How to Measure AI Visibility: The Right Metrics for the Recommendation Era

Why rank tracking breaks the moment you point it at an AI

For twenty years, SEO measurement rested on one quiet assumption: that the system was stable. You typed a query, the engine returned a ranked list, and that list held still long enough to be tracked. Position 4 today was position 4 tomorrow, give or take. The whole apparatus of rank tracking, share-of-search, and CTR-by-position depended on the answer being a fixed object you could check on a schedule.

Generative engines violate that assumption at the root. They do not retrieve a fixed list. They generate a fresh answer each time, token by token, and that generation is probabilistic. The same prompt produces a different answer on the next run, and the difference is not noise around a stable mean. It is the mean moving.

The clearest evidence comes from SparkToro and Gumshoe, who ran the most comprehensive public test of AI recommendation consistency to date. Rand Fishkin and Patrick O'Donnell had 600 volunteers run 12 brand-recommendation prompts through ChatGPT, Claude, and Google's AI a combined 2,961 times in late 2025. The categories were ordinary commercial questions: chef's knives, headphones, cancer-care hospitals, digital marketing consultants, science fiction novels. The result is the single most important measurement finding of the year.

Across tools and prompts, the odds of getting the same list of recommended brands twice were under 1 in 100. The odds of getting that same list in the same order were closer to 1 in 1,000.

Sit with what that does to rank tracking. If a brand's "position" in an AI answer changes on essentially every run, then any single observation of position is a coin flip dressed up as a metric. A tool that reports "you rank #3 in ChatGPT for this prompt" is reporting the result of one die roll and calling it a standing. Run it again and you might be #1, #7, or absent. The position is not a property of your brand. It is a property of that one generation.

The instability is not confined to the order of brands. It reaches the sources too. Profound's 2025 volatility study measured "citation drift," the share of domains that appeared in a platform's answers one month but not the next for the same prompts. The drift is enormous: Google AI Overviews 59.3%, ChatGPT 54.1%, Microsoft Copilot 53.4%, Perplexity 40.5%. Roughly 40 to 60 percent of the domains an AI cites this month will be different next month, for identical questions. The ground under any citation-rank metric is moving by half every thirty days.

Two forces drive this. The first is non-determinism at inference: unless a model is run at temperature zero with strict deterministic settings, which commercial closed models almost never expose, the model introduces randomness selecting each next token. The second is the retrieval layer: live web grounding means the corpus the model reads is itself changing as the web changes, as the index refreshes, and as the platform reweights its sources. You are tracking a moving target with a moving ruler.

This is why the central error of 2026 GEO measurement is importing the SEO scoreboard wholesale. Teams stand up an "AI rank tracker," watch the number jump around, and conclude either that the tool is broken or that AI search is unmeasurable. Neither is true. The tool is faithfully reporting a quantity that was never stable enough to track at the resolution of a single observation. The fix is not a better rank tracker. It is a different unit of measurement.

From position to probability: what actually holds still

If a single answer is a coin flip, the way you measure a coin is not to flip it once and record "heads." You flip it many times and report the rate. The same move rescues AI visibility measurement. Stop asking "what position am I in," which is unstable by run, and start asking "across many runs of the prompts that matter, how often do I appear, and how prominently." That rate is stable even when every individual answer is not.

This is the conceptual core of the recommendation era. AI does not hand a buyer a ranked list to scroll. It hands them a short consideration set, composed on the spot, of the brands the model most strongly associates with that intent. Your job is not to hold a rank. It is to be in the set often enough that you are part of the buyer's shortlist before a human ever describes you. Appearing often beats appearing first, because there is no durable "first" to win.

The metrics that survive this shift are aggregate, not point-in-time. Four of them carry the load.

Presence rate, also called recommendation rate. The share of runs, across your tracked prompt set, in which your brand appears at all. AirOps frames the equivalent as the North Star: answers mentioning your brand divided by total relevant answers. This is the metric the SparkToro data validates, because while any one answer is random, presence rate across dozens or hundreds of runs is a real, repeatable signal of how strongly the model associates you with the intent. Benchmark from the measurement corpus: aim to appear in 30% or more of answers for your core category prompts; the strongest brands clear 50%.

