--- title: "Claude AI visibility baseline case study: what one audit revealed" slug: "claude-ai-visibility-case-study" category: "ai-visibility" canonical_path: "/articles/ai-visibility/claude-ai-visibility-case-study" meta_title: "Claude AI Visibility Baseline Case Study — Prime AI Visibility" meta_description: "A Claude AI visibility baseline case study on an anonymized tourism brand: what one first measurement of a fixed prompt set showed about mentions and citations." author: "Bob Generale" reviewer: "Alex Mannine" date: "2026-08-01" last_updated: "2026-08-01" read_time: "13 min" keywords: - Claude AI visibility - AI visibility baseline - brand mentions vs citations - answer engine audit - share of citation featured_image: "/brand/articles/ai-visibility/claude-ai-visibility-case-study.png" featured_image_alt: "One solid navy baseline bar followed by faint dotted outline bars of unknown height, crossed by a thin citrine reference line" og_image: "/brand/articles/ai-visibility/claude-ai-visibility-case-study.og.png" cta_mid_headline: "Establish your Claude visibility baseline" cta_mid_body: "Prime AI Visibility runs a fixed prompt set across Claude and the other major answer engines, then records who each answer mentions and which sources it cites — so you have a dated first measurement to remeasure against later." cta_mid_button: "Measure your baseline" cta_bottom_headline: "Start with a documented first measurement." cta_bottom_body: "Create a workspace, bring your buyer prompts, and capture a saved Claude baseline — mentions and citations, dated and repeatable — before you change anything." cta_bottom_button: "Establish your baseline" --- # Claude AI visibility baseline case study: what one audit revealed A Claude AI visibility baseline audit measures how Claude describes and cites a brand at a single point in time, before any changes are made. In an anonymized study of a national tourism organization, a fixed prompt set run across five AI platforms showed the brand was mentioned in roughly a third of answers, its own site was cited in roughly half, and owned pages were a small fraction of all cited sources. No improvement is claimed here. > **Who this is for:** marketing and communications leaders who want to understand what a rigorous first measurement of Claude AI visibility looks like — and what it does not prove — before commissioning work. > **This is a baseline, not an improvement study.** This page documents a first measurement of a fixed prompt set at defined settings and dates. It claims no gains, ranking changes, or improvements of any kind. When the same fixed prompt set is remeasured after documented changes, this page will be updated with those later results and their dates. ## Claude AI visibility baseline: the short answer 1. **A baseline is a dated snapshot, not a result.** It records what Claude and other engines said about a brand at defined settings and dates, so future measurements have something honest to compare against. 2. **Mentions and owned-site citations are different measurements.** In this anonymized study they diverged sharply — a brand can be named in an answer without its own pages being cited, and vice versa. 3. **AI outputs vary.** The study represents a defined prompt set, engine settings, and dates; it is not a universal or permanent statement of how any engine behaves. ## The business question The organization behind this anonymized study is a national tourism body — the kind of entity that spends heavily on brand awareness and wants to know whether that investment shows up when travelers ask AI assistants for recommendations. The question its leaders brought was narrow and answerable: *when someone asks an answer engine about destinations, activities, or trip planning in our category, does our brand get named, and does the engine cite our own official pages?* That is a measurement question, not a marketing one. The temptation in this situation is to reach for a promise — "we will get you cited by Claude." No responsible audit does that, because no engine documents or guarantees which brands it names or which sources it cites. The honest first step is to find out what the engines are actually saying right now, capture it in a way that can be checked again later, and resist drawing conclusions the data cannot support. That discipline is the whole point of a baseline. If you are new to the practice, our overview of [how a measurement-first program is structured](https://primeaivisibility.com/articles/ai-visibility/ai-visibility-strategy) frames where a baseline sits in the wider workflow. ## Why Claude mattered for this audience A Claude AI visibility baseline is worth isolating because Claude is one of several assistants this audience uses, and it behaves differently enough to warrant its own look. Anthropic's assistant can answer from its training or, depending on the interface and settings, draw on browsing — and the way it surfaces sources is not identical to a search-native engine. For a travel-research audience that asks open, planning-style questions ("where should I go for X," "help me plan a week doing Y"), Claude's conversational, synthesis-heavy answers are exactly the format where a brand either gets woven into the recommendation or gets left out entirely. Because engines do not document their brand-selection logic, we treated Claude as one measured surface among five rather than a special case to be reverse-engineered. The value was in comparing Claude's behavior against the other platforms on the identical prompt set — not in claiming