AI visibility tracking is the practice of measuring how often, how prominently, and how favorably your brand appears in answers from ChatGPT, Gemini, Perplexity, Copilot, and Google's AI Overviews. Unlike traditional rank tracking, there is no fixed position to check and no impression count the platforms hand you by default, so you build the measurement yourself: a panel of buyer questions, a sampling method, and a small set of metrics tracked over time. This guide gives you that framework for measuring generative engine optimization performance, plus the analytics and log-reading techniques that turn AI visibility from a guess into a number you can report.
Why AI Visibility Tracking Is Different From Rank Tracking
Traditional SEO tracking rests on a stable object: a ranked list of ten links for a given query, checked from a neutral location, refreshed on a predictable schedule. AI answers break that model in three ways. They are generative, not retrieved, so two identical prompts run seconds apart can return different wording, different sources, and occasionally a different verdict. Most consumer AI apps also personalize on account history, memory, and location, so the answer your CMO sees is not the answer a fresh browser session sees.
The third break is the absence of impression data. Google Search Console has told site owners how many times a page appeared for a query since 2015. No AI platform offered anything similar until 2026, when Google added a generative AI performance report to Search Console covering AI Overviews and AI Mode (more in our guide to getting cited in Google AI Overviews and AI Mode). Even that report gives impressions only, with no clicks, no click-through rate, and no query detail.
ChatGPT, Claude, Perplexity, and Copilot publish no comparable dashboard for site owners at all. Until they do, AI visibility tracking is something you build, not something you are handed.
How to Build a Prompt Panel for AI Visibility
A prompt panel is the fixed set of questions you run, on a schedule, through each AI platform you care about. It is the single most important design decision in the whole exercise, because a sloppy panel produces noisy numbers no matter how carefully you analyze them afterward.
Cover Every Funnel Stage, Not Just Your Brand Name
Prompts that name your brand directly ("Is Agentixly a good software house?") test something narrow: whether the model already has an opinion about you. Category prompts ("best software development company in Israel for a Series B startup") test the question that actually predicts revenue: whether you show up when nobody asked for you by name. A useful panel weights heavily toward category and comparison prompts, with only a handful of direct brand prompts.
Write Prompts the Way Buyers Actually Type Them
Skip keyword-stuffed phrasing like "top cybersecurity vendor 2026 best price." Real buyers write specific, conversational questions with context: budget, industry, team size, urgency. Specific prompts also produce more stable, comparable answers across runs than generic ones, because they narrow the model's retrieval to a smaller, more consistent set of sources.
Build the panel in this order:
- List four to six buyer personas relevant to your business, defined by role and company stage, not by demographics.
- For each persona, write five to ten prompts spanning awareness ("what is X"), consideration ("best X for Y"), and decision ("X vs Y") stages.
- Add three to five head-to-head prompts naming your two or three closest competitors alongside you.
- Add a small number of direct brand prompts to track reputation and factual accuracy separately from discovery.
- Freeze the panel for a full measurement cycle before you touch it again, so period-over-period changes reflect visibility, not panel drift.
- Retire and replace 10 to 20 percent of prompts each quarter as your market and competitors change.
A panel of 40 to 100 prompts is enough for most mid-market companies to see meaningful patterns without turning measurement into a full-time job.
The Metrics That Define AI Share of Voice
Once you have a panel and a set of platforms, decide what you are actually counting. These six metrics cover the questions stakeholders ask most often:
| Metric | What It Measures | How to Calculate It | |---|---|---| | Mention rate | How often your brand appears anywhere in the answer | Runs where your brand appears, divided by total runs | | Citation rate | How often your brand appears as a clickable source, not just a word | Runs with a live link to your domain, divided by total runs | | Prominence | Whether you are the first source or buried in a list | Average position among all sources cited, first mention ranks highest | | Sentiment | Whether the mention reads as neutral, favorable, or negative | Manual or model-assisted read of the surrounding sentence, on a three-point scale | | Share of voice | Your visibility relative to named competitors | Your mention count divided by the combined mention count of you and your tracked competitors | | Cited source domains | Which sites the model actually pulled facts from | A tally of every domain that appears as a source across the full panel |
Illustrative scenario: assume you run a 50-prompt panel about penetration testing vendors through ChatGPT, three times each, for 150 total runs. Your brand appears in 60 of those runs, a 40 percent mention rate, and carries a live citation link in 24 of them, a 16 percent citation rate. Two named competitors appear a combined 90 times across the same runs. Your share of voice is 60 divided by 150 (60 plus 90), or 40 percent.