Share of citation, or AI share of voice. Your presence relative to competitors, not in isolation. This is the metric the tooling has standardized on. Semrush defines AI share of voice using both how often your brand is mentioned and how prominently it sits within each answer, and updated the metric in October 2025 to weight prompts by how often they are actually searched. Ahrefs Brand Radar defines it as the percentage of brand impressions you capture out of total impressions across responses that mention any tracked brand, modeled over a database of 199 million search-backed prompts across six engines. Share of citation answers the only question an executive cares about: in the answers buyers see, what fraction of the recommendation is us versus them.

Sentiment. Presence is necessary but not sufficient. Being mentioned as a cautionary tale is not the same as being recommended. Sentiment tracks how positively or negatively the model frames you when it does mention you. The measurement corpus benchmark is 70% or more positive across platforms, with recurring negative themes flagged for content and PR response. Semrush's 2026 enterprise stack pushes this further with an LLM sentiment analyzer that simulates thousands of customer personas and reports the percentage of those conversations that ended in a recommendation for your brand versus a competitor. Sentiment is where presence converts to preference.

Source attribution. The list of domains the model actually pulled from when it built the answer. This is the bridge between measurement and action, because it tells you why you are present or absent. If competitors are being assembled from Reddit threads, G2 reviews, and a Wikipedia entry you do not have, your absence is explained and your fix is named. Attribution turns a score into a worklist.

Position asks where you placed in one answer. Presence, share of citation, sentiment, and attribution ask how the system thinks about you across thousands of answers. Only the second question has a stable answer.

Note the discipline these four share. Each is computed over a body of runs, not a single response. Each is a rate or a distribution, not a rank. That is what makes them robust to the non-determinism that destroys position tracking. You are no longer measuring the outcome of a die roll. You are measuring the loading of the die.

The prompt set is the instrument, not an afterthought

Every aggregate metric in the previous chapter inherits its validity from one thing: the set of prompts you run. Get the prompt set wrong and presence rate, share of citation, and sentiment all measure the wrong universe with perfect precision. The prompt set is the measurement instrument. It deserves the same rigor a survey methodologist gives a sampling frame.

The first principle is that a prompt set is a panel, not a keyword list. SE Ranking's 2026 guidance frames a tracking set as a representative panel of 20 to 40 prompts covering core topics, personas, and journey stages, functioning like a brand-tracking panel that yields directional visibility data. The number is deliberately modest. The goal is not coverage of every possible phrasing; it is a stable, repeatable sample you can run on a cadence and trust to move only when reality moves.

The second principle is that prompts are not keywords, and translating one to the other naively breaks the measurement. People do not type "best CRM" into an AI. They type full, qualified, conversational questions. The average conversational query in 2026 runs about 23 words, dense with the qualifiers that push a model out of explanation mode and into recommendation mode. "What's a good CRM for a 15-person agency that already uses Slack and hates long onboarding" is a prompt. "CRM software" is a keyword. Only the first one triggers the consideration-set behavior you are trying to measure.

The third principle is fan-out. Modern AI does not answer the prompt you typed. It decomposes it into sub-questions, retrieves against each, and synthesizes. The 2026 tracking practice reflects this: the majority of brand citations come from the secondary and tertiary fan-out branches, not the literal root query. Ahrefs Brand Radar models this explicitly, expanding seed prompts into related sub-questions using Google's People Also Ask corpus and a fan-out system before measuring. A prompt set that tracks only root queries measures the front door and misses the rooms where most citations are actually decided.

The practical construction follows from these principles:

• Source prompts from real language. Pull phrasing from sales-call transcripts, support tickets, and community forums, not from a keyword tool. The words buyers actually use are where the recommendation behavior lives.
• Span the journey. Cover problem-aware prompts ("how do I reduce onboarding churn"), category prompts ("best onboarding tools for B2B SaaS"), comparison prompts ("X vs. Y vs. Z"), and branded prompts ("is X any good"). Each stage exposes a different facet of presence and sentiment.
• Include unbranded competitive prompts. Share of citation is only meaningful on prompts where you are not named. Branded prompts tell you how the model talks about you; unbranded prompts tell you whether the model reaches for you at all.
• Run each prompt many times, per platform, per market. Because any single run is a coin flip, the unit of measurement is the rate across runs. AI answers also vary by geography and language, so multi-market brands must sample per country.
• Hold the set stable. The panel only yields trend data if it stays fixed. Changing prompts every month makes month-over-month comparison meaningless. Version the set; change it deliberately and document when you do.

