# How to Measure Your Brand AI Visibility Baseline: A Repeatable Framework > Measuring AI visibility takes more than a few ad-hoc prompts. Here is a repeatable framework to benchmark mentions, citations, rankings, and competitor share. Source: https://www.starbox.work/en/insights/geo-guide/geo082/ (Starbox AI Insights, updated 2026-09-18) ## A Single AI Answer Does Not Reflect Your True Visibility When marketing teams first explore Generative Engine Optimization (GEO), they often start by typing their company name into an AI chatbot. If the brand appears, they assume their visibility is fine. If it does not, they scramble to publish more content. Both reactions are premature. Generative AI outputs shift based on model versioning, real-time web retrieval results, account history, geographic IP, and the specific interface used. A single mention or a single omission tells you very little about your brand's underlying presence in AI discovery. GEO focuses on ensuring that when prospective buyers ask AI engines for solutions, your brand is accurately identified, recommended as a qualified option, and backed by credible source citations. A reliable AI visibility baseline must answer four basic questions: 1. Does the AI engine mention your brand proactively without being prompted with your brand name? 2. When your brand is mentioned, where does it rank in the recommendation list? 3. Which web pages does the AI engine cite to back up its claims? 4. How often do direct competitors appear across the exact same query set? Without tracking these four points systematically, you cannot know whether changes made to your website, documentation, or digital PR actually influence how AI engines represent your business. To run a reliable baseline, you must separate two metrics: - **Brand Mention:** Your brand name appears in the AI's generated response text. - **Page Citation:** The AI engine provides a clickable reference link back to a specific URL. A brand mention means the model recognizes your company. A page citation means a specific URL served as verifiable evidence for the answer. Both matter, but they reflect different stages of optimization and must be tracked as separate metrics. GEO metrics also cannot be lumped in with traditional Search Engine Optimization (SEO). SEO tracks page-level rankings and click-through rates on search engine results pages (SERPs). GEO tracks whether generative answer engines recommend your brand, include you in buyer shortlists, and reference your digital assets. While their measurement frameworks differ, their technical foundations overlap: your web pages must be easily crawlable, clearly structured, and packed with easily verifiable factual entities. ![How to Measure Your Brand AI Visibility Baseline](/static/img/insights/geo-guide/geo082-2.jpg) ## Step 1: Build a Fixed Prompt Library You do not need hundreds of prompts to begin. A practical starting point is a standardized set of 30 questions distributed across five distinct query categories: brand, category, scenario, comparison, and entity queries. Running 30 questions across four different AI engines produces 120 baseline observation points. When you run the exact same prompts in future audits, you have an objective comparison. The number 30 is a guideline. Enterprises with multiple product lines, regional markets, or languages can expand the library to 50 or 100 queries. Early-stage teams can begin with 15. The volume can vary, but once you finalize your baseline set, the wording must remain locked. If you ask "What is the best [product category]?" in month one and change it to "Which [product category] vendor is most reliable?" in month two, your results are no longer comparable because the underlying prompt changed. | Query Type | Suggested Count | Core Objective | Example Prompt Format | |---|---:|---|---| | Branded | 4–6 | Verify what the AI knows about the brand and spot factual errors | "Tell me about [Full Brand Name]." | | Category | 6–8 | Test whether the brand enters broad unbranded category recommendations | "What are the top [Category] service providers to consider in 2026?" | | Scenario | 6–8 | Check if the brand is associated with specific use cases and problems | "How should an [Industry] enterprise solve [Specific Problem]?" | | Comparison | 6–8 | Test shortlist inclusion against primary rivals | "How do [Brand A], [Brand B], and [Brand C] compare for [Use Case]?" | | Entity | 4–6 | Verify foundational facts: founding year, leadership, core services | "When was [Company Name] founded, and what are its primary products?" | Write prompts using the natural phrasing real buyers use, rather than internal corporate jargon. For example, an industrial service provider should not only test "recommended industrial digital transformation solutions." It should also test questions like "What should I look for when sourcing a [specific manufacturing process] vendor?", "Which manufacturers hold [specific industry certification]?", and "Can [process] suppliers handle low-volume custom runs?" These phrases reflect actual procurement conversations. ![How to Measure Your Brand AI Visibility Baseline](/static/img/insights/geo-guide/geo082-3.jpg) ## Step 