--- title: "How to Measure Brand Visibility in Claude" slug: "measure-brand-visibility-in-claude" category: "claude" canonical_path: "/articles/claude/measure-brand-visibility-in-claude" meta_title: "Measure Brand Visibility in Claude — Prime AI Visibility" meta_description: "How to measure brand visibility in Claude: exact formulas and denominators for mentions, recommendations, and citations, the web-search-on/off split, and a worked example." author: "Bob Generale" reviewer: "Alex Mannine" date: "2026-08-06" last_updated: "2026-08-06" read_time: "12 min" keywords: - measure brand visibility in Claude - brand mentions and citations in Claude - Claude visibility metrics - AI answer measurement - share of citation featured_image: "/brand/articles/claude/measure-brand-visibility-in-claude.png" featured_image_alt: "A three-by-three grid of nine equal squares, each filled to a different level with citrine against navy, like a row of small gauges" og_image: "/brand/articles/claude/measure-brand-visibility-in-claude.og.png" cta_mid_headline: "Create a Claude measurement baseline" cta_mid_body: "Prime AI Visibility runs a fixed prompt set through Claude at recorded settings, then counts mentions, recommendations, and citations against explicit denominators — a dated first number you can honestly remeasure against later." cta_mid_button: "Create a baseline" cta_bottom_headline: "Start measuring, not guessing" cta_bottom_body: "Bring ten buyer prompts, fix the run conditions, and capture a saved Claude measurement baseline with every raw answer attached." cta_bottom_button: "Create your baseline" --- # How to Measure Brand Visibility in Claude To measure brand visibility in Claude, run a fixed set of buyer prompts at recorded settings and count three separate things: how often Claude names your brand (mentions), recommends you as the answer (recommendations), and cites a page as a source (citations). Each metric needs an explicit denominator, a disclosed web-search state, and repeated runs — one answer is an anecdote, not a measurement. > **Who this is for:** analysts and marketing operators who need defensible numbers for how Claude describes their brand, and want the formulas rather than a black-box score. > **This is a metric dictionary, not a tool tour.** This page defines what to count and how; it does not compare products or walk the full audit workflow. For the software that should hold these numbers, see the [Claude AI visibility reporting tool features](https://primeaivisibility.com/articles/claude/claude-ai-visibility-reporting-tool-features) buyer's guide; for the repeatable test that produces them, see the audit protocol linked below. ## Measure brand visibility in Claude: the short answer 1. **Count three things, never one.** Mentions, recommendations, and citations are distinct outcomes with distinct denominators; a single "visibility score" hides which one moved. 2. **Every metric needs a denominator.** A rate is meaningless until you state what it is divided by — usually the number of *relevant* answers in your prompt set. 3. **Web-search state is part of the measurement.** With search on, Claude can cite live pages; with it off, it cannot. Never mix the two. 4. **One run is noise.** Repeat each prompt and report consistency, because Claude's answers vary by run, wording, location, and time. ## Web search on versus web search off Before you count anything, fix the setting. Claude's web search can be enabled or disabled; with it enabled, Claude can retrieve and cite live web pages, and with it disabled the answer reflects training data with no live citations [[1]](#references). These produce two different populations of answers. Owned-domain citation share, for example, is only meaningful for web-search-on answers, because search-off answers rarely cite anything. Record the state per answer and report each population separately. ## The Claude Observation Ledger: nine metrics with formulas This is the original framework this page contributes: a ledger of nine metrics, each with an explicit formula and denominator. Record them per prompt family, per web-search state, and per run, then roll them up. 1. **Valid-response count.** The denominator for most rates. It is the number of answers that actually addressed the prompt — excluding refusals, errors, and off-topic replies. `Valid responses = total runs − (refusals + errors + off-topic)` 2. **Observed mention rate.** How often Claude names your brand. `Mention rate = answers naming the brand ÷ valid responses` 3. **Explicit recommendation rate.** How often Claude recommends you as *the* answer, not merely lists you. `Recommendation rate = answers recommending the brand ÷ valid responses` 4. **Visible citation rate.** How often a page is cited as a source (web-search-on answers only). `Citation rate = answers citing any of your pages ÷ valid web-search-on responses` 5. **Owned-domain citation share.** Of all cited sources, what fraction are yours. `Owned citation share = your cited URLs ÷ all cited URLs` 6. **Competitor recommendation share.** Of answers that recommend someone, how often it is a competitor. `Competitor rec share = answers recommending competitor X ÷ answers with any recommendation` 7. **Factual fidelity.** Of answers that make factual claims about you, how many are correct against your approved source of truth. `Factual fidelity = accurate answers ÷ answers making a checkable claim` 8. **Answer consistency.