to know why Claude named what it named. ## Company and market context (anonymized) To protect the client, every identifying detail is withheld: the organization is not named, nor is its country, region, specific campaigns, or the phrasing of any prompt that could reveal it. What can be shared safely is the shape of the market. This is a destination-marketing category with a small number of recognizable competing brands (other national and regional tourism bodies), a long tail of commercial intermediaries (booking platforms, tour operators, travel-guide publishers), and a very large volume of user-generated and editorial content that engines can draw on. That structure matters: in a category this content-rich, an engine has many non-official sources to cite, which is precisely why owned-site citation cannot be taken for granted. All figures and patterns below are **anonymized demonstrations**. They describe the *shape* of what one audit found, not precise public statistics you can lift and quote as audited fact. ## Prompt-set construction A baseline is only as good as its prompt set. The set here was built to mirror how real travelers phrase research to an assistant, spanning several intents: broad discovery ("best places to…"), comparison ("X versus Y for…"), practical planning ("how many days for…"), and brand-adjacent checks (questions where the organization's own remit is directly relevant). The goal is coverage of genuine buyer questions, not a set engineered to make the brand look good. Two rules kept it honest. First, the prompt set is **fixed** — the exact same questions are saved so they can be re-run verbatim later. A baseline you cannot reproduce is not a baseline. Second, prompts were categorized and disclosed by intent, so the later comparison can be read category by category rather than as one blurred average. Building a defensible set is closely related to the discipline covered in our guide to [scoping and running an audit for a real business](https://primeaivisibility.com/articles/ai-visibility/ai-business-context-strategic-visibility). ## The methodology categories For every saved answer, the audit recorded a consistent set of fields. We describe the *categories* here; we do not fabricate the specific values, because those belong to the client's private dataset. - **Model.** Which assistant produced the answer (Claude and four other platforms), captured per answer rather than assumed. - **Answer mode.** Whether the response was a direct synthesized answer or a browsing-augmented one, since the mode affects what sources appear. - **Browsing setting.** Whether retrieval/browsing was enabled, recorded as a fixed setting so re-runs match. - **Location.** The geographic context the query was issued from, held constant across the set. - **Date.** The date each answer was captured — non-negotiable, because engine behavior drifts and an undated answer is uncheckable. - **Saved answers.** The full response text preserved, not summarized, so a later reviewer can verify what was actually said. - **Citations.** Every source the answer surfaced, recorded separately from the prose so mentions and citations never get conflated. The point of listing these is not ceremony. It is that a baseline anyone can trust must be *reproducible* and *dated*. Fixing the settings and saving the raw outputs is what separates a defensible measurement from a screenshot someone took once and can never re-create. ## What Claude said at baseline (qualitative) Within the Claude subset, the answers were fluent, confident, and organized as planning-style recommendations — the format that makes Claude useful to travelers and consequential to brands. Qualitatively, the brand appeared in a **minority** of the saved Claude answers, and it earned a citation to its own official site in **fewer still**. When the brand was named, it was usually folded into a broader recommendation rather than presented as the primary authority. When it was absent, Claude typically leaned on general knowledge or third-party sources rather than the organization's own material. None of this is presented as good or bad. It is the starting line. The value is that these observations are saved and dated, so a future re-run can show movement — or its absence — instead of relying on memory. ## Brand mentions versus owned citations The single most useful lesson from this audit is that **brand mentions and owned-site citations are different measurements that diverge in practice.** Across the full five-platform baseline, the brand was mentioned in roughly a third of answers, while its own site was cited in roughly half of the answers that carried citations — and its owned pages were only a small fraction of all cited sources across the set. Read carefully, those numbers do not line up the way intuition expects, and that is the point. | Measurement | What it counts | What it does not tell you | |---|---|---| | Brand mention | Whether the brand name appears in the answer prose | Whether the engine used the brand's own content | | Owned-site citation | Whether one of the brand's own pages is cited as a source | Whether the brand was actually recommended | | Share of all cited sources | How much of the engine's sourcing is the brand's own pages | How the brand compares on prose recommendations | A brand can be recommended warmly while the engine cites a travel-guide publisher instead of the brand's site; equally, an engine can cite an official page as a fact source while recommending a competing destination in the prose. Tracking only one number gives