To measure AI citations, fix a list of the questions your buyers ask, run each one several times on every AI engine that matters to you, and log every answer: whether your brand was named, whether your site was linked, and who appeared instead. From that log you report two numbers: your appearance rate (the share of answers that name you) and your share of AI voice (your share of all brand mentions across your competitive set). Then add the first-party data the platforms now give you for free: Google Search Console’s generative AI report, Bing Webmaster Tools’ AI Performance report, and AI referral traffic in GA4.
This page is the measurement method. For tool prices and features, see Best AEO Tools in 2026. For turning the numbers into a business case, see The ROI of GEO. For a one-hour first pass you can do today, use the AI citation audit checklist.
Why your normal analytics miss AI citations
When someone asks ChatGPT “what’s the best accounting software for a small agency?” and it recommends Brand X, Brand X’s analytics record nothing unless the person clicks a link. No impression, no keyword position, no session. The recommendation happened, and it may have shaped the purchase, but it left no trace in a standard dashboard.
The audience is large: OpenAI said in October 2026 that more than 1.2 billion people use ChatGPT each week. And where AI answers appear in Google, they take clicks from the links below them (see AI Overviews Cut Clicks by 40% Where They Appear). So measurement has two halves: sampling the answers yourself, the only way to see what engines say about you and your competitors, and reading the first-party data from Google, Microsoft and your own analytics.
Mentions, citations and sources: get the terms straight
Most confusion in AI visibility reporting starts with three different things all being called “citations.” Log them separately.
| Term | What it means | Example |
|---|---|---|
| Mention | The answer names your brand in its text | ”Brand X and Brand Y are popular choices…” |
| Citation | The answer links to a page on your domain as a source | A footnote or source card pointing to your pricing page |
| Source | Any page the engine retrieved or cited, whoever owns it | A review site or directory listing that names you |
They move independently. You can be mentioned without being cited (the engine names you but links to a review site), or cited without being mentioned (your guide supports a generic answer that never names you).
The metrics to track
Appearance rate
The share of answers that mention your brand.
Appearance rate = answers that mention you ÷ total answers sampled
Run 50 prompts twice each on one engine and you have 100 answers; named in 40, your appearance rate is 40%. Track it per engine and per prompt group, because a gap on one engine points to a specific cause.
Share of AI voice
Your share of all brand mentions across the competitors you track.
Share of AI voice = your mentions ÷ mentions of all tracked brands (yours included)
The two are easy to confuse. Here are the same 100 answers with three tracked competitors:
| Brand | Answers that name it | Appearance rate | Share of AI voice |
|---|---|---|---|
| You | 40 | 40% | 40 ÷ 150 = 26.7% |
| Competitor A | 55 | 55% | 55 ÷ 150 = 36.7% |
| Competitor B | 35 | 35% | 35 ÷ 150 = 23.3% |
| Competitor C | 20 | 20% | 20 ÷ 150 = 13.3% |
| Total | 150 mentions | — | 100% |
Appearance rates don’t add up to 100% because one answer can name several brands. Shares of voice always do. If you tracked only Competitor A, your share would be 40 ÷ 95 = 42.1% from exactly the same answers. That’s the catch: share of voice depends on which competitors you choose to track, so fix the competitor list at baseline and state it next to the number.
Count each brand at most once per answer, and decide up front what counts as a mention. Bing applies the same idea at the URL level: its “Citation Share” is your site’s percentage of all citations shown across all sites for a given grounding query (Bing, June 16, 2026).
Citation rate
The share of answers that link to your domain, the metric that can produce traffic. Log which URLs get cited; important pages that never appear are candidates for a rewrite.
Prominence
Being named first is different from being the seventh item in a list. The GEO paper by Princeton and IIT Delhi researchers (Aggarwal et al., KDD 2024) measured visibility with a “position-adjusted word count,” which weights sentences earlier in an answer more heavily. Log whether you were named first, in the middle or last, but don’t report an “average rank”: as the next section shows, order is the least stable part of an AI answer.
Accuracy and framing
Label each answer about you correct, outdated, incomplete, inaccurate or missing, and note how you’re described (“the budget option”). A rising appearance rate with wrong prices attached is not progress.
