GEO Metrics: How to Measure AI Visibility, Share of Voice and ROI

GEO metrics measure whether AI platforms mention your brand, describe it correctly and link to you, whether their crawlers can reach your site, and whether any of it sends visitors. Track each one per platform and per mode (answered from memory or with web search), because a single blended "AI visibility" number hides where it came from.

GEO is generative engine optimization: changing how answer engines such as ChatGPT, Perplexity, Gemini, Google AI Overviews and Claude describe you (our guide to what generative engine optimization is covers the basics). This post covers what to measure and how, what no tool can measure, and how to compare competitors and estimate ROI without inventing numbers.

The GEO metrics that hold up

The first five come from asking AI platforms questions and reading the answers. The last four come from your own logs and the platforms' own reports. You need both kinds: the first says what AI says, the second says what reaches you.

Metrics from your prompt set

Every answer-based metric starts with the same sample: a fixed set of prompts, asked on each platform, several times, with the date and mode logged. Twenty prompts on five platforms, run three times, is 300 answers. Change the prompts and you have a new sample, not a trend.

Mention rate

Mention rate = answers that name you ÷ all answers. Report it per platform and per mode, never only as one blended figure. Being named by ChatGPT from memory and being named by Perplexity after a live search are different events with different causes.

Accuracy rate

Accuracy rate = facts stated correctly ÷ facts stated. Check founding year, headquarters, CEO, ownership, prices, plans and features against a fact sheet you wrote. Keep it separate from mention rate: a brand named with the wrong price is worse off than one not named, and a blended score hides that.

Citation rate

Citation rate = answers that link to your domain ÷ all answers. Expect it to vary by platform for structural reasons: Perplexity, Gemini with Google Search grounding and AI Overviews cite sources, while ChatGPT and Claude answering from memory seldom link anything. A low ChatGPT citation rate from memory answers is normal, not a failure.

Sentiment

Classify each answer as positive, neutral or negative with a written rubric, and keep the rubric fixed. Sentiment drifts with whoever classifies it, so the same person or the same model should score every run.

First-party data: what the platforms report themselves

Google Search Console: Generative AI performance report

The report shows impressions for AI Overviews and AI Mode, by page, country, date and device. As of August 31, 2026, Google says it has rolled out to all websites. It does not show clicks, click-through rate or position.

Clicks from those features exist, but Google counts them in the ordinary Performance report: AI features traffic is reported within the "Web" search type, mixed with classic results. You cannot separate them.

Bing Webmaster Tools: AI Performance

Bing's report counts citations in Microsoft Copilot, Bing's AI summaries and select partner integrations, with page-level counts and grounding queries: the phrases the AI used when it retrieved your content. Microsoft says the grounding query data is a sample and that citation counts do not indicate placement within an answer.

Grounding queries are the closest thing anywhere to the real queries an AI ran. Compare them with your prompt set. If they barely overlap, your prompts are measuring a different question from the one buyers ask.

Google Analytics 4: the AI Assistant channel

Since May 13, 2026, GA4's default channel group has an AI Assistant channel for visits from sources such as ChatGPT, Gemini, Deepseek, Copilot and Grok. The same page says it excludes Google's AI Overviews and AI Mode, which GA4 files under Organic Search. OpenAI's FAQ for its now-retired ChatGPT Atlas browser also stated that ChatGPT adds utm_source=chatgpt.com to referral URLs (archived October 4, 2026); no current OpenAI page repeats it. GA4 assigns the channel by the visit's referrer, not by a UTM tag.

So AI referral traffic is always undercounted. AI Overviews clicks arrive as plain Google organic traffic, a memory answer with no link sends no click at all, and any click that arrives without a referrer, as some app clicks do, lands in Direct.

Read the channel as a floor, and never compare platforms by it: a platform whose clicks keep their referrer looks bigger than one whose clicks lose it.

The leading indicator: crawler access

Access is the only GEO metric with a documented cause and effect. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, and its help center makes eligibility conditional on your CDN allowing OpenAI's published IP addresses too. Track it as a table, one row per bot:

Any red cell explains a drop in the answer metrics before you look anywhere else. Our GEO checklist lists the access checks in order.

What no GEO metric can measure

Some numbers sold as GEO metrics have nothing behind them. Do not pay for them.

• Your rank inside an AI answer. Google states that third-party tools don't have access to our internal ranking data, on a page that names "AEO" and "GEO" tools. OpenAI says only that it ranks results "using multiple factors" and that placement is not guaranteed. There is no ground truth to check a position against.
• The real queries. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries before it searches. Bing's grounding queries are the only first-party view of what was actually run.
• Which mode a buyer got. ChatGPT decides per question whether to search, and Google says AI Overviews often don't trigger at all.
• One cross-platform truth. Google says AI Overviews and AI Mode "may use different models and techniques." Any single score across platforms is a model someone chose, not a measurement.

Report every number with its platform and mode

A GEO report should let a reader see where each number came from. Use one row per platform and mode:

Two rules make the table honest. Report counts next to percentages, because 2 of 3 and 200 of 300 are not the same evidence. And skip queries where a platform did not answer, or where Google showed no Overview, rather than scoring them as misses.

