How to Measure AI Visibility: Metrics, Prompts and Baselines

To measure AI visibility, build a fixed set of buyer prompts and run them repeatedly across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews. For each run, record whether your brand is mentioned, cited and recommended, and how it is described. Then compare those rates against competitors and against a week-0 baseline. AI visibility is how often and how prominently your brand appears inside AI-generated answers. AI influence is the effect of those answers on the buyer's shortlist, follow-up search and purchase, often without a click to your site. As of October 2026, most enterprise teams already track rankings and traffic well. What they lack is a repeatable method. This guide covers how to measure AI visibility with defined prompts, formulas, baselines and a reporting model you can run every week.

Why Traffic Alone Misses Part of the Buyer Journey

AI answers now deliver the comparison and the shortlist before any click happens, and traffic data cannot see that step.

For years the chain was simple. You ranked for a query, earned a click, measured the session and attributed the conversion. Traffic worked as the lead KPI because it stood in for three things at once:

• Visibility: if people clicked, you were on the page they saw.
• Relevance: click-through rate showed whether your listing matched intent.
• Demand capture: sessions fed directly into pipeline reporting.

AI search breaks that chain. Google AI Overviews summarise an answer above the organic results. ChatGPT and Claude can explain a category, name four vendors and weigh them without linking anywhere. Perplexity cites sources, but the reader often gets what they need from the summary alone.

Consider a buyer who reads a ChatGPT comparison, drops two vendors and later types your brand name into Google. Your analytics credit branded organic search, and the decisive moment never shows up. This is the gap between visibility and clicks. Your brand can be persuasive in an answer that produces no session. It can also be missing from an answer that removes you from the shortlist before anyone reaches your site.

What traffic still measures well

• Demand that reaches your site, including AI-referred sessions from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai.
• Engagement and conversion quality once a visitor arrives.
• Branded search lift, which often reflects upstream AI exposure.

Where traffic under-reports AI influence

• Answers read without a click, so the exposure never registers.
• Google AI Overview clicks, which Search Console includes in standard web search data instead of reporting them as a separate surface.
• Assistant apps that strip referrers, so their visits land as direct traffic.
• Delayed journeys, where an AI answer shapes the shortlist and the conversion arrives weeks later through branded search or sales.

Traffic tells you who arrived. AI visibility metrics tell you who was in the room when the decision started.

How to Measure AI Visibility: Metrics, Prompts and Baselines

AI visibility measurement has three layers. The prompt set defines what you test, the metrics define what you record, and the baseline defines what counts as progress. If any layer is missing, the numbers stop meaning anything. The method applies the same discipline as tracking whether AI answers mention you, with enough rigour to report to a CMO.

Layer 1: Prompt set design

Your prompt set is the full list of questions you test every cycle, and it should mirror how buyers actually ask. That is the core of choosing prompts that represent your buyers. Build the set from sales call notes, search query data, Reddit threads, support tickets and competitor comparison pages. Then sort every prompt into a category.

Sizing rules for a first program:

• Start with 75–150 prompts for one product category. With fewer than 50, each prompt swings the totals too much.
• Keep roughly 70–80% of prompts non-branded and the rest branded or competitor-branded.
• Tag every prompt by funnel stage, product line, persona and geography so you can segment results later.
• Duplicate geography-sensitive prompts per market (US, UK, DACH) instead of mixing markets in one prompt.
• Freeze the set for a quarter. Add new prompts as a tagged cohort so trend lines stay comparable.

Standardizing prompts so the test is fair

When you write prompts for measurement, the job is to remove bias from the question. Leave the answer alone. A prompt like "Why is [Your Brand] the best CSPM tool?" only measures the model's willingness to agree.

1. Use neutral wording, such as "Which CSPM tools suit a 2,000-person fintech?", and avoid "top-rated" or "most trusted".
2. Exclude brand terms from every non-branded discovery prompt, including your own product names.
3. Write two or three phrasings per intent (short query style, full question, persona-framed) and average the results.
4. Document each variant with an ID, the exact text, its category, the date added and the platforms it runs on.
5. Run prompts in clean sessions (logged out, or with memory and custom instructions disabled) and set location deliberately.