Chart mention rate, citation rate, and share of voice against the same competitor set every cycle so the trend line means something.
How to Sample AI Answers Without Fooling Yourself
A single query proves nothing. Build sampling discipline into the process or the resulting numbers will bounce around for reasons that have nothing to do with your actual visibility.
Run each prompt multiple times per cycle, not once, since model responses vary run to run even with identical input. Where a platform's product supports it, test from more than one account state (logged out, and a fresh account with no history), since personalization from memory and past conversations skews what gets surfaced. If your budget allows, sample from more than one geography, since sources and even language can shift with location signals.
Most GEO teams script their panel against each provider's API (the OpenAI API, the Anthropic API, Perplexity's API, and Google's Gemini API) rather than clicking through consumer apps by hand, because it is faster and repeatable. Treat API results as a directional proxy, not a mirror. Consumer apps like the ChatGPT app or Google's AI Overviews apply web grounding, memory, and UI-level source curation that the raw API does not always replicate, and Google AI Overviews specifically has no public API at all. Spot-check the consumer surfaces by hand each cycle to confirm the API panel still reflects what real users see.
A weekly cadence suits volatile, competitive categories; monthly is enough for stable, less contested ones. Keep the interval fixed so period-over-period comparisons are fair.
How to Track AI Referral Traffic in Google Analytics
Prompt panels tell you whether you are being cited. Referral data tells you whether anyone actually clicked through afterward, and it is the closer of the two metrics to revenue.
GA4 added a native AI Assistant channel to its Default Channel Group, which automatically tags a session with the medium ai-assistant when the referrer matches a recognized assistant such as ChatGPT, Gemini, Deepseek, Copilot, or Grok, keeping that traffic separate from Organic Search and Referral. Google's own AI Overviews and AI Mode clicks are not included in this channel; they are counted under Organic Search instead. Claude and Perplexity are not part of the default recognized list either, so if you want full coverage of the platforms your prompt panel tracks, build a custom channel group with your own regex.
Create that custom channel above Referral in your channel group priority order, with a condition that matches session source against your own AI hostnames:
chatgpt\.com|chat\.openai\.com|claude\.ai|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com
This pulls Claude and Perplexity sessions out of generic Referral so you can report on them by name, without disturbing how the built-in AI Assistant channel or Organic Search handle the platforms they already own. Google documents the default behavior and the custom channel group setup directly, alongside the AI Assistant channel definition itself. Treat every number this produces as a floor, not a ceiling. AI apps frequently strip or rewrite referrer headers, especially in mobile in-app browsers, so a portion of real AI-driven visits will always land in GA4 as direct traffic with no way to reattribute them after the fact.
How to Read Your Server Logs for AI Crawler Activity
Referral traffic proves a human clicked through. Server logs prove an AI system could read your page at all, which is the leading indicator: you cannot be cited from a page a crawler never fetched.
Each major provider runs more than one crawler, and the distinction matters. An indexing crawler builds the corpus a platform's search feature draws on and generally respects robots.txt. A live-fetch crawler visits a specific page in response to a specific user question and, because a person triggered it, may ignore robots.txt entirely. A training crawler feeds the model's underlying knowledge and has no direct connection to a single citation.