How the 99Visibility score combines them (our model)

Our visibility score is a weighted average of Mention Rate 40%, Accuracy 30%, Sentiment 15% and Citation Rate 15%, multiplied by 0.2 + 0.8 × coverage. Coverage is the market-weighted share of the audited platforms that mentioned you at least once, with weights of ChatGPT 36%, Google AI Overviews 18%, Perplexity 18%, Google AI Mode 10%, Gemini 9% and Claude 9%. A brand no platform mentions scores 0.

It is our model, not an industry measurement, and we state how each platform is read: ChatGPT and Claude through their APIs with web search on, each answer labelled as searched or from memory; Perplexity with web search on; Gemini with Google Search grounding (without it when grounding fails); and AI Overviews and AI Mode from Google's live results. Our what is GEO lesson covers the basics behind it.

AI share of voice: how to compare yourself with competitors

"AI share of voice" means two different things depending on who uses it. Some tools mean mention rate. The useful definition for comparison is:

AI share of voice = your mentions ÷ mentions of you plus every tracked competitor, across the same answers.

Step 1: Find your AI competitors

Run your category prompts and list every brand the answers name. These are your AI competitors, and they are often not the companies your sales team names.

Step 2: Score each competitor the same way

Mention rate, citation rate and sentiment, on the same prompts, platforms and modes. For accuracy, check what AI says about them against their own published pricing and About pages.

Step 3: Find the gaps

• Platform gaps: a competitor named on ChatGPT but absent on Perplexity, or the reverse.
• Query gaps: high-intent questions where no brand is named consistently.
• Accuracy gaps: AI states a competitor's facts wrong. Publish an accurate comparison page; do not repeat the error.
• Access gaps: robots.txt is public, so read your competitors'. One that blocks PerplexityBot, or loads its prices by JavaScript, cannot be read where you can.

Step 4: Prioritize by intent

Rank the gaps by how close the question sits to a purchase, then by how many runs show the gap. A gap that shows in one run of three is noise until it repeats.

Read share of voice with care

If a few large brands fill your category answers, compare yourself with brands of a similar size, or the metric reports their scale, not your work. Share of voice can also fall while your own mentions rise, when a new competitor enters the answers.

Microsoft frames its own Citation Share the same way, as an observational metric, not a ranking system or a competitive scoreboard.

Paid 99Visibility plans track one to three competitors alongside you; see competitor comparison.

GEO ROI: how to estimate it without inventing numbers

Be skeptical of any GEO ROI figure that arrives without a method. A critical survey of 45 GEO studies found no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability. Conversion multiples quoted for "AI visitors" come from whatever traffic happened to be tagged, which is not a random sample.

The floor calculation

1. Attributed sessions: GA4 AI Assistant channel sessions for the period.
2. Attributed value: those sessions × your conversion rate for that channel × average customer value.
3. Cost: tool fees plus the hours spent on fixes, at your loaded hourly rate.
4. ROI: (attributed value − cost) ÷ cost.

This is a floor, not an estimate. It misses AI Overviews clicks (counted as organic search), untagged clicks, and every answer that named you without a link. Add two softer signals beside it, clearly labeled: signups whose "How did you hear about us?" answer names an AI assistant, and the trend in branded queries in Search Console.

The test that holds up: date your fixes

The most defensible ROI evidence is a before-and-after on a dated fix. If you unblocked OAI-SearchBot on a known day, compare ChatGPT search mentions, citations and AI Assistant sessions before and after, on that platform only. Note anything else that changed in the window, such as a model release, and say so in the report.

Start for free with 99Visibility. See the current pricing for paid-plan details.

How often to measure

Weekly for a short set of the prompts that matter most, monthly for the full prompt set and competitor comparison, and immediately after a price change or a new model release. Keep the prompts, platforms and modes fixed between runs. Our AI visibility monitoring routine fits the weekly part into 15 minutes, and the AI visibility tracking template gives you the log.

For a single platform, the step-by-step check in what does ChatGPT say about my brand shows how to collect the answers by hand.

Frequently asked questions

How is GEO measured?

Two ways at once. Ask a fixed set of prompts on each AI platform and record whether you are named, whether the facts are right and whether you are linked. Then read the first-party data: Search Console's AI impressions, Bing's AI citations, GA4's AI Assistant channel and your server logs for AI crawler access.

What are the most important GEO KPIs?

Accuracy rate and crawler access, because they are the two you can verify and fix directly. Mention rate and citation rate per platform and mode come next. Treat AI referral sessions as a floor, since AI Overviews clicks count as organic search and many AI mentions carry no link.

What is AI share of voice?

It is your brand's share of all brand mentions across a fixed set of AI answers, compared with the competitors you track. Calculate it as your mentions divided by your mentions plus your competitors' mentions. It favors large brands, so compare yourself with brands of similar size.

How do I calculate the ROI of GEO?

Multiply GA4 AI Assistant sessions by that channel's conversion rate and your average customer value, subtract tool and labor costs, and divide by cost. Treat the result as a floor, because AI Overviews clicks and unlinked mentions are not counted. For stronger evidence, compare metrics before and after a dated fix.

Can I track my brand's position in ChatGPT or AI Overviews?

Not meaningfully. Google says third-party tools have no access to its internal ranking data, OpenAI says placement is not guaranteed, and the order of brands in a generated answer shifts between runs. Track whether you are mentioned, cited and described correctly instead.

To get mention, accuracy, sentiment and citation rates per platform, with each AI's mode disclosed, run a 99Visibility audit.