Sources cited instead of you
For every answer where competitors appear and you don’t, log the cited domains: competitor pages, review sites, directories, publications, forums. That list is where the improvement work comes from; the method is in How to Do Competitor Analysis for AI Citations.
Log sources per engine, because engines lean on different kinds of sources. In Yext’s October 2025 study of 6.8 million citations for location-based queries in four industries (collected July–August 2025), Gemini drew 52.15% of its citations from brand-owned websites, while OpenAI drew 48.73% from listings, mostly Google’s. That’s one study of local queries, but it shows how a brand can be strong on one engine and missing on another for structural reasons.
Why one run isn’t enough
AI answers change between runs of the same prompt. In a study published in January 2026, SparkToro and Gumshoe had 600 volunteers run 12 recommendation prompts through ChatGPT, Claude and Google’s AI Overviews or AI Mode, 2,961 times in total, during November and December 2025. ChatGPT and Google’s AI returned the same list of brands less than 1 time in 100, and the same list in the same order about once in 1,000 runs. How often a brand appeared was far steadier: City of Hope appeared in 69 of 71 ChatGPT answers but was named first in only 25. Rand Fishkin concluded that visibility percentage across many prompts and runs is a reasonable metric, while ranking position is not. The study isn’t peer-reviewed, and his co-researcher works for an AI tracking company.
Two practical consequences follow.
Run each prompt more than once. The study averaged about 80 runs per prompt per tool, Fishkin suggests asking 60–100 times, and he says the number of runs needed for statistical soundness is still an open question. For routine tracking, three to five runs per prompt per engine is a workable floor. Adding distinct prompts helps more than repeating the same prompt, because repeats of one prompt are correlated with each other.
Know your margin of error. A simple binomial estimate shows the scale. At a 40% appearance rate:
| Answers sampled | Approximate 95% margin of error |
|---|---|
| 50 | ±13.6 points |
| 150 | ±7.8 points |
| 500 | ±4.3 points |
The calculation is 1.96 × √(0.4 × 0.6 ÷ n). Because repeated runs of one prompt aren’t independent, the real uncertainty is somewhat wider. In practice, a move from 38% to 44% on 150 answers is not yet a trend. Look for changes that hold across two or three measurement cycles.
Build a prompt set you can re-run
Write prompts the way buyers ask. Mix these groups and track each separately:
- Branded: “What is [brand]?”, “Is [brand] good for [use case]?”, “[brand] pricing”
- Category: “Best [category] for [type of customer]”, “Top [category] providers in [market]”
- Problem: “How do I [problem you solve]?”
- Comparison: “[brand] vs [competitor]”, “Alternatives to [competitor]”
- Attribute or location filters, where they matter: “[category] with [feature]”, “[category] near [city]”
Twenty to fifty prompts covers most single-brand programs; split larger sets by product line or market.
Lock the list before the first run. Don’t swap prompts because you dislike the answers; the fixed list is what makes Day 30 comparable with Day 0. If you must change it, re-baseline.
Control the session. In ChatGPT, start a temporary chat and choose Unpersonalized before the first message; OpenAI says that mode doesn’t use memory, custom instructions or plugins. Use logged-out or fresh sessions elsewhere. Keep location constant too: Yext reports that models from Google, OpenAI and Perplexity confirm they use a person’s location when answering questions about businesses and services, so a run from another city is a different test.
Record non-answers as data. If Google shows no AI Overview, log “no AI Overview,” not a miss. If an engine declines to recommend businesses, log that too.
What to record for each answer
One row per prompt, per engine, per run:
| Field | What to record |
|---|---|
| Prompt, group, engine, date, run number | The identifiers |
| Mentioned (Y/N) | Is your brand named in the answer text? |
| Position | First, middle or last mention; blank if not named |
| Cited URL | Which of your pages was linked, if any |
| Competitors named | Every tracked brand in the answer |
| Domains cited | Every source domain shown |
| Accuracy | Correct, outdated, incomplete, inaccurate or missing |
| Answer type | Recommendation list, comparison, single-brand description, refusal, no AI answer shown |
| Snippet | The sentence where you appear, copied exactly |
A spreadsheet is enough to start. Fifty prompts on five engines with three runs each is 750 rows per cycle, which is about where most teams start paying for a tool.
Consumer apps, APIs and tracking tools
There are three ways to collect answers, and they don’t return identical results.