A prompt set is a measurement instrument. Twenty to forty real, conversational, journey-spanning prompts, run repeatedly across platforms and markets, is the difference between a brand panel and a vanity dashboard.

This is also where most "AI visibility scores" quietly go wrong. A score computed over a thin or unrepresentative prompt set is precise and meaningless. The number can be tracked to two decimal places and still describe a universe your buyers never inhabit. The instrument has to be right before the readings can matter.

The four signals behind the GEO Score

Presence, share of citation, sentiment, and attribution tell you what is happening. They are the outcome layer: lagging, observed, and largely outside your direct control on any given run. They answer "are we visible." They do not answer "why," and they do not tell you what to change tomorrow. That gap is the entire reason the Visigent GEO Score exists.

GEO score measures the inputs that cause visibility, not just the visibility itself. It is built from four signals, each chosen because it is something a team can actually execute against, and each grounded in the citation-factor evidence rather than in the volatile output.

Signal one: entity clarity. Generative engines resolve entities before they cite strings. The model has to know with confidence who you are, what category you belong to, and how you connect to your people and products, before it can reach for you in a consideration set. The evidence that this is the gating factor is direct: in 2026 entity-authority analyses, brand mentions correlate with AI visibility at 0.664 versus 0.218 for backlinks — roughly three times more predictive. Entity clarity measures whether you are a resolved node the model trusts or an ambiguous string it cannot place. If the model cannot tell who you are, nothing downstream matters.

Signal two: citation density. Visibility flows through the corpus the model reads, and that corpus is concentrated. LLMs typically cite only 2 to 7 domains per response, far fewer than Google's ten blue links; ChatGPT averages around five. Citation density measures how present your brand is across the specific third-party surfaces that feed those few slots — the review sites, the community threads, the reference pages the models actually pull from. The structural reality makes this decisive: 57% of branded-query citations go to reviews, listicles, forums, social media, and case studies, not to brand-owned pages. Density off your own domain is where most citations are won.

Signal three: structured proof. Once the model reaches your content, it has to be able to lift a claim cleanly. The Princeton GEO study established the levers: adding statistics lifted visibility by up to 41%, the single most effective technique tested; adding quotations from named sources lifted citations 28%; citing authoritative sources lifted lower-ranked pages by as much as 115%. Structured proof measures whether your facts are written in shapes a retrieval system can quote without rewriting — discrete statistics, clean tables, self-contained answer passages, named attribution. Prose has to be summarized. Proof can be quoted, and the quotable content is what gets cited.

Signal four: corroborating mentions. No model trusts a single source for a recommendation. The dominant new signal is co-occurrence: an analysis of 2.2 million prompts found that models now cross-reference multiple independent sources before citing, and roughly 60% of AI answers come from training data with no live web search at all, meaning brands not consistently mentioned across authoritative sources before the training cutoff are simply absent from the majority of answers. Corroborating mentions measure whether your brand shows up, consistently, across enough independent and authoritative places that the model treats you as a safe, well-attested recommendation rather than an unverified claim.

The outcome metrics tell you whether you are being recommended. GEO score's four signals — entity clarity, citation density, structured proof, corroborating mentions — tell you why, and which one to fix first. One is a scoreboard. The other is a control panel.

The reason GEO score is built on inputs is the same reason rank tracking fails on outputs. The outputs are volatile by design; you cannot pull a lever directly on a number that changes every run. But the inputs are stable and controllable. You can build the entity, earn the corroborating mentions, restructure the proof, and seed the surfaces that feed citation density. Move those, and the volatile outputs move with them, on average, over time. GEO score measures the things you can actually change.

A measurement framework teams can adopt

Putting this together yields a framework that any team can run without a research department. It has three layers — instrument, outcomes, drivers — and a cadence that matches how fast each layer actually moves.

Layer one: build and freeze the instrument. Construct a representative prompt set of 20 to 40 conversational, journey-spanning prompts, sourced from real buyer language, including unbranded competitive prompts and fan-out sub-questions. Decide your platforms (ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini, Copilot, increasingly Grok) and your markets. Freeze it. Version it. This is your measurement instrument and it must hold still to be trusted.

