content and third-party coverage move the answer quickly. A base model answers from training data with a cutoff, so a product launched after that date can be invisible no matter what you publish. Reporting one blended visibility figure hides exactly the difference you would act on.
Is high visibility always good?
No. Visibility multiplied by wrong claims is worse than invisibility, because a buyer receives a confident, specific and incorrect statement about your product with no indication it should be checked. Track claim accuracy alongside mention and recommendation rates rather than after them.
How often should it be measured?
Monthly suits most categories, with an extra run after a model update or a significant content release. More frequent measurement mostly surfaces the natural variance of non-deterministic answers rather than real change.
See what AI currently says about your brand
Perciva runs buyer-intent prompts across ChatGPT and Gemini and shows you the answers verbatim — including which competitor gets recommended instead of you.
Get a free snapshot