AI visibility isn’t measured in rankings but in frequencies. If someone offers you your “position in ChatGPT”, be wary: a model’s answer changes every time you ask. The metric that does make sense is the percentage of answers your brand appears in (your share of voice), calculated over dozens of runs, per engine and against your competitors. Here’s how to set up that measurement: which metrics matter, which free sources you have today, which paid tools exist and how to build your own system step by step.
Why measuring AI isn’t like measuring SEO
In classic SEO a keyword has a stable position. In generative engines, every answer is a sample from a probability distribution: the model picks brands and sources from a different set on every run. Three independent studies confirm it:
- SparkToro + Gumshoe (January 2026, 2,961 runs): there’s less than a 1 in 100 chance that ChatGPT or Google’s AI will repeat the same list of brands in two answers, and roughly 1 in 1,000 that they’ll repeat it in the same order.
- Ahrefs (43,000 keywords): an AI Overview has a 70% chance of changing between two observations, and lasts on average just 2.15 days.
- SE Ranking (AI Mode, 10,000 keywords three times on the same day): only 9.2% of URLs matched.
But there’s good news: even though the list shifts, how often you appear is stable. In the SparkToro study, one brand appeared in 97% of ChatGPT’s answers, even though it was the first mention in only one in three. That consistency is what you can measure.
“When a client shows me a screenshot of ChatGPT recommending them and wants to celebrate, I ask them to hold on. A screenshot proves nothing, good or bad: with less than a 1% chance of the answer repeating, that capture is a snapshot of a moment that won’t happen again. The only question that counts is: out of 100 times someone asks this, how many times do you show up, and how many times do your competitors?” – Rafa Villaplana, SEO consultant and CEO of Linking & Growing
The metrics that really matter
| Metric | What it measures | Reliability |
|---|---|---|
| Presence rate / visibility | % of answers you appear in | High, with enough runs |
| Share of voice | Your weight against a fixed set of competitors | High; the most actionable |
| Citation share | Whether your domain appears as a linked source | Medium: citations rotate a lot |
| Average position in the answer | Where in the list you appear | Low: only use it aggregated |
| Sentiment | How the model describes you | Medium: review it manually |
A key nuance: a mention and a citation aren’t the same thing. Seer Interactive (541,213 answers) documented “ghost citations”: cases where the model uses your content as a source but recommends a competitor. Measure them separately, and optimise to be mentioned, not just linked.
And be careful with small changes: below roughly 5–7 percentage points, the difference is usually statistical noise, not a trend.
The official, free sources
Google Search Console: generative AI report. Available for all websites since 31 August 2026. It shows impressions in AI Overviews and AI Mode, by page, country and device. Its limits: no clicks, no CTR, no queries, and it doesn’t separate AI Overviews from AI Mode. Use it by cross-referencing AI impressions with clicks from the main performance report at URL level, so you can see which pages get a lot of AI exposure but lose clicks (remember that the presence of an AI Overview is associated with a 58% lower CTR in position one).
Bing Webmaster Tools: AI Performance. The only official source that gives you queries (the “grounding queries”). It shows citations in Copilot and Bing’s summaries, it’s essential for measuring Copilot, and it’s free.
Google Analytics 4: the “AI Assistant” channel. Since May 2026, GA4 has a native channel that captures ChatGPT, Gemini, Copilot and others. Two important caveats:
- Traffic from AI Overviews and AI Mode is NOT here: it counts as Organic Search and can’t be told apart from organic.
- Perplexity may not be on the official list, so check whether it falls under Referral.
It’s worth creating a custom “AI” channel group with a regex condition on the source: chatgpt\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|deepseek\.com.
October 2026 notice: after the GPT-6 rollout, many ChatGPT links stopped including their source parameter, so GA4 classifies those visits as “Direct” instead of AI traffic. OpenAI has confirmed it and it may be reversed. If you see a sharp drop in ChatGPT referrals these months, it’s probably this rather than a real loss of visibility, so add an annotation in GA4.
Paid tools
Almost all of them work the same way: they run prompts against LLMs on a schedule, detect mentions, citations and sentiment, and aggregate the results. They differ in where the prompts come from, which engines they cover and how much statistical transparency they offer. By category:
- Indexes inside SEO suites (Ahrefs Brand Radar, Semrush AI Visibility, SE Ranking): a market benchmark without writing prompts, which you can cross-reference with search demand.
- Dedicated prompt trackers (Peec AI, Otterly, GEOMetrics, LLMPulse): your own prompts, run daily.
- Enterprise platforms (Profound, Scrunch): more engines and analysis of real conversations.
- Explicitly statistical measurement (Gumshoe): simulates conversations and publishes its margins of error.