Layer two: measure outcomes as rates, never as ranks. Run the prompt set repeatedly — many runs per prompt, per platform, per market — and report the four aggregate metrics:

• Presence / recommendation rate — share of runs you appear in. Target 30%+ on core prompts, 50%+ to lead.
• Share of citation — your presence versus competitors on unbranded prompts. The competitive number that travels to the executive team.
• Sentiment — how you are framed when mentioned. Target 70%+ positive; flag recurring negatives.
• Source attribution — the domains feeding each answer, which doubles as your action map.

Never report a single-run position as a metric. If a number cannot be expressed as a rate or distribution over many runs, it is a coin flip, not a measurement.

Layer three: diagnose with the four GEO score signals. When an outcome metric is low, the input signals tell you why. Low presence with strong content usually means an entity-clarity or corroborating-mention gap — the model does not resolve you or does not trust you yet. Present-but-not-cited usually means a citation-density gap on the third-party surfaces that feed the slots. Cited-but-paraphrased-away usually means a structured-proof gap — your facts are not in liftable shapes. The diagnosis points straight at the lever.

The cadence follows the physics of each layer. Outcomes move slowly and noisily, so do not overreact to weekly wobble; the SparkToro and Profound data say a single bad week is mostly randomness. A sensible rhythm:

• Weekly: presence rate, share of citation, sentiment, per platform. Watch trend, not single points.
• Monthly: competitive share-of-citation deep dive, source-attribution analysis, GEO score signal review.
• Quarterly: full strategic audit. Benchmark: a 10%+ quarter-over-quarter improvement in visibility indicates the strategy is working.

A final honesty about testing. Classic A/B testing largely does not work here. Feedback loops are slow, model updates arrive in batches rather than continuously, and the output noise is large enough to swamp small effects. The practical substitute is before-and-after measurement on the GEO score input signals — change the entity, the proof, the mentions — with enough time between measurement windows (4 to 8 weeks) to let the noise average out. You are not running split tests on the answer. You are improving the inputs and watching the rates move over quarters.

Track the instrument so it holds still. Measure outcomes as rates, never ranks. Diagnose with the input signals. React on quarters, not weeks. That is the entire discipline.

Why this is the standard, not a preference

It would be easy to read all of this as one vendor's opinion about which dashboard to buy. It is not. It is the only measurement model consistent with how these systems actually behave, and the behavior is now thoroughly documented.

The case is short. AI generation is non-deterministic, so any single answer is one sample from a distribution. SparkToro's 2,961-run study put the odds of the same list twice at under 1 in 100. The sources are volatile, so any single citation snapshot decays fast. Profound's drift study put the monthly turnover at 40 to 60% of cited domains. Together these two facts make point-in-time rank tracking not merely imperfect but structurally invalid — it measures a quantity that does not hold still long enough to be a measurement. The only signals that survive are aggregates: rates and distributions computed over many runs of a stable, representative prompt set.

The stakes make the precision worth it. This is no longer a fringe channel. ChatGPT crossed 800 million weekly users in late 2025 and 900 million by early 2026. Forrester's 2026 buyers'-journey survey found twice as many buyers named generative AI or conversational search their most meaningful research source as named any other source, outranking vendor websites, product experts, and sales reps. 73% of B2B buyers now use AI tools in purchase research. The consideration set the model assembles is, increasingly, the consideration set the buyer brings to the table. Measuring your place in it badly is worse than not measuring it, because a confident wrong number drives confident wrong decisions.

This is the contrast that runs through everything Visigent publishes. Tracking a rank is measuring a thing you cannot change, that does not hold still, in a system that does not rank. Measuring presence, share of citation, sentiment, and attribution — and driving them through entity clarity, citation density, structured proof, and corroborating mentions — is measuring the things you can actually influence, in the units the system actually produces.

Visibility is what you measure. Influence is what you execute. The right metrics make the second one possible, because you cannot improve what you have mismeasured.

The recommendation era did not make AI visibility unmeasurable. It made the old measurements obsolete and demanded better ones. The brands that adopt the new standard — rates over ranks, a frozen prompt panel as the instrument, input signals as the diagnosis — will know exactly where they stand and exactly what to change. The brands still running an AI rank tracker and refreshing it daily will be staring at noise, mistaking the jitter of a probabilistic system for a scoreboard. The mention is the signal. Measure the signal, not the roll of the die.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.