Names and behavior change, so verify against the source before you edit robots.txt: OpenAI's crawler overview, Anthropic's crawler policy, Perplexity's bot documentation, and Google's crawler list each publish current names and IP ranges for verification.
| Platform | Crawler | Purpose | Respects robots.txt | |---|---|---|---| | OpenAI | GPTBot | Collects content to train future models | Yes | | OpenAI | OAI-SearchBot | Indexes pages for ChatGPT's search features | Yes | | OpenAI | ChatGPT-User | Fetches a page live when a user's prompt requires it | User-initiated, may not apply | | Anthropic | ClaudeBot | Collects web content to train Claude | Yes | | Anthropic | Claude-SearchBot | Indexes pages to improve Claude's search results | Yes | | Anthropic | Claude-User | Fetches a page live when a user's prompt requires it | User-initiated, may not apply | | Perplexity | PerplexityBot | Indexes pages to surface in Perplexity answers | Yes | | Perplexity | Perplexity-User | Fetches a page live when a user's prompt requires it | User-initiated, typically ignores it | | Google | Google-Extended | Governs use of crawled content for Gemini training and grounding | Yes, as a separate token |
Google-Extended is worth a specific callout: Google states explicitly that it carries no weight as a ranking signal in ordinary Search, so blocking or allowing it will not move your organic rankings either way.
Pull the relevant lines out of your access logs to see which pages each crawler is actually reaching:
grep -Ei "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|Claude-User|PerplexityBot|Perplexity-User" access.log \
| awk '{print $1, $7}' \
| sort | uniq -c | sort -rn | head -20
Example: a spike in ChatGPT-User hits on a single product page during the same week your mention rate for related prompts rises from 20 to 35 percent is a strong signal that page is the source being cited. Confirm it by checking the link the AI answer actually shows, then reinforce that page rather than spreading the next content sprint thin.
Choosing an AI Visibility Tracking Tool
Most teams start with the free methods above and add software once the manual process becomes the bottleneck, not before. The category is new enough that naming specific winners here would be a disservice. Instead, know the three shapes the tooling takes:
| Category | Best For | Trade-off | |---|---|---| | Manual or spreadsheet panels | Small panels, full control over prompts and scoring | Time-intensive past roughly 50 prompts run on more than one platform | | Dedicated GEO platforms | Automated multi-platform sampling, dashboards, historical trend lines | Newer vendors, uneven coverage of every model and region, ongoing cost | | Brand monitoring suites with an AI module | Teams that already track earned media and social mentions in one place | AI-answer tracking is often an add-on, not the core product, with less prompt-level control |
Whichever shape you choose, ask any vendor exactly which platforms they sample, how many runs per prompt, from which locations, and whether they use each provider's API or something closer to the consumer product. A platform's marketing page rarely volunteers these details unprompted, and they change what the resulting numbers actually mean. For a broader checklist on vetting a partner rather than just a tool, see our guide to choosing a GEO agency.
How Agentixly Approaches AI Visibility Tracking
Agentixly builds AI visibility tracking as part of a broader GEO engagement, not as a one-off report. The methodology has five phases:
- Baseline audit. We build a persona-based prompt panel, typically 60 to 100 prompts, and run it across ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews to establish your starting mention rate, citation rate, and share of voice.
- Competitor and source mapping. We document which competitors are cited alongside you and which domains the models treat as authoritative in your category, which usually points directly at content gaps worth closing.
- Analytics instrumentation. We configure GA4 channel groups and Search Console's generative AI performance report so AI-driven traffic and impressions show up in your existing reporting, not a separate tool nobody checks.
- Crawler and log review. We check server logs and
robots.txtto confirm the pages you most want cited are actually reachable by the crawlers that feed each platform. - Ongoing measurement. We re-run the panel on a fixed cadence, report the same six metrics every cycle, and tie the findings back into the content and technical roadmap.
Every engagement also looks at entity consistency across the web, since a model is more likely to cite a brand it can describe the same way twice; our guide to entity SEO covers that groundwork in depth.
Next Steps
AI visibility tracking is not a one-time audit. Model providers update retrieval, ranking, and grounding behavior continuously, and a panel measured once in the first quarter tells you nothing about the third. Build the panel, fix the sampling method, instrument GA4 and your logs, and put a number on the board every month.
If you want that framework built and run for you, alongside the content and technical work that actually moves the numbers, talk to Agentixly about a GEO engagement scoped to your category and competitors.