Measuring your AI visibility: share-of-voice tracking across engines

Why "Rank" Doesn't Apply to AI Answers

Google rank tracking counts a position: your page is #3 for a keyword, or it isn't. That model assumes a stable, ordered list you can screenshot and diff day over day. AI answer engines don't produce that list. ChatGPT, Perplexity, Claude and Google AI Overviews each synthesize a single answer per query, assembled from whichever retrieved passages the model judged most relevant at that moment, and the same prompt run twice can retrieve different sources, especially on engines with continuous crawling like Perplexity.

The consequence: a single query, run once, tells you almost nothing. It's one sample from a probabilistic process. The metric that survives this is share-of-voice: not "am I #1," but "across a representative panel of prompts my buyers actually ask, in what share of runs do I appear at all, and how does that compare to my named competitors, tracked over enough repetitions and enough time to be a trend, not a coin flip."

What Share-of-Voice Actually Measures

Share-of-voice is a ratio, but the numerator hides three distinct events that get conflated if you don't separate them:

1. Mention

Your brand name appears somewhere in the answer text, with or without a link. This is the loosest signal, it means the model's training data or retrieved context contains your brand, but says nothing about whether the user could click through to you.

2. Citation

The answer includes a source reference (a footnote, an inline link, a "Sources" panel entry) pointing at your domain. This is the commercially meaningful event: it's the one that can produce a click. Perplexity, Google AI Overviews and Bing-backed ChatGPT answers expose citations structurally; you can parse them. Claude's citation behavior is more conversational and less consistently structured, which is itself a measurement finding worth logging.

3. Sentiment / framing

When you are mentioned or cited, is the framing accurate and favorable, neutral, or wrong? A brand can have high share-of-voice and still lose deals if the engine consistently describes it with an outdated price, a discontinued feature, or a direct "however, X is better for most users" qualifier. Track this qualitatively, it's the field a pure-numeric dashboard misses.

Report all three separately. A brand with a 40% mention rate but a 5% citation rate has a visibility problem in the model's synthesis, not its knowledge, different fix than a brand with low mentions altogether.

Building a Measurement Methodology

The methodology is the whole game, a good panel run inconsistently is worse than a smaller panel run rigorously. Four decisions determine whether your numbers mean anything.

The prompt panel

Build 20-50 prompts that mirror how real buyers actually ask, not your target keywords verbatim. Include category questions ("best tools for X"), comparison questions ("X vs Y"), and problem-first questions ("how do I do X"). A panel built from keyword-research habits will overweight phrasing no one uses conversationally and undercount your real exposure.

The competitor set

Fix a named list of 3-8 direct competitors before you start measuring, and don't change it mid-quarter, swapping competitors resets your baseline and makes trend lines meaningless. Usethe ranked index or your own sales-loss reasons to pick the set that actually contests your deals.

Repeats per prompt

Run each prompt multiple times per measurement window (5-10 runs is a reasonable floor) because engine output varies run to run, especially on Perplexity and ChatGPT. A single run per prompt per week will show you noise, not a trend, it can swing 20 points from a genuinely stable underlying rate purely from sampling variance.

Cadence

Weekly is the practical floor for a fast-moving competitive set; monthly is enough for a mature category with slow content churn. Match cadence to how often the underlying content (yours and competitors') actually changes. Measuring daily on a panel that never moves just spends budget on noise.

Per-Engine Measurement Notes

Each engine retrieves from a different source and exposes citations differently, so the same panel needs engine-specific handling, not one script that assumes a shared response format. See the companion guide on ranking mechanicsfor how each engine chooses sources in the first place, this section covers how to measure the outcome.

ChatGPT

Bing-backed browsing responses include structured citations you can parse programmatically; plain (non-browsing) responses draw on training data and won't reflect anything published after the model's cutoff, which will understate your true share-of-voice if you only sample the non-browsing mode. Sample both modes and report them separately.

Perplexity

Every answer cites sources by default, which makes Perplexity the easiest engine to measure precisely, but its continuous crawl also means the highest run-to-run variance of the four, because a page published an hour ago can outrank one published last year. Budget more repeats per prompt here than on any other engine.

Claude

Without live web access in most consumer contexts, Claude draws on training-time knowledge and is the slowest engine to reflect new content, measured share-of-voice here changes over months, not weeks. Treat a flat Claude trend line as expected, not a tracking failure; look for step changes around model updates instead of week-to-week movement.

Google AI Overviews

Overviews draw on Google's existing index, so share-of-voice here correlates with (but isn't identical to) your organic rank; it's the one engine where a Search Console query-level view can cross-check your AI panel. Discrepancies between the two are themselves a finding: content that ranks in Search but doesn't get pulled into an Overview usually has an answer-first structure problem, not a relevance problem.

Reading the Data Correctly

A share-of-voice dashboard is only useful if you resist three common misreadings.

• One bad week isn't a trend. Given run-to-run variance, treat single-window drops under ~10 points as noise until a second consecutive window confirms it.
• A competitor's spike often has a dated cause. Check for a launch, a press cycle, or fresh content from them before assuming the model's weighting changed generally.
• Citation share and traffic aren't the same curve. A citation can sit below the fold of a long AI answer and get almost no clicks; correlate citation-share trend with actual referral traffic (most AI engines now send a small but growing amount) before treating a citation win as a business outcome.

Tools like Rankscale andNightwatch plot the trend line for you; the reading discipline above is what keeps you from acting on a single noisy data point either way.

Implementation Checklist

Tick items off as you set this up, this page remembers your progress on this device.

Frequently Asked Questions