The NIST AI RMF playbook notes that transparency enables meaningful visibility into AI workflows and reduces information asymmetry. A written prompt register does the same for your reporting, because anyone can rerun the test and get comparable results.

Layer 2: What to record for every run

• Whether your brand is mentioned (yes/no), and its position among the brands named.
• Whether your domain is cited (yes/no), plus which URLs and domain types are cited.
• Recommendation status: top pick, listed, mentioned with a caveat, or absent.
• A representation score, using the rubric in the next section.

Layer 3: Baseline methodology

A baseline is the week-0 snapshot of every metric, taken before new content, Reddit activity or PR work begins. Split each metric by branded vs non-branded prompts and by platform. Measure share of voice against three to five named competitors. Add two outside signals: 12 months of branded search history from Search Console and Google Trends, and AI-referred sessions and conversions in GA4. Run the full set three to five times across two weeks and average the results. A single run mostly captures noise, so it cannot give you a baseline.

AI Visibility Metrics, Formulas and Scoring

AI visibility metrics turn raw answers into ratios you can trend, compare against competitors and tie to outcomes. Use the same definitions across every platform and reporting period.

Discovery metrics

• Brand mention rate = prompt runs where your brand is mentioned ÷ total prompt runs. Example: mentioned in 42 of 300 non-branded runs = 14%.
• Share of voice = your brand mentions ÷ total mentions of all tracked brands across the prompt set. Example: 42 of 210 brand mentions = 20%.
• Competitive visibility gap = competitor mention rate − your mention rate. Example: a competitor at 31% and you at 14% = a 17-point gap.
• Average answer position = the mean ordinal position of your brand among the brands named, counted only when you are mentioned. A position of 1.8 means you are usually named first or second.

Citation metrics

• Citation rate = prompt runs where your domain is cited ÷ total prompt runs.
• Citation share = citations to your domain ÷ all citations in the prompt set. Example: 18 of 400 citations = 4.5%.
• Source mix = the share of all citations going to your site, competitors, Reddit, review sites, media and reference sources. Source mix shows which third parties the engines trust for your category.

Recommendation and representation metrics

• Recommendation rate = runs where your brand is the top pick or explicitly recommended ÷ runs where it is mentioned.
• Branded accuracy rate = branded prompts with no factual errors about your product ÷ total branded prompts.
• Representation score = a rubric score out of 10, as defined below.

Each month, have two people score the same 20-answer sample independently. If their scores differ by more than one point on average, tighten the rubric wording before you trust the trend.

Many teams combine these metrics into a composite AI visibility score for executive reporting. A composite works as a headline only if the component metrics stay visible underneath it, because a composite can rise while discovery falls.

Outcome metrics for triangulation

Direct attribution from an AI answer to revenue is rarely possible. Instead, look for three signals moving together:

• AI-referred sessions and conversions in GA4. Set up a custom channel group with a regex on referrers such as chatgpt\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai.
• Branded search impressions in Search Console, trended against non-branded mention rate with a four- to eight-week lag.
• Self-reported attribution. Add "AI assistant (ChatGPT, Perplexity, etc.)" as an option to "How did you hear about us?" on demo forms.

How Measurement Differs by Platform and Business Model

Each AI platform rewards a different kind of presence, so metric priorities should change by platform and by how your business sells. A single blended number hides those differences.

Controlling volatility

AI answers are non-deterministic, so the same prompt can name different brands on consecutive runs. NIST's work on deployed AI systems states that post-deployment measurement and monitoring is necessary to confirm that a system behaves as expected in real-world use. For visibility tracking, that means repeated, ongoing measurement instead of spot checks.