you half the picture. This is the core idea behind [share of citation as a metric](https://primeaivisibility.com/articles/geo/share-of-citation-explained): presence in the answer and presence in the sources are separate things worth measuring separately. ## Competing brands and sources (types only) Because the audit recorded every brand and source named, it produced a map of the competitive field — described here by *type* only, never by name. Competing brands that recurred were other destination-marketing bodies and, in comparison-style prompts, adjacent regions. The cited sources clustered into recognizable types: large travel-booking and aggregator platforms, editorial travel-guide publishers, encyclopedic reference sites, community and forum content, and — less often than any of those — the organization's own official pages. The takeaway is structural: in a content-rich category, an engine has abundant non-official sources to reach for, so an owned page competes for citation against a crowded field of intermediaries. Understanding *which types* of sources dominate is more actionable than any single count, because it points to where the engine is currently getting its information. ## Accuracy and framing findings (types of issues) Beyond presence, the audit noted the *types* of accuracy and framing issues that appeared in answers — again described as categories, not attributed to any named answer: - **Staleness.** Some answers reflected older information, the kind that shifts as seasons, offerings, or logistics change. - **Conflation.** Occasional blending of the organization's remit with adjacent or neighboring entities. - **Framing gaps.** Recommendations that omitted context the organization considers central to how it wants to be understood. - **Source mismatch.** Facts sometimes attributed to third-party sources where an official page would have been more authoritative. These are logged as observations to re-check, not as failures to fix under a promise. Whether they change over time is itself a measurement. ## What a true improvement case study requires This page is deliberately *not* an improvement study, and it is worth being precise about what one would actually require. Too many "case studies" are a single flattering screenshot. A defensible before-and-after needs a minimum dataset, agreed before any work begins: | Element | Minimum standard | Why it matters | |---|---|---| | Fixed questions | 30–100 saved prompts, run verbatim both times | Reproducibility; too few questions is anecdote | | Disclosed categories | Prompts tagged by intent and shared | Lets reviewers read results by segment, not one blurred average | | Saved baseline responses | Full answer text preserved, not summarized | A later reviewer can verify what was actually said | | Predefined checks | Mention, owned citation, source type defined up front | Prevents moving the goalposts after seeing results | | Documented changes | A dated log of what was actually done | Separates correlation from wishful attribution | | Final measurement date | The re-run captured and dated | Engine drift makes undated comparisons meaningless | | Stated limitations | AI-output variance and scope disclosed | Keeps conclusions inside what the data supports | Even with all seven in place, a rigorous study reports *what changed between two dated measurements* — never a guaranteed outcome, because engine behavior is undocumented and probabilistic. If you are commissioning execution work, our review of [how to evaluate AI search visibility services](https://primeaivisibility.com/articles/ai-visibility/ai-search-visibility-services) covers what to demand of any provider before you start. ## Cross-model comparison note Running the identical prompt set across five platforms is what turns a single-engine curiosity into a signal. In this baseline, Claude's mention and citation behavior was not identical to the search-native engines in the set — a reminder that "AI visibility" is not one number but a distribution across surfaces that behave differently. A brand can be present on one engine and thin on another for the same question. That is exactly why measuring one engine in isolation is misleading, and why the cross-model view is part of the baseline rather than an add-on. ## Limitations and the next test The limitations here are not fine print; they are the frame. The study covers a **defined prompt set** at **defined settings and dates** for **one anonymized organization** in one category. AI outputs vary between runs, engines update without notice, and personalization and location shift results. Nothing here generalizes to a universal state of any engine, and nothing here is a promise about outcomes. The next test is straightforward: re-run the identical fixed prompt set after a documented period, at the same settings, and record the new dated measurement alongside this one. Only then can anyone speak honestly about change — and even then, only as an observation between two dates, not a guarantee. ### A note on execution opportunity (not Claude-specific evidence) Separately, and *not* as evidence about Claude, one public example of the kind of narrative work that earns organic reach is this [LinkedIn post on career vulnerability](https://www.linkedin.com/posts/bobgeneraleinteractivemedia_early-in-my-career-i-was-terrified-of-being-ugcPost-7485716316033765376-VSgn). It illustrates content that resonates with a human audience; it is explicitly not a Claude visibility result and should not be read as one. ## Correction and update policy This