2: Select Testing Engines by Target Market Your choice of AI engines depends on where your prospective buyers research solutions. For international global markets, focus on ChatGPT, Google AI Overviews, Perplexity, Gemini, and Claude. For the Chinese domestic market, monitor Doubao (ByteDance's AI assistant), DeepSeek, Kimi (Moonshot AI's assistant), Yuanbao (Tencent's AI assistant), and ERNIE Bot (Baidu's AI assistant). You can supplement these with Tongyi Qianwen (Alibaba's AI model), Metaso (an AI search engine), or Zhihu Direct Search (the AI search engine from Zhihu, China's largest Q&A platform). Different engines pull from different source ecosystems and format answers uniquely. In the Chinese market, for example, industry practitioners observe that Yuanbao frequently cites WeChat ecosystem content, ERNIE leans on Baidu properties, Doubao references ByteDance platforms, and Kimi and DeepSeek frequently surface answers from public Q&A communities like Zhihu. However, these platform habits are industry observations rather than rigid rules. Never assume an engine will only look at one source pool; evaluate actual generated citations without bias. Even within a single engine, a single test cannot serve as a permanent judgment. Model updates, index refreshes, personalization parameters, and real-time web fluctuations all alter responses over time. Industry tracking by Profound has shown that Google AI Overviews—the AI-generated summary panels in Google search results—shift frequently for identical queries, regularly swapping out cited sources. Run every prompt through each selected engine once during round one to establish your starting baseline, then re-test under the exact same parameters on a set cadence. The purpose of a baseline is not to prove that a single response is permanently accurate; it is to establish a repeatable benchmark for comparison. ## Step 3: Standardize Your Testing Conditions Testing conditions are the most frequently overlooked variable in AI audits. For every testing cycle, record: - Account status: Logged in or anonymous/incognito; - Geographic location and IP address of the tester; - Interface language; - Web search toggle: Enabled or disabled; - Access method: Web browser, mobile app, or search engine integration. Never test half of your prompt set while logged in and the other half anonymously. Never run the first audit in Chinese and the follow-up audit in English while attempting to compare the scores directly. Complete each evaluation round within a tight time window. Lock in your 4 to 6 engines and execute the prompts in the exact same sequence. Copy and paste each prompt word-for-word. Do not submit follow-up prompts, provide conversational background, or rephrase questions mid-test to coax the AI into mentioning your brand. If a model asks clarifying questions, log the response as "Clarification requested." If it explicitly states it lacks verified information, record that verbatim. Those behaviors are valid data points. For enterprise teams, have two team members cross-check a sample of the audit. For instance, assign 20% of the prompts to a second tester working independently. If the results diverge significantly, check for discrepancies in account login state, IP location, interface mode, or prompt text before attributing the difference to model variance. ## Step 4: Log 9 Data Points Per Response Recording simple "yes/no" columns for brand appearances will not give you enough context to troubleshoot weaknesses. Log these 9 data points for every prompt response in a centralized spreadsheet: | Field | Logging Format | Analytical Purpose | |---|---|---| | Timestamp | Date and approximate time of day | Prevents conflating responses from different model release windows | | Engine & Interface | Engine name, model version, and access entry (web, app, search) | Distinguishes between different interfaces under the same provider | | Raw Prompt | Exact text, unedited | Guarantees identical execution in future re-tests | | Brand Mentioned | Yes / No | Calculates overall Brand Mention Rate | | Shortlist Position | #1, Top 3, Other, Not Listed | Evaluates recommendation prominence when the brand is selected | | Cited Source URLs | Full URLs of all clickable references | Identifies the specific web assets the model relies on | | Information Accuracy | Accurate, Partially Inaccurate, Severely Inaccurate | Identifies and prioritizes factual brand hallucinations for remediation | | Competitors Mentioned | Competitor brand names and mention counts | Calculates Competitor Share of AI Visibility | | Evidence Capture | Screenshot file link or shared chat link | Maintains permanent source proof for verification | Define your "Shortlist Position" criteria in advance. If an AI answer lists 10 vendors, counting a mention at position #9 as equivalent to position #1 will distort your findings. Group positions into clear tiers: Top Choice (#1), Top 3, Other Shortlisted, or Not Listed. Being suggested second carries far greater buyer weight than being buried at the bottom of a ten-item list. Do not merely log root domains for citations. Categorize each URL by type: official product