• Run each prompt three to five times per platform per cycle and record the mean.
• Report on a four-week rolling average. Treat week-over-week moves under five points as noise unless they persist.
• Fix location and language per market, because retrieval-based answers change by region.
• Log the model and mode used (for example, ChatGPT with search on vs off), because API answers and consumer app answers can differ.
• Mark platform model updates on the timeline, so a step change is not misread as your own win or loss.

Metrics by business model

• Enterprise B2B and SaaS: weight commercial-investigation and competitor-branded prompts. Lead with share of voice and recommendation rate, and check them against demo-form self-attribution.
• Consumer brands and ecommerce: weight "best X for Y" prompts. Track product-level mentions, review-site citations and the accuracy of price and availability.
• Local and multi-location businesses: duplicate prompts by city. Track mention rate per market and the accuracy of addresses and opening hours.
• Publishers: lead with citation share and URL-level citations, because being the cited source is the product.
• Healthcare and higher education: weight accuracy and representation score heavily. An incorrect claim carries regulatory and reputational risk beyond lost visibility.

Reading the Data: A Worked Dashboard, False Positives and Next Steps

A useful AI visibility dashboard shows a small set of comparable fields, and each field has a clear meaning when it moves. The example below uses illustrative figures for a cloud security vendor that tracks 120 prompts across four platforms. It compares week 0 with week 12.

Bad baselines and false positives

How to act when a metric falls

• Falling mention rate: the usual causes are weak entity association, thin topical coverage or outdated comparison content. Publish answer-first comparison and use-case pages, and refresh pages that describe old features.
• Falling citation share: competitors or third parties publish more quotable material. Add original data, clear definitions and structured headings, and check that robots.txt does not block GPTBot, PerplexityBot or ClaudeBot.
• Source mix shifting to Reddit or review sites: the engines trust community voices for your category. Build a genuine presence in the threads being cited and keep review profiles current.
• Falling representation score: the cause is usually inconsistent facts across your site and third-party profiles. Align Organization and Product schema, update directory listings and correct outdated coverage.
• Worse answer position: a rival has stronger "best for" positioning. State explicitly which segments and use cases you fit.

Where Tellr Fits in an AI Visibility Program

Tellr is a premium earned-visibility agency. It uses measurement to direct the work that moves AI visibility and does not sell a dashboard. Each week it tracks the category's queries on Google and records the organic results, the AI Overview and the discussions block. It labels every cited domain as your site, a competitor, Reddit, social, review sites or references, and shows week-over-week movement. A senior team then works on those gaps through one governed program, with approvals, an audit trail and a monthly program review. Tellr needs a few weeks to map the category and agree the brief, so it does not suit a three-week launch.

• Weekly citation share by domain type for Google and AI Overviews.
• Guideline-checked Reddit replies behind an approval gate, reported as threads found, replies placed and whether each reply is still live.
• Content written to be quoted by ChatGPT, Perplexity, Gemini, Claude and AI Overviews, with rankings and citations reported per article.

An Implementation Checklist for Your First 90 Days

A first AI visibility program can be running and reporting within 90 days if you build the prompt set, the baseline and the reporting cadence in that order.

1. Weeks 1–2: collect buyer questions from sales, search data, Reddit and support. Draft 75–150 prompts and tag them by category, funnel stage, persona and market.
2. Week 2: standardize wording, write two or three variants per intent and log every variant in a prompt register.
3. Weeks 3–4: run the full set three to five times per platform and record week-0 values for every baseline metric.
4. Week 4: set up the GA4 AI-referral channel group, add an AI option to demo-form attribution and pull 12 months of branded search history.
5. Weeks 5–12: rerun the set weekly, report four-week rolling averages, mark model updates and score representation monthly with two reviewers.
6. Week 12: compare against the baseline by platform and by branded vs non-branded prompts, and assign a next action to every declining metric.

Knowing how to measure AI visibility comes down to a stable prompt set, consistent formulas and an honest baseline. With those in place, you can judge every new page, thread reply or PR placement by one test: whether it moves the numbers in the answers your buyers actually read.

FAQ