- By hand in the consumer apps. Closest to what buyers see. Slow, and hard to repeat at volume.
- Through the engines’ APIs. OpenAI’s Responses API, for example, has a web search tool that returns cited URLs plus the full list of sources consulted (OpenAI). But search only runs when the model decides to unless you require it, and API answers can differ from the app; Yext lists this as a limitation of its own API-collected study.
- Through an AI visibility tool. Most automate one or both of the above. Check which engines are included at your price, how many prompts run and how often, and whether you can export raw answers. Options and prices are compared in Best AEO Tools in 2026.
Whichever you use, keep the method fixed between cycles, or your numbers will move for reasons that have nothing to do with your visibility.
Free first-party data you should be reading
Google Search Console: the generative AI performance report
On June 3, 2026, Google launched Search Generative AI performance reports in Search Console, and its help page says they reached all websites worldwide as of August 31, 2026. The Search report shows how often URLs from your site appeared in generative AI features such as AI Overviews and AI Mode, with impressions by page, country, device and date (hourly to monthly). A separate report covers generative AI features in Discover. An impression counts each time a link to your site is shown to a user in one of these features (Search Console Help).
It shows which of your pages Google’s AI features link to and whether that’s growing, but not what the answer said, whether you were named without a link, or who appeared instead. The same AI Overviews and AI Mode data also stays inside the main Performance report under the “Web” search type (Google Search Central), so don’t add the two together.
Bing Webmaster Tools: AI Performance
Microsoft launched AI Performance in Bing Webmaster Tools as a public preview on February 10, 2026. It counts how often your pages are cited as sources in Microsoft Copilot, AI-generated summaries in Bing and select partner integrations, lists the cited pages, and shows the “grounding queries” the AI used to retrieve your content. Microsoft doesn’t name the partner integrations, so don’t read it as ChatGPT data. Bing counts citations while Google counts impressions, so the two numbers aren’t comparable.
GA4: AI referral traffic
OpenAI’s publisher FAQ said ChatGPT adds utm_source=chatgpt.com to referral URLs from its search results, so those visits are identifiable in Google Analytics (OpenAI Help Center, archived August 27, 2026; page since retired). The tag went missing for about a week after GPT-6 launched in October 2026 and was back by October 15 (Search Engine Roundtable). Since May 13, 2026, GA4 also sorts visits from recognized assistants into an AI Assistant channel in the Default Channel Group (medium ai-assistant), which Google describes as covering “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok” and which excludes AI Overviews and AI Mode (Analytics Help). Check the Referral channel for perplexity.ai and other assistant domains; if any land there, build a custom channel group that combines them with the AI Assistant channel, since default channel groups can’t be edited (Analytics Help). Then compare that channel’s engagement and key-event rate with organic search.
Connecting citations to business results
AI visibility is a means, not the outcome. Track three layers, and expect the later ones to lag:
- Visibility: appearance rate, share of AI voice and citation rate by engine and prompt group, plus Search Console impressions and Bing citations.
- Traffic and engagement: AI-assistant sessions, pages per session, key events.
- Business signals: branded search demand, direct traffic, and leads who say they found you through an AI assistant. Add an “AI assistant” option to the “How did you hear about us?” question on your forms.
Treat conversion claims about AI traffic with care. The most-quoted figure comes from Seer Interactive’s June 2025 case study of a single client site: in GA4 data from October 1, 2024 to April 30, 2025, ChatGPT referrals converted at 15.9% against 1.76% for Google organic. The same study found AI sources made up only about 0.07% of that site’s organic traffic (just under 11,000 sessions against almost 14 million). One site is a hypothesis, not a benchmark. Your own GA4 comparison is the number that counts.
Report impact conservatively: if branded search is rising across your industry, part of your lift isn’t yours. Where you can, compare pages or prompt groups you worked on against ones you didn’t. The full ROI method is in The ROI of GEO.
Reporting cadence by audience
| Audience | How often | What they need |
|---|---|---|
| The person doing the work | Weekly | Appearance rate by engine, citations won or lost, competitor movement, technical issues found and fixed |
| Marketing lead | Monthly | Appearance rate and share of AI voice by prompt group, cited pages, Search Console and Bing trends, AI referral sessions and key events |
| Executives or clients | Quarterly | Share of AI voice against named competitors, business signals, what changed and why, next quarter’s priorities |
Weekly numbers will be noisy (see the margins above). Make decisions on monthly and quarterly trends.
