LLM Visibility: The Metrics That Matter in AI Search

LLM visibility measures how often your brand appears in AI answers, whether it is used as a source, and how accurately it is described. Five metrics carry it: mention rate, citation rate, share of voice, sentiment, and source overlap. Everything else on a visibility dashboard is decoration until those five are defined and stable.

Provider and research citations below were checked in August 2026.

ChatGPT passed 900 million weekly active users by February 2026, on OpenAI's own count. The same systems are sensitive to how a question is arranged. A 2023 study by Pouya Pezeshkpour and Estevam Hruschka measured a performance gap of roughly 13% to 75% across benchmarks when the answer options were reordered.

Enormous reach, unstable arrangement. That pairing is why this channel needs a metric set rather than a screenshot.

Key Takeaways

• Five metrics carry the report: mention rate, citation rate, share of voice, sentiment, and source overlap. Each answers a different question your CMO will ask.
• Definitions come before dashboards: a metric nobody can state in one sentence gets argued with in the meeting, and the argument lands on whoever presented it.
• Contrarian: position inside an AI answer is the most over-reported number in this discipline. Drop it from the board slide first.
• Band Reporting beats point readings: run the prompt set twice and report the range. Volatility stops reading as your error.
• Source overlap tells you what to do next. The other four tell you where you stand.

What Is LLM Visibility?

LLM visibility is the measurable presence of a brand inside the answers language models produce for a defined set of buying questions. It has three parts, and a report missing any one of them has a hole a CMO will find.

• Appearance: whether you show up at all.
• Evidence: whether a source you own or influence was used to build the answer.
• Accuracy: whether the description of you is right.

Teams reach this in stages. The stage decides which metric is worth presenting. A single llm visibility checker run answers the first question on the ladder below and none of the rest, and what visibility in AI search depends on covers the mechanism underneath it.

The Five Core LLM Visibility Metrics

Five metrics, one sentence each. The sentence is the point. A definition your finance team can repeat back is a definition nobody relitigates in month four.

The two numeric columns are working defaults from category practice. No measured distribution stands behind them yet. Raise them the moment your own runs disagree with each other.

Mention Rate

What it counts: the share of prompts where the brand name appears anywhere in the answer. It is the cheapest of the five to collect. It is also the easiest to overstate, because a mention buried in a list of eleven vendors scores the same as a mention in the opening line.

Healthy against unhealthy: a more stable reading holds within a relatively narrow range across repeated runs of the same set. Large swings are a signal to inspect the prompt set, run conditions, and sample size before treating movement as real.

Citation Rate

What it counts: the share of prompts where a domain you own or influence is named as a source. It is harder to move and it means more. A citation shows that the source was surfaced alongside the answer, which gives you a visible link between the response and an external source.

The structural catch: citation rate is only measurable where the engine shows its sources. On engines that hide them, the metric is absent rather than zero. A dashboard printing zero in that cell is telling you something false about your programme.

Share of Voice

What it counts: your named appearances as a proportion of all brand appearances across the same answers. It needs a fixed competitor list. That list has to survive contact with the engines rather than reflect who sales names in a deal review.

How to keep it defensible: fix the list once a quarter and publish it beside the number. A share-of-voice figure whose denominator moved mid-quarter is two different numbers wearing one label.

Sentiment

What it counts: the share of answers describing the brand favourably against a published rubric. The rubric is the metric. Without one, two analysts score the same paragraph differently, and the result is an opinion with a percent sign attached.

How to report it: as a distribution across three bands, never as a single average. Averages hide the single answer calling your product unsuitable for regulated buyers. That is the answer you needed to see.

Source Overlap

What it counts: the share of the domains an engine cited that also mention your brand. Collect the domains that appear across your prompt set, check each for a mention of you, then divide. A low reading on a domain the engines lean on constantly is a work order with an address on it.

Why it earns its slot: this is the metric that converts a report into a task list. Changing what those domains say is a separate discipline, and how models choose what to cite sets out that side end to end.

Share of Voice vs Mention Rate

The two look interchangeable on a dashboard, and they answer different questions. Mention rate asks whether the category knows you exist. Share of voice asks how much of the answer space you occupy against named rivals.

They diverge whenever engines name several brands per answer. In mature categories that is most of the time. In narrow ones it is not.

Where the Two Diverge

A worked example. A vertical SaaS platform for construction firms runs a 40-prompt set. The brand appears in 16 answers, so mention rate is 40%. Across those same 40 answers the engines named 100 brand appearances in total, 16 of them yours, so share of voice is 16%.

The gap is a category fact. In this example the engines name about two or three vendors per answer. Against that backdrop, 40% presence alongside 16% share describes a brand doing well and sharing the podium. In a category that names ten vendors per answer, the same mention rate cuts the share to roughly a quarter.