** Across repeated runs of the same prompt, how stable the outcome is. `Consistency = most-common outcome count ÷ number of runs` 9. **Framing distribution.** The spread of tone/context across answers — recommended, listed among alternatives, cautioned — reported as a distribution, not a single average. The vocabulary here is deliberately aligned with the site's [mention-based definition of share of citation](https://primeaivisibility.com/articles/geo/share-of-citation-explained): an answer "cites" a brand when it names it at least once, a conservative denominator that keeps the number honest. Prime uses the same entity resolution when it [runs a repeatable Claude audit](https://primeaivisibility.com/articles/claude/claude-ai-visibility-audit) so that "Prime AI Visibility" and its variants resolve to one entity. ## The denominator, in plain terms Most arguments about Claude visibility numbers are denominator arguments. If you divide mentions by *every* prompt, including ones where naming a brand is off-topic, you get a low, misleading rate. If you divide by *relevant* answers — prompts where naming at least one brand is on-topic — you get a fair one. Decide the denominator before you run, write it down, and keep it fixed across remeasurements. Changing the denominator between runs is the fastest way to fake a trend. ## One-run variability and repeated tests A single answer tells you what Claude said once, under one set of conditions. It does not tell you what Claude usually says. Run each priority prompt several times, on separate occasions, and report the *consistency* metric alongside the rate. A brand mentioned in three of five runs is a different finding from one mentioned in five of five, even though both could be summarised as "mentioned." This is also why comparing your Monday number to your Friday number, run once each, proves nothing — the swing may be run variance, not your work. ## Citation visibility versus source influence A visible citation tells you Claude *showed* a source in that answer. It does not tell you the source *caused* the answer, or how much it influenced the wording. Claude may synthesise from material it does not cite, and cite material it barely used. Measure and report citation visibility honestly — it is what you can observe — but do not upgrade it to "influence." That distinction keeps a report defensible and keeps a remediation plan grounded in what you can actually change: the accuracy and reachability of your owned pages. ## Prompt-family and persona segmentation Roll the ledger up by prompt family, because a blended average hides where you are strong and weak. Useful families include branded prompts ("what is [brand]"), category prompts ("best tools for X"), comparison prompts ("[brand] vs [competitor]"), and use-case prompts ("how do I do X"). Segment further by persona when your buyers differ — a practitioner and a procurement lead ask different questions and get different answers. Regulated categories push this further still; a [healthcare AI visibility audit](https://primeaivisibility.com/articles/healthcare/healthcare-ai-visibility-audit) segments by patient, caregiver, clinician, and procurement audiences because the accuracy stakes differ by persona. The same fixed families should carry across every remeasurement so the trend is real, and each metric can then feed a [correction workflow](https://primeaivisibility.com/articles/automation/ai-visibility-crm-content-workflows) when a number falls out of tolerance. ## Rolling the metrics into a scorecard Nine metrics is a dictionary, not a report. Roll them into a small scorecard that a stakeholder can read in a minute, using word ratings rather than invented composite numbers. For each prompt family, rate presence (mention rate) as strong, partial, or none; rate recommendation as strong, partial, or none; rate accuracy as strong, partial, or none; and note the dominant framing. A family that is *strong* on presence but *none* on recommendation is a very different problem from one that is *partial* on both — the first is a positioning gap, the second is a coverage gap. Resist the urge to average the three into one figure: the whole point of separating mention, recommendation, and citation is lost the moment they collapse back into a single score. Keep the raw metrics and their denominators in an appendix beneath the scorecard so any rating can be re-derived — the same evidence-appendix discipline agencies use when they [report client AI visibility](https://primeaivisibility.com/articles/agencies/agency-client-ai-visibility-reporting). Commerce teams apply the identical split at the SKU level in an [e-commerce AI visibility audit](https://primeaivisibility.com/articles/ai-visibility/ecommerce-ai-visibility-audit), where product-level presence and recommendation diverge sharply. ## Confidence labels Attach a confidence label to every rolled-up number, driven by three inputs: how many valid responses backed it, how consistent the runs were, and whether the web-search state was clean. A mention rate from forty consistent, single-state runs is *high* confidence; the same rate from four mixed-state runs is *low*. Reporting confidence is not hedging — it is what stops a stakeholder from over-reacting to a number that a larger sample would not support. This is the discipline behind Prime's [executive visibility reporting](https://primeaivisibility.com/articles/measurement/ai-visibility-executive-reporting), where the confidence label travels with the metric. ## A worked example (illustrative, fictional brands) The