How to Measure Your Brand's Share of Voice in ChatGPT, Gemini and Perplexity: Step-by-Step Guide

TL;DR

Measuring your brand's Share of Voice in AI engines is not the same as measuring organic traffic in Google Analytics. It requires defining a universe of strategic prompts, running them systematically across relevant models, recording presence, position and cited sources — and repeating it frequently enough to detect variations. This guide explains the full process in 5 steps, which metrics to record in each execution and how to automate the process with GEO Metrics across the 9 main models simultaneously.

When someone searches your category in ChatGPT, Gemini or Perplexity "best GEO tools", "what health insurance do you recommend in Argentina", "best airlines to fly to Mexico" — your brand either appears or it doesn't.

If it appears, at what position? How often? With correct information or with errors? In which models yes and in which no?

AI Share of Voice answers those questions. And measuring it correctly is what separates a GEO strategy with data from a GEO strategy with assumptions.

Why Measuring AI Share of Voice Is Different From Traditional SEO

In SEO, the object of measurement is clear: a keyword, a URL, a Google ranking. Search Console gives you impressions and clicks. Ahrefs or Semrush give you the position. The metric is stable.

In the AI ecosystem, the object of measurement is more complex:

Models don't always respond the same way to the same prompt. ChatGPT may mention your brand today and not mention it tomorrow for the same query. Perplexity crawls the web in real time — its response changes with every new media mention. Claude works from its training corpus — its response is more stable but also harder to move.

There is no fixed position. On Google, your page is at position 4 for a given keyword. In an AI response, your brand can appear as the first recommendation, as the third option in a list, as a mention at the end — or not appear at all. Position within the response matters.

There are 9 distinct models with distinct behaviors. The same prompt in ChatGPT and Perplexity can generate completely different responses — with different sources, different positions and different brands. Measuring only ChatGPT is measuring 10-15% of the ecosystem.

There is no native measurement tool. Google Search Console exists because Google wants you to optimize for its search engine. ChatGPT, Claude and Perplexity have no equivalent. Measurement requires specific tools or manual methodology.

The 4 Metrics to Record

Before getting into the process, it is important to be clear on what to measure — because not all metrics are equal.

1. Share of Voice (%)

The percentage of times your brand appears in a model's responses for a set of strategic prompts. It is the central GEO metric.

Formula: (number of responses that mention your brand ÷ total responses executed) × 100

A SoV of 20% means your brand appears in 1 out of every 5 relevant responses from that model. A SoV of 2% means you appear rarely.

2. Average Position

Where in the response your brand appears when it does appear. Position 1 means first mention. Position 5 means fifth in a list.

This metric is as important as Share of Voice — or more so. A SoV of 15% at position 1 is radically different from a SoV of 15% at position 7. The first implies leadership; the second, residual presence.

3. Accuracy Rate (%)

What percentage of responses that include your brand are factually correct — aligned with the verified information your brand declares as truth.

This metric requires configuring specific verification prompts with the expected response. Without it, you know you appear but you don't know whether what they say is correct.

4. Domain Citations

Which URLs from your domain are being referenced by models when generating their responses. This metric connects GEO with SEO — it shows which owned pages are being used as sources.

The Process: 5 Steps to Measure AI Share of Voice

Step 1: Define Your Strategic Prompt Universe

Prompts are the equivalent of keywords in SEO — but they are not keywords. They are conversational questions of 10 to 20 words that represent the real queries your audience asks AIs about your category.

Criteria for choosing the right prompts:

• High intent: prompts that represent purchase or recommendation decisions ("best GEO tools for agencies", not "what is GEO")

• High chatbot frequency: queries people ask more in an AI than in Google — GEO Metrics' Chatbot Preference Score identifies which ones

• Representative of your category: they must cover the different angles from which someone might search for what your brand offers

The recommended range: between 10 and 40 prompts, grouped into contexts by product, service or intent. Fewer than 10 does not give sufficient statistical representativeness. More than 40 dilutes the analysis without adding proportional information.

How to discover the right prompts:

• Use GEO Metrics' Fan-out de consultas a IAs Chrome extension to extract the real sub-queries ChatGPT generates internally when processing searches in your category

• Analyze Bing AI Performance grounding queries to see which queries Microsoft Copilot uses in your sector

• Combine with Google Search Console data to identify which traditional search queries also have high chatbot intent

Step 2: Run the Prompts Across the Relevant Models

With the prompt universe defined, the next step is running them in the AI models and recording the results.

The 9 main models to cover: ChatGPT · Gemini · Perplexity · Claude · Copilot · DeepSeek · AI Mode · AI Overviews · Grok

The minimum frequency: daily for models with real-time web access (Perplexity, AI Overviews, AI Mode, Copilot). Perplexity can change its response within hours following a new press or Reddit mention. Without daily monitoring, those changes are invisible.

The manual execution problem: running 40 prompts across 9 models daily means 360 executions per day. Done manually, with manual recording of results, that is hours of daily work — for a single brand. Systematic measurement requires automation.

Step 3: Record the Metrics per Execution

For each (prompt, model) pair, record:

• ✅ / ❌ Does the brand appear?