Which Number Goes in Front of a CMO

Lead with mention rate for two quarters. It moves earliest, and it needs no agreement about who counts as a competitor. A challenge bounces off it, because the denominator is your own prompt set.

Introduce share of voice once the competitor list is published, then show the pair from that point on. Presence rising while share falls means the category got more crowded. That is a strategy conversation rather than a reporting failure.

Why Position Matters Less Than You Think

Position looks like the natural import from rank tracking, and it is the weakest number on offer. An AI answer is prose. Three vendors named in one sentence have no defensible first, second, and third, and a tool that assigns them one has made a scoring decision it rarely publishes.

Two forces move the ordering underneath you. Neither has anything to do with your programme.

What Position Measures

Arrangement, first. Pezeshkpour and Hruschka's 2023 experiment on multiple-choice tasks found a performance gap of roughly 13% to 75% across benchmarks when the options were reordered. It is published evidence that these systems are sensitive to sequence in the prompt. Their test measures input order rather than output order, so read it as a caution and not as proof.

Sampling, second. Run the same prompt twice. The same three brands can arrive in a different sequence with nothing underneath having changed. Order is a property models are sensitive to, not a verdict they deliver.

Band Reporting

Band Reporting is the fix, and it costs one extra run. Execute the prompt set twice per period. Record the high and the low for every metric. Put the band on the slide instead of the midpoint.

Two things follow. Volatility becomes a stated property of the medium instead of something your audience discovers alone. A move that clears the observed run-to-run band is more defensible than one that sits inside it. Changes inside the band should be treated cautiously rather than automatically reported as movement. Single-engine practice has its own conventions, and scoring position and sentiment on one engine is worth learning before you average anything across five.

Sentiment and Misattribution

Sentiment gets filed as the soft metric, and it carries the expensive failures. A brand absent from an answer loses a deal it never hears about. A brand described wrongly loses a deal it was already in, and the buyer never says why.

Llm visibility monitoring earns its budget on the second case. Absence is recoverable with publishing. A confident wrong statement circulating across several engines can require repeated corrections across the underlying source set before it disappears consistently.

Sentiment Without a Published Rubric

Write the rubric before the first reading. Keep it to three bands, each defined by what appears in the answer rather than by how it reads. Favourable recommends the brand for a named use case. Neutral lists the brand without a judgement. Unfavourable states a limitation, a risk, or a reason to choose something else.

Hold the rater still. Apply the rubric with the same classifier or the same person for the whole period. Changing rater mid-quarter produces a trend line that measures your staffing.

Misattribution Is the Expensive Failure

Track four error classes separately from tone: wrong pricing model, wrong category placement, a competitor's feature described as yours, and a discontinued product presented as current. Count them as incidents instead of as a percentage. One is worth escalating. Eleven is worth a project.

Misattribution does not trend. It appears, persists across runs until the underlying source changes, then disappears. An incident log serves it better than a chart.

Vanity Metrics to Ignore

Four numbers show up on AI visibility dashboards, look authoritative in a screenshot, and collapse under a follow-up question. Name them now and nobody asks you to defend them later. A dashboard built only from these is where monitoring stops short of telling anyone what to do on Monday.

Impressions and Estimated Reach

No engine publishes per-answer view counts to brands, so every impression figure is modelled. Ask for the method behind the number before it reaches a slide. A figure nobody outside the tool can reproduce is specific enough to feel like information and impossible to check.

Raw Prompt Counts

"We tracked 4,000 prompts this month" describes your spend. A count with no denominator, no competitor set, and no repeat runs measures how much the tool did. Report it once in the methodology note. Never on the summary slide.

Single-Engine Screenshots

One good answer from one engine on one day is an anecdote. Keep it as evidence for a specific claim, in an appendix. Screenshots mislead the moment they stand in for a rate, because nobody screenshots the run where the brand was missing.

Composite Scores You Cannot Decompose

A visibility score is useful when you can break it back into its inputs and reproduce it. A score whose formula is not published is a number you cannot defend when someone asks how it is calculated. Ask any vendor to decompose one, and treat the answer as part of the evaluation.

How to Build an LLM Visibility Dashboard

Build it as six panels. One owner each, and one decision each panel is allowed to trigger. Whatever your llm visibility tools export, assembling the raw readings into something a leadership team reads in ninety seconds is your job rather than the vendor's.

Stand the referral panel up in month one. HubSpot now tags assistant traffic as its own source category, which HubSpot's New AI Referrals Signal sets out. Tagging starts the day you switch it on, so the panel cannot be backfilled from last quarter's analytics.

What to Track Weekly

Almost nothing. Weekly readings on a monthly metric generate four data points inside the noise band. They invite four conversations about movement that has not happened.