numbers below are invented for arithmetic clarity. The brands *Lumora*, *Kestrel*, and *Vanta* are fictional and do not describe any real measurement. - Prompt family: "best project tools for small agencies," web search **on**, 20 valid responses. - Lumora named in 9 answers → **mention rate = 9 ÷ 20 = 45%**. - Lumora recommended as the answer in 4 → **recommendation rate = 4 ÷ 20 = 20%**. - Of 60 total cited URLs across those answers, 6 were Lumora's pages → **owned citation share = 6 ÷ 60 = 10%**. - Of 15 answers containing any recommendation, Kestrel was recommended in 8 → **Kestrel recommendation share = 8 ÷ 15 ≈ 53%**. - Lumora's "best tools" prompt run 5 times returned "mentioned" in 4 → **consistency = 4 ÷ 5 = 80%**. Read together, this fictional picture says Lumora is present but rarely the pick, its own pages are a thin slice of the sources Claude leans on, and a competitor owns the recommendation. Each number is re-openable because the raw answers and cited URLs were kept — which is the whole point of measuring this way. Contrast this with a manual approach in [tracking AI visibility manually versus with a tool](https://primeaivisibility.com/articles/ai-visibility/tracking-ai-visibility-manually-vs-with-a-tool), where the denominators tend to drift between runs. ## What Claude does not expose publicly Claude does not publish which pages influenced an answer beyond the citations it chooses to show, does not expose an internal ranking, and does not guarantee the same answer twice. Anthropic documents the web-search behaviour and separate crawler controls [[1]](#references)[[2]](#references), but not a brand-visibility feed. So every metric here is an *observation* of Claude's outputs, not a readout of its internals. Report accordingly. ## A note on variability AI answers can vary by platform, model or product, search state, location, prompt wording, time, and repeated run. Results describe a defined observation method, not a permanent universal rank. ## Methodology and sources This article was authored by Bob Generale; the methodology was reviewed by Alex Mannine. It defines observation metrics for a third-party measurement of Claude's outputs and is grounded in Anthropic's published web-search documentation. All example figures are explicitly fictional and illustrate arithmetic only; no real brand, run, or result is described, and no ranking or citation promise is made. > **Create a Claude measurement baseline** > > Prime AI Visibility runs a fixed prompt set through Claude at recorded settings, then counts mentions, recommendations, and citations against explicit denominators — a dated first number you can honestly remeasure against later. > > **[Create a baseline](https://app.primeaivisibility.com/sign-up)** ## References 1. Anthropic / Claude Support, *Enable and use web search* (2026). 2. Anthropic / Claude Support, *Does Anthropic crawl data from the web, and how can site owners block the crawler?* (2026). 3. Google Search Central, *AI features optimization guide* (2026). ## Next steps 1. **[Choose the tool that keeps this evidence](https://primeaivisibility.com/articles/claude/claude-ai-visibility-reporting-tool-features)** so your metrics stay auditable, not just charted. 2. **[Run the measurement as a repeatable audit](https://primeaivisibility.com/articles/claude/claude-ai-visibility-audit)** with a fixed prompt protocol and a retest schedule. 3. When you are ready, **[create a Prime AI Visibility workspace](https://app.primeaivisibility.com/sign-up)** and bring 10 buyer prompts. ## Frequently asked questions **What is the difference between a mention, a recommendation, and a citation in Claude?** A mention means Claude named your brand in the answer text. A recommendation means Claude put you forward as the answer, not just listed you. A citation means a page appeared as a source, usually only when web search is on. They move independently and need separate counting. **What denominator should I use for a mention rate?** Relevant answers — the count of valid responses to prompts where naming at least one brand is on-topic. Fix it before you run and keep it identical across remeasurements, because changing the denominator is the easiest way to fabricate a trend. **Does web search need to be on to measure citations?** Yes for citation metrics. With web search off, Claude rarely cites live pages, so citation rate and owned-domain citation share are only meaningful for web-search-on answers. Keep the two populations separate. **How many runs make a number trustworthy?** Enough that the consistency metric is stable — usually several runs per priority prompt across separate occasions. Report the sample size and consistency with every rate so stakeholders can judge confidence. **Can I tell which of my pages influenced a Claude answer?** You can see which pages Claude *cited* when web search is on, but citation visibility is not proof of influence. Claude may synthesise from uncited material. Measure what you can observe and label it as visibility, not influence. **Does Claude give the same answer every time?** No. Claude's answers vary by run, wording, location, web-search state, and time. That is why repeated runs and confidence labels are part of the method rather than optional extras. > **Start measuring, not guessing** > > Bring ten buyer prompts, fix the run conditions, and capture a saved Claude measurement baseline with every raw answer attached. > > **[Create your baseline](https://app.primeaivisibility.com/sign-up)**