• Position within the response (first mention = 1, second = 2, etc.)

• Mention text (how does the model describe the brand?)

• Cited sources (which domains does the model use as reference?)

• Total number of competitor brands mentioned in the same response

With that data you can calculate Share of Voice, average position and the competitive context for each prompt and each model.

Step 4: Aggregate and Analyze the Data

With multiple executions accumulated, the data starts to tell a story:

By model: in which models do you have the highest SoV? Where are you at position 1 and where at position 8? In which models do you not appear at all?

By prompt: which types of queries generate the most citations? Which are the prompts where your competitor leads and you don't appear?

By period: is SoV rising or falling? Has a recent PR action moved Perplexity citations? Has a content change improved position in AI Overviews?

By sources: which domains does the model cite when responding about your category? Are they yours or your competitor's? Those sources are the map of where you need to build presence.

Step 5: Compare With Your Competition

Share of Voice only makes sense in a competitive context. A SoV of 8% can be leadership in a competitive market or marginal presence in a concentrated one.

For each prompt and model, also record the presence of your main competitors. The gap between your SoV and your strongest competitor's defines the urgency of intervention.

The Manual Measurement Problem and How to Solve It

The process described above is methodologically correct. The problem is scale.

A company with 20 strategic prompts wanting to measure across 9 models daily needs 180 executions per day. With manual recording of position, sources and response text, that is hours of daily work — for a single brand.

An agency with 10 clients and 20 prompts per client needs 1,800 daily executions. It is operationally impossible without automation.

GEO Metrics automates the full process:

• Configure your prompts once and the platform runs them daily across all 9 models

• Automatically records Share of Voice, position, response text and cited sources

• Generates competitive benchmarking — measuring your competitors for the same prompts without consuming additional executions, thanks to the Comparison view

• Detects day-to-day variations and alerts on significant changes

• Integrates data with Looker Studio for combined dashboards with GA4 and GSC

• Generates prioritized recommendations based on detected gaps

For agencies, the unlimited brands model from the first plan makes it viable to manage the full client portfolio from a single dashboard with no additional cost per brand.

Set up your first GEO Metrics project → trygeometrics.com

A Real Example: How Data Varies by Model

To illustrate the real variation between models, here is data from a GEO tool in the Spanish-speaking market for the prompt "solutions to measure share of voice AI" — extracted from GEO Metrics on August 1, 2026:

The pattern is clear: the tool appears in some models but at low positions, and is completely invisible in Perplexity, Claude, Copilot, AI Overviews and DeepSeek for this prompt. The top cited domains reveal that hubspot.es (142 citations) and trueranker.com (70 citations) are dominating the space.

That analysis defines exactly what needs to be done: appear in the sources these models are citing — hubspot.es, trueranker.com, cyberclick.es, dageno.ai — so that AIs start including the brand in their responses for this prompt.

Without measuring, that gap was invisible.

The Warning Signals You Only Detect by Measuring

Continuous AI Share of Voice monitoring detects anomalies that would be invisible otherwise:

Active hallucinations. If a model starts generating incorrect information about your brand — wrong prices, services you don't offer, associations with competitors — the Accuracy Score detects it before it reaches your clients.

Position drops with no apparent cause. If your Perplexity SoV drops 30% in a week without you having done anything, a competitor may have secured mentions in sources Perplexity prioritizes. Without daily data, you wouldn't know it happened.

Uncaptured opportunities. If a new high-intent prompt in your category generates responses where your competitor appears at position 1 and you don't appear at all, that is a direct opportunity. Without monitoring that prompt, you would never know.

PR action impact. If you publish a press release with a proprietary data point in an authority outlet, how long does it take to move SoV in Perplexity? With daily monitoring, you can correlate action and result. Without it, it's impossible.

Frequently Asked Questions

How many prompts do I need for representative data? Between 10 and 40 is the recommended range, grouped into contexts by product, service or intent. The exact number depends on the breadth of your category. A SaaS tool with a single product can start with 10-15 prompts. An agency managing multiple brands across different sectors may need 30-40 per client.

How often should I measure? Daily for models with real-time web access (Perplexity, AI Overviews, AI Mode, Copilot). Weekly at minimum for Claude and base ChatGPT, which update less frequently. Strategic data review can be monthly, but data collection must be continuous.

Can I measure my competitors' Share of Voice? Yes. You can record your competitors' presence in the same executions where you measure your brand. GEO Metrics does this automatically in the Comparison view — without consuming additional executions.

What do I do if my Share of Voice is 0% across all models? It is the most common starting point. Before working on content or PR, make sure the entity fundamentals are in order: consistent name across all channels, correct JSON-LD, llms.txt at the domain root. Without that foundation, content and PR have limited impact.

Is AI Share of Voice correlated with web traffic? Partially. Perplexity and AI Overviews generate referral traffic because they include links to cited sources — that traffic shows up in GA4. Base ChatGPT and Claude do not generate direct web traffic. But high SoV on high-intent prompts impacts brand consideration and purchase decisions even without generating direct sessions.

Want to see your brand's Share of Voice across the 9 AI models right now?

Start measuring with GEO Metrics → trygeometrics.com