page, help documentation, third-party press release, industry wiki, community discussion thread, or competitor comparison page. Categorizing citations helps diagnose root causes quickly: - If a prompt fails to mention your brand because the AI engine almost exclusively cites industry trade associations, regulatory bodies, and technical whitepapers, publishing more blog posts will not solve the issue. - If your brand is mentioned, but the citation leads to an outdated news article from three years ago, your priority is updating third-party factual references, not writing fresh top-of-funnel content. ## Step 5: Standardize 4 Core Baseline Metrics You do not need a complex BI dashboard to start, but you must calculate these four core baseline metrics consistently. ### 1. Brand Mention Rate Formula: `Prompts mentioning the brand ÷ Total valid prompts tested × 100%` If your brand appears in 9 out of 30 tested prompts, your Brand Mention Rate is 30%. If an engine experiences a server timeout, refuses to answer, or fails to generate an output, exclude that prompt from the denominator for that specific engine and flag it as an invalid test. ### 2. Consideration Rate (Shortlist Inclusion) This metric evaluates only category, scenario, and comparison queries that prompt the engine for vendor recommendations, software comparisons, or provider shortlists. Formula: `Prompts including the brand in recommendations ÷ Total valid recommendation prompts tested × 100%` The Consideration Rate tracks your presence during actual vendor evaluations far better than simple branded search queries. ### 3. Citation Coverage Rate Formula: `Prompts citing at least one owned domain URL ÷ Total valid prompts tested × 100%` In addition to the overall rate, track the absolute count of unique cited URLs and unique cited domains. Ten responses that all cite the same corporate overview page represent a very fragile baseline compared to ten responses citing ten distinct, deep product and technical pages. ### 4. Competitor Share of AI Visibility For comparative prompts, count the total number of times each competitor brand is mentioned across the test run: `Specific brand mentions ÷ Total mentions across all evaluated brands × 100%` This metric represents share of voice within your specific prompt set, not absolute market share. Clearly state the tested prompt set, engines, and test dates in internal reports so stakeholders understand the scope. In addition to these four metrics, track your Factual Error Rate separately. If an AI engine states the wrong founding year, misidentifies your headquarters, confuses your primary line of business, or invents features you do not offer, a high Brand Mention Rate is not a positive signal. Inaccurate outputs usually stem from stale news coverage, low-quality web directories, or scraped aggregator pages that require targeted factual cleanup. ## Three Common Visibility Gaps to Address First Once your baseline is established, resist the urge to publish generic content immediately. Identify which structural gap you are dealing with first. ### Gap 1: Branded queries return accurate answers, but category and scenario queries ignore you This pattern indicates the model recognizes your entity, but lacks strong contextual associations between your brand and specific customer problems, industries, or workflows. Publishing more general corporate profiles will not help. Instead, build focused landing pages that address real-world buyer scenarios: which problems your product solves, where your solution is not a good fit, clear operational boundaries, direct comparison parameters, and verifiable technical specifications. ### Gap 2: The AI mentions your brand, but never cites your owned web properties This indicates the model learned about your company through its foundational training corpus or secondary third-party coverage, but does not use your official domain as a primary retrieval reference. Audit your technical and structural foundations: - Confirm your site architecture permits search engine bots and AI crawlers without blocking key paths in `robots.txt`; - Ensure critical product specifications and value propositions are rendered in server-side HTML rather than buried behind complex client-side JavaScript; - Check that pages feature clear H2/H3 headers, transparent author attribution, and publication dates; - Provide structured FAQs and unambiguous factual data blocks. Google's 2025 public documentation regarding AI search surfaces reinforces that effective technical SEO remains the prerequisite for AI discoverability. A page must be crawlable and comprehensible before it can be cited. ### Gap 3: Your URLs are cited frequently, but the AI generates inaccurate brand details Fix your public corporate entities across the web before attempting to expand coverage. Audit your legal company name, operating names, executive leadership, founding date, primary business segments, professional certifications, and office locations. Ensure complete factual consistency across your corporate site, PR announcements, Wikipedia or Wikidata entries, and primary social channels. Generative models assemble responses from multiple web