Common measurement mistakes
- Treating one run as a result. A single answer is one sample. Repeat prompts and judge trends.
- Reporting an average “rank” in AI answers. Order is the least stable part of an answer.
- Confusing appearance rate with share of voice. Show both, and name the competitor set behind the share.
- Only logging when you’re cited. The answers you’re missing from, and the sources cited there, are your roadmap.
- Adding Google impressions to Bing citations. They measure different things.
Want a measured baseline without building it yourself? Our free AI Visibility Snapshot runs 10 prompts on ChatGPT, Google AI Overviews and Perplexity and sends a short findings note. The Citation Audit tests 30–50 locked prompts across ChatGPT, Google AI Overviews, Perplexity, Gemini, Claude and the OpenAI Responses API, with competitor source mapping and a prioritized 60–90 day roadmap in 5 business days. It costs $1,500, fully credited toward an Implementation Sprint or retainer signed within 30 days. We don’t guarantee citations or rankings; we measure the same locked prompts the same way each time so you can judge the trend.
Request a free Snapshot or a Citation Audit →
FAQ
What is the difference between appearance rate and share of AI voice?
Appearance rate is the share of AI answers that mention you: named in 40 of 100 answers is 40%. Share of AI voice is your share of all mentions of the brands you track: if those answers contain 150 such mentions and 40 are yours, your share is 26.7%.
How many times should I run each prompt?
More than once: in SparkToro and Gumshoe’s January 2026 study, ChatGPT and Google’s AI returned the same list of brands less than 1 time in 100. Three to five runs per prompt per engine is a workable floor, and adding distinct prompts helps more than repeating the same ones.
Can I measure AI citations for free?
Partly. Search Console’s generative AI report shows your impressions in AI Overviews and AI Mode, Bing Webmaster Tools’ AI Performance report shows your citations in Copilot and Bing’s AI summaries, and GA4 can isolate AI-assistant traffic. To see what answers say and who appears instead of you, run a fixed prompt list by hand and log it in a spreadsheet.
Does Google Search Console show AI Overviews data?
Yes. Google’s generative AI performance report, available to all sites since August 31, 2026, shows impressions from AI Overviews and AI Mode by page, country, device and date. The same data is also included in the main Performance report under the Web search type, so don’t add the two together.
Is there a benchmark for a good AI appearance rate?
No reliable cross-industry benchmark has been published, and the number depends on your prompt set and competitor list. Compare yourself with your own baseline and with the competitors that appear in the same answers.
Sources
- OpenAI (October 7, 2026). GPT-6 and Intelligent UI for everyone (more than 1.2 billion weekly ChatGPT users).
- OpenAI. Web search tool documentation (inline citations, sources list, optional search).
- OpenAI Help Center. Publishers and Developers FAQ (utm_source=chatgpt.com; archived August 27, 2026, page since retired) and Temporary chat in ChatGPT.
- Google Search Central Blog (June 3, 2026). Introducing Search Generative AI performance reports in Search Console.
- Search Console Help. Generative AI performance report (Search).
- Analytics Help. Default channel group.
- Google Search Central. AI features and your website.
- Analytics Help. Custom channel groups (AI assistants example).
- Microsoft Bing (February 10, 2026). Introducing AI Performance in Bing Webmaster Tools Public Preview; (June 16, 2026) New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare.
- SparkToro (Rand Fishkin, with Gumshoe), January 2026. AIs are highly inconsistent when recommending brands or products. 600 volunteers, 12 prompts, 2,961 runs, November–December 2025.
- Aggarwal, P. et al. (2023; KDD 2024). GEO: Generative Engine Optimization. Princeton University / IIT Delhi.
- Yext Research (October 9, 2025). AI Citations, User Locations, & Query Context. 6.8 million citations, July 1–August 31, 2025.
- Seer Interactive (June 3, 2025). Case Study: 6 Learnings, 1 Site — How Traffic from ChatGPT Converts.
All sources checked November 10, 2026.
Related reading: Best AEO Tools in 2026, The ROI of GEO, How to Do Competitor Analysis for AI Citations, AI Citation Audit Checklist, The 15-Point AI Citation Readiness Checklist.