Before signing up for any of them, ask these four questions: How many times do you run each prompt? Do you use the API or the user interface? Can you set the country and language? Do you publish confidence intervals? If they can’t answer, you’re paying for one-off screenshots dressed up as data.
How to build your own system (6 steps)
1. Prompt library (50–200 to start). Organise it as a matrix of topic cluster × funnel stage: discovery (“what is…”), consideration (“best SEO agencies in Valencia”, “alternatives to X”), decision (“X vs Y”, “X pricing”). Include branded prompts (how it describes you) and unbranded ones (your real visibility, which is what matters). Write them in the language and local variant your customers use, and at the length a real user would. Fix a competitive set of 3–8 brands and don’t change it.
2. How many runs. For conclusions per prompt, at least 60–100 runs (±10 pp margin); around 400 for ±5 pp. For a global KPI per cluster or engine, you can aggregate.
3. Frequency. Full batches weekly or every two weeks, comparing 4-week periods. A single-day monthly measurement is noise.
4. Control the variables. Same location, same language, and record the mode (with or without web search) and the model. And always segment by engine: ChatGPT, Gemini, Perplexity, AI Overviews, AI Mode and Copilot behave differently (AI Mode and AI Overviews share only 13.7% of their citations).
5. Logging schema. One row per answer: date, engine, model, prompt, cluster, answer text, brands mentioned (in order), whether you appear and in what position, sentiment, URLs cited and whether any of them are yours.
6. Read the trend. Work with trends and confidence intervals. A change is only “real” if it exceeds the margin of error of both periods. Always compare the gap with your competitors, not just your own number.
“You can build this with a spreadsheet and scripts against the APIs, and in fact I recommend starting that way to understand the mechanics before paying for a tool. It’s exactly the problem that led me to build my own measurement platform: I needed to run the same prompts dozens of times, segment by engine and for the Spanish market, and see the real trend against competitors, not a pretty screenshot. Measure first, invest later.” – Rafa Villaplana, SEO consultant and CEO of Linking & Growing
Measuring real traffic
With the custom GA4 channel, measure conversions by channel, landing pages and engagement rate. Keep it in context: AI traffic is small but tends to convert better. Adobe found it converted 42% better than other traffic in retail (March 2026); on Ahrefs’ own site it was 0.5% of traffic but 12% of sign-ups. But it’s not a universal law: a study in Marketing Science covering 973 ecommerce sites found that ChatGPT converted worse than organic. Measure it on your own site, and add a self-reported attribution question (“how did you hear about us?”) to your forms: it’s the best defence against the traffic GA4 can’t see.
For scale, in Spain: AI brought in 0.30% of visits between January and April 2026 (ChatGPT 70%, Gemini 12%, Perplexity 12%), while organic fell from 52% to 40.7% of traffic (SE Ranking). Low volume, but the trend is clear.
Common measurement mistakes
- Trusting a single screenshot. With <1% repeatability, a screenshot proves nothing.
- Not controlling location or language. A prompt run from the US in English doesn’t represent a user in another market.
- Confusing “the model knows me” with “the model cites me”. Without web search, it answers from its training (entity recognition); with search, it retrieves and cites. They’re different metrics.
- Not segmenting by engine. A global average hides opposite behaviours.
- Adding Search Console’s AI impressions to the total. They’re already included in the “Web” search type.
- Treating ±3-point changes as trends. That’s noise.
Frequently asked questions
Can I find out my “position” in ChatGPT?
Not reliably, because model answers change on every run. What you can measure is your visibility rate: how many of many answers you appear in. Any tool that gives you a fixed “position” is showing you a one-off snapshot.
How many times do you need to run a prompt to get reliable data?
Between 60 and 100 runs per prompt and engine for a ±10-point margin of error, and around 400 for ±5 points. A single query has no statistical value.
Does Search Console tell me how many clicks I get from Google’s AI?
No. Its generative AI report only shows impressions in AI Overviews and AI Mode, not clicks. On top of that, click traffic from Google’s AI is counted as organic in GA4, indistinguishable from the rest.
How do I see ChatGPT or Perplexity traffic in GA4?
GA4 has an “AI Assistant” channel that captures ChatGPT, Gemini and Copilot (but not AI Overviews or AI Mode, which count as organic). Perplexity is worth checking. The most robust approach is to create a custom channel group with a regex for the AI domains.
Is it worth paying for an AI tracking tool?
It depends on the size of your business. You can start for free with a spreadsheet and the official sources. If you sign up for a tool, insist that it tells you how many runs it does per prompt and that it publishes confidence intervals; otherwise you’re paying for screenshots with no rigour.