page will be corrected if any figure or characterization here is found to be inaccurate, and it will be updated when the same fixed prompt set is remeasured after documented changes — with the new dated results published alongside these baseline observations. Because the underlying answers are engine outputs that vary and drift, we treat this page as a living record: dated, reproducible, and open to revision rather than a fixed claim. ## Methodology and sources All examples on this page are **anonymized demonstrations** describing the shape of one audit, not precise public statistics; no client, country, region, or campaign is identified. **AI outputs vary** between runs, engines, settings, and dates, so this baseline represents defined questions, settings, and dates — not a universal or permanent state of any engine. This article was authored by Bob Generale; the methodology was reviewed by Alex Mannine. Prime AI Visibility provides measurement and diagnosis; managed execution is handled separately by Percepture. Disclosure: Prime AI Visibility has a commercial relationship with Percepture — a firm founded in 2004, a five-time Inc. 5000 honoree (2017–2021), and an NMSDC-certified minority-owned business ([Percepture's Inc. profile](https://www.inc.com/profile/percepture)). > **Establish your Claude visibility baseline** > > Prime AI Visibility runs a fixed prompt set across Claude and the other major answer engines, then records who each answer mentions and which sources it cites — so you have a dated first measurement to remeasure against later. > > **[Measure your baseline](https://app.primeaivisibility.com/sign-up)** ## References 1. Google Search Central, *AI features and your website* (2024). 2. Google Search Central, *Creating helpful, reliable, people-first content* (2024). 3. OpenAI, *Overview of OpenAI crawlers (OAI-SearchBot and GPTBot are distinct)*. 4. OpenAI, *Publishers and developers FAQ*. 5. Inc., *Percepture company profile*. ## Next steps 1. **[Frame the measurement-first strategy](https://primeaivisibility.com/articles/ai-visibility/ai-visibility-strategy)** so a baseline sits inside a plan rather than existing as a one-off screenshot. 2. **[Avoid the common brand-visibility mistakes](https://primeaivisibility.com/articles/ai-visibility/ai-brand-visibility-mistakes)** that make audits look better than the data supports. 3. When you are ready, **[create a Prime AI Visibility workspace](https://app.primeaivisibility.com/sign-up)** and bring 10 buyer prompts to capture a saved, dated Claude baseline. ## Frequently asked questions **What is a Claude AI visibility baseline?** It is a dated first measurement of how Claude describes and cites a brand across a fixed set of buyer prompts, captured at defined settings before any changes are made. It exists so future measurements have an honest, reproducible reference point to compare against, rather than relying on memory or a single screenshot. **Does this case study prove the audit improved anything?** No. This is explicitly a baseline, not an improvement study, and it claims no gains of any kind. Improvement can only be discussed after the same fixed prompt set is remeasured following documented changes, and even then only as an observed difference between two dated measurements — never as a guaranteed outcome. **Why were the numbers described only roughly and anonymized?** The organization is a real client whose identity and details are protected, so the patterns are presented as anonymized demonstrations of the shape of one audit rather than as precise public statistics. Presenting rough, labeled figures keeps the lesson intact while avoiding any claim that these are auditable public facts. **Why do brand mentions and owned-site citations differ so much?** They measure different things. A mention counts whether the brand name appears in the answer prose; an owned-site citation counts whether the engine used one of the brand's own pages as a source. An engine can recommend a brand while citing a third-party site, or cite an official page while recommending a competitor, so the two numbers routinely diverge. **Why measure Claude alongside other engines instead of on its own?** Because AI visibility is a distribution across surfaces that behave differently, not a single number. A brand can be present on one engine and thin on another for the same question, so running the identical prompt set across several platforms turns a single-engine curiosity into a comparable signal and prevents misleading conclusions. **Can you guarantee Claude will cite our site if we do the work?** No. No engine documents or guarantees which brands it names or which sources it cites, and their behavior is probabilistic and changes over time. A baseline measures what is happening now; any later work is reported as an observed change between dated measurements, never as a promised result. **What would a genuine before-and-after study require?** A minimum dataset agreed up front: 30–100 fixed questions with disclosed categories, saved baseline responses, predefined checks, a documented log of changes made, a dated final measurement, and stated limitations including AI-output variance. Without those elements, a "case study" is an anecdote rather than a defensible comparison. > **Start with a documented first measurement.** > > Create a workspace, bring your buyer prompts, and capture a saved Claude baseline — mentions and citations, dated and repeatable — before you change anything. > > **[Establish your baseline](https://app.primeaivisibility.com/sign-up)**