sources. When external sources contradict each other, the model is significantly more likely to hallucinate inaccurate information. ## Establish a 30-60-90 Day Retesting Cadence A baseline audit is not a one-time project. Implement a structured quarterly cadence: - **Day 0:** Establish your full foundational baseline across all engines and prompts; - **Day 30:** Re-run the identical prompt set across the same engines to evaluate natural baseline movement; - **Day 60:** Re-run the core prompt set, and introduce a secondary tier of high-intent prompts; - **Day 90:** Conduct a comprehensive quarterly review and assess the impact of on-site updates and PR campaigns. Never delete poorly performing prompts simply to make month-over-month metrics look better. If your team has limited bandwidth, monitor a core subset of 10 commercial-intent prompts weekly, while saving the full 30-prompt baseline for monthly execution. Core prompts should focus on active buying decisions: "how to choose," "vendor reliability," "suitability for [industry]," and "direct product comparisons." Use weekly checks to spot sudden volatility, monthly audits to track operational trends, and quarterly reviews to allocate content and PR budgets. A sudden one-week spike does not prove a single blog post moved the needle, nor does a temporary dip mean your strategy failed. Look for sustained directional trends over two to three evaluation cycles, corroborated by indexation patterns and third-party web citations. GEO requires ongoing observation of how engines crawl, parse, and cite your brand assets over time. ## Ready-to-Use Spreadsheet Headers To set up your baseline tracking sheet in Excel or Google Sheets, copy and paste this standard header row: `Date | Target Market & Language | Engine & Entry Point | Account Status | Prompt ID | Raw Prompt | Brand Mentioned (Y/N) | Shortlist Position | Owned Cited URLs | Citation Source Type | Factual Accuracy | Competitors Mentioned | Competitor Mention Count | Screenshot / Link | Notes` Capture clean, unedited source data before building automated scrapers or visualization dashboards. If an AI visibility metric cannot be traced back to the exact prompt, raw output text, and cited reference URLs, your team will struggle to diagnose why the numbers moved or what actions to take next. ## Frequently Asked Questions ### Can we measure our AI visibility baseline just by searching our brand name? No. Branded queries only show whether an AI model knows your company exists. They tell you nothing about whether your brand appears when buyers search for solutions, compare vendors, or evaluate specific use cases. Your prompt set must include category, scenario, comparison, and entity queries. ### Does a baseline prompt set have to contain exactly 30 questions? No. Thirty is simply a manageable starting point that yields statistically useful observations across multiple engines. Fewer than 15 queries usually creates blind spots, while more than 100 becomes labor-intensive to test manually. Choose a realistic volume that covers all five query types. ### How should we record AI responses that provide no source citations? Log them as "No citation" without guessing where the model found the information. If the model mentions your brand without a link, count it toward your Brand Mention Rate, but not your Citation Coverage Rate. Keeping these separate is critical for diagnosing performance. ### Why does the same AI prompt return different answers from one day to the next? Live retrieval indices, model weight updates, routing changes, and personalized retrieval parameters cause regular fluctuations in generative answers. Rather than treating any single response as definitive, maintain standardized test conditions and look for sustained trends across multiple evaluation rounds. ### If competitors have a high share of AI visibility, should we immediately build more backlinks? Not necessarily. Check which specific URLs the engines cite for competitors first. If competitors win citations through technical specification sheets, industry benchmark reports, or clear product fact sheets, build equivalent credible reference assets. More backlinks alone will not guarantee AI citations. ### Can we use the same prompt set for international and domestic Chinese markets? The underlying buyer intent can remain aligned, but the actual wording, terminology, and target AI engines must be customized. International buyers often ask about compliance certifications, minimum order quantities (MOQs), and delivery terms, whereas domestic buyers focus more on specific ecosystem integrations and local case studies. ### What is considered a passing score for an AI visibility baseline? There is no universal passing score across industries. Early-stage startups, niche B2B industrial suppliers, and consumer brands start from entirely different baselines. The goal of round one is not to achieve an arbitrary score, but to establish an honest reference point to track ongoing improvements. A baseline is only as reliable as the raw data behind it. Keep your records consistent, and your team will have clear, defensible benchmarks for every GEO decision going forward.