Two exceptions belong on a weekly cadence. Alerting on a total disappearance from any engine, which is worth checking as a tracking fault before it is read as a visibility event. And the misattribution log, which is incident-driven and does not wait for a reporting cycle.

What to Track Quarterly

Source overlap, sentiment distribution, and the competitor list itself. Each is expensive to collect and slow to move. Each is better served by a deliberate review than by a line on a monthly chart.

Re-derive the prompt set quarterly too. Buyer language shifts and categories get renamed, so a set untouched for a year measures an older market than the one you sell into. Ask whether you can edit the set at all, which is a buying criterion when you compare prompt-level tracking across models.

Benchmarks by Category

The most requested number in this discipline is a category benchmark. It is also the least transferable one on offer. Share of voice is a fraction of a denominator that changes completely between categories, so a good reading in construction software and a good reading in a category with ten times the vendor count share a unit and nothing else.

Four structural factors set the range before any programme work happens:

• Size of the competitive set. A category with six serious vendors gives every brand a higher arithmetic ceiling than one with sixty.
• Volume of third-party discussion. Active forums, practitioner communities, and frequent comparison content give engines more to retrieve, which lifts citation rates for every brand in the category.
• Review-platform depth. Where a category has mature review coverage, engines lean on it. A brand with thin profiles carries a structural disadvantage no on-site work fixes.
• Category naming stability. When buyers and analysts have not settled on one name, prompt sets fragment and every metric reads lower for reasons unrelated to the brand.

Benchmark against yourself. Take a baseline in month one. Hold the prompt set and the competitor list still, then judge the programme against its own first reading. The construction SaaS above learns more from three consecutive quarterly readings of its own than from any published cross-category average, because the denominators finally match.

What Klarivo Monitor Reports, and What It Leaves Out

Klarivo Monitor runs targeted queries across ChatGPT, Claude, Perplexity, Gemini, and Grok, tracks each engine independently, and updates on a schedule the client controls. It reports mention rate at the query level, competitor share of voice, sentiment, per-provider reliability, and Top 10 Citation Sources. That last surface is the one that turns a report into work, because it names the domains the engines drew on for your category.

Four limits belong in the same breath:

• Citation reporting is domain-level. The surface names the domain an engine drew on. Page-level and thread-level attribution sits outside what it can claim.
• Every reading describes the queries it ran. Questions outside that set go unmeasured, so the query set is the scope of the answer.
• There is no counterfactual. A rise tells you the number moved. It cannot tell you what would have happened otherwise.
• A single headline score is never the whole answer. By the standard set three sections above, any composite is only as defensible as the numbers you can break it back into, which is why the five metrics above are the ones to report.

Running tracking and publishing as one programme exists for the loop between them. When mention rate moves, the citation-source list names the domains behind the move, so next quarter's outreach comes from evidence instead of from a planning session. Teams already publishing off-site every week should buy monitoring on its own and keep the capacity.

Ready to see what your five numbers look like? Book a Klarivo discovery call. Fifteen minutes, a slot you choose, and instant confirmation, with no form to fill in first. The tracking side runs in Klarivo Monitor.

Frequently Asked Questions

How much movement is worth re-reporting to leadership?

Prioritise changes that clear the observed run-to-run band. Treat changes inside that range as inconclusive unless they persist across subsequent readings. Say two runs of the same prompt set on the same day differ by six points. A five-point month-over-month change is then not a result and should not be presented as one. Publish the bandwidth once, at the top of the methodology note, and the question stops being asked.

Can you compare visibility across engines, or only within one?

Compare within an engine over time. Compare across engines only on the same prompt set, run in the same window. Absolute levels should be compared across engines cautiously, because engines differ in how many brands they name, how they retrieve sources, and whether they expose citations. Cross-engine comparison is legitimate for direction and misleading for level.

What cadence should an llm visibility tracker run at?

Monthly for the full prompt set, with two runs per reading so the band is real. Add weekly alerting for one condition only: a brand disappearing from an engine entirely, which should be checked as a broken integration before it is read as a visibility collapse. Anything more frequent produces readings you have already agreed not to act on.

How do you handle a competitor named inside your own brand's answer?

Count it in both metrics and read it as a signal. The brand slot enters the share-of-voice denominator, and the answer still counts as a mention for you. A rival appearing consistently inside answers about you means the engines see you as substitutable. That is a positioning finding for product marketing rather than a tracking anomaly.

What sample size makes a reading trustworthy?

Enough prompts that a single answer cannot swing the metric enough to change the story. The working default used here is at least 30 to 40 for presence metrics, and more for anything with a competitor denominator. The better test is empirical. Run the set twice and look at the spread: a set that disagrees with itself by ten points is too small, whatever its size.