AI answers change from one run to the next, and no engine hands you a ranking report. That leaves most teams guessing: someone asks ChatGPT about the brand once, sees a good answer, and files it as success.
A single answer is an anecdote. A measurement is a fixed set of prompts, run the same way, logged in a table, and read as a trend. This guide gives you that method, the four numbers worth keeping, and a plain list of what nobody can measure yet. If you also want the fixing side, see our AI visibility audit and the AI visibility checker.
01The four numbers to track
Track four numbers: presence rate, citation share, accuracy and AI referral sessions. Between them they cover being named, being linked, being right and sending visitors. Most dashboards blend them into one score, and that blended score hides which part moved.
Adobe's training page on this makes the same point: a visibility score is an aggregate, every tool calculates it differently, and it is hard to act on alone, so it should be paired with mentions, citations, agentic traffic and referral traffic. Supermetrics lists a similar set: AI impressions, citation share, AI referral sessions and prompt-level presence rate.
| Number | How to get it | What it cannot tell you |
|---|---|---|
| Presence rate | Share of your fixed prompts where the answer names your brand | Whether the person who saw it acted on it |
| Citation share | Share of answers where your URL is linked as a source | Mentions without a link, which are common |
| Accuracy | Share of mentions where price, location, services and ownership are right | Why a wrong claim appeared, which needs a source check |
| AI referral sessions | Analytics visits from assistants such as chatgpt.com or perplexity.ai | Visits that arrive untagged, or people who see you and search your name later |
02Build a prompt panel of 20 real prompts
A prompt panel is the fixed list of questions you run every time. Build it from what buyers actually ask, not from what you want to rank for. Pull phrasing from sales calls, support tickets, onboarding forms and the questions in your own inbox, then group them by intent.
| Prompt type | How many | Example for a SaaS tool for dentists |
|---|---|---|
| Category, unbranded | 6 | What is the best scheduling software for a small dental practice? |
| Problem, unbranded | 4 | How do I cut missed appointments at my dental clinic? |
| Comparison | 4 | Compare [your brand] with [a named alternative] |
| Branded facts | 4 | What does [your brand] cost, and who owns it? |
| Reputation | 2 | Is [your brand] any good? What do reviews say? |
Keep the panel stable for at least a month, then add prompts rather than swapping them. Changing prompts changes the numbers, and you lose the trend. Write down any wording change the day you make it.
03Sample the answers properly
Run each prompt several times per engine, because the answer is not fixed. Supermetrics recommends running each prompt three to five times per engine in logged-out sessions. Do the same, record the date of the run in your log (not on your public pages), and note whether web search was on.
Pick the engines your buyers use
At minimum ChatGPT, Perplexity, Gemini, Google AI Overviews and AI Mode, and Claude. Add Copilot if your buyers use Microsoft products.
Run in a logged-out or clean session
Personal history and memory shift answers. Use a fresh browser profile or a logged-out mode so the answer is closer to what a new buyer sees.
Run each prompt three to five times
Record each answer. A brand named in one of five runs is a different position from one named in five of five.
Record four things per answer
Whether you are named, whether you are linked, where you appear in a list, and whether each claim about you is right.
Save the evidence
Screenshot or copy the answer text with the prompt, the engine and whether search was on. You will need it to show a client or colleague what changed.
04Use the analytics you already have
Search Console, Bing Webmaster Tools and your web analytics each see part of the picture. Use them for what they can see and do not stretch them past it.
Google says traffic from AI Overviews and AI Mode is included in the Search Console Performance report under the Web search type, and suggests combining it with Google Analytics to track conversions and time on site. Supermetrics notes that Search Console does not separate out clicks, click-through rate or queries for Google's AI features. Search Engine Journal reports that Google opened a pilot of reports with AI impressions and share of voice for AI Mode and AI Overviews for a limited number of US accounts. Those reports show impressions by page, country, device and date, with no click or query data.
Search Console
Watch impressions on the question pages you rewrote. Use the page filter to compare the group before and after, and keep the 16 months of data Supermetrics says Search Console retains.
Bing Webmaster Tools
Supermetrics describes an AI Performance report in Bing Webmaster Tools with a Citation Share view for Copilot. Check whether your account has it.
GA4
Look for an AI assistant channel under Traffic acquisition, or build a custom channel group that matches source names such as chatgpt, perplexity, gemini, claude and copilot, placed above the generic referral channel.
A form question
Add 'How did you hear about us?' with an option for AI assistants. It catches visitors who never arrive with a tag.
Untagged visits are the norm. Profound's analysis of more than two million US conversations found that over 97% of brand visits following a ChatGPT mention carried no UTM tag. An Arcalea client case showed that 49% of the revenue from ChatGPT-touched orders came from orders where ChatGPT was not the last touchpoint, so a last-click report would credit it to Google Ads, organic or direct. Profound sells visibility tools, and the Arcalea case is one client's first-party data, so read both as signs of undercounting, not as benchmarks.
05Why clicks alone mislead
Click data is not enough because AI answers reduce clicks while still shaping who buys. Pew Research found that people clicked a traditional result in 8% of visits when an AI summary appeared and in 15% when none did, and clicked a link inside the summary in about 1% of visits. Ahrefs found AI Overviews are linked to a 58% lower click-through rate for the top-ranking page, in desktop data.
Seer Interactive's 53-brand data adds the other half. Organic click-through on queries that showed an AI Overview rose from 1.3% to 2.4% over two months after a low point, and brand-cited pages earned roughly double the organic click-through of uncited ones across the year. So being cited matters, and a falling click count does not by itself mean falling influence.
06Choose tools without trusting one score
A tracking tool is a faster way to run your panel, not a replacement for it. The IAB's measurement framework says more than 20 companies sell AI visibility measurement, and it separates decision-grade measurement from directional measurement, with disclosure of method. Use that split: treat a single-run score as directional, and use a decision-grade standard, many prompts and repeated runs, before you change budget.
Questions to ask any AI visibility tool
0 of 8 done07What nobody can measure yet
Several things stay unmeasurable today, and a report that claims them should be questioned. You cannot see how many people were shown your brand in a ChatGPT answer. You cannot see the clicks and queries from Google's AI features separately in Search Console, per Supermetrics. You cannot prove that a mention caused a sale, because the later visit often arrives as brand search or direct traffic.
A monthly prompt panel suits you if
- Buyers ask AI for recommendations in your category
- You can spare a few hours each month
- You want a trend you can show a team
Skip heavy tooling if
- You have no ranked pages yet, so fix that first
- You expect a single score to prove revenue
- Nobody will act on the findings
08A monthly routine that fits in a morning
Baseline
Build the 20-prompt panel, run it three to five times per engine, and log presence, citation and accuracy. Do nothing else this round.
Fix and rerun
Correct the wrong claims at their sources, rewrite the answer pages, and rerun the same panel after recrawls. Compare to the baseline.
Add and report
Add ten prompts, re-read analytics, and report presence rate, citation share, accuracy and AI referral sessions as four separate lines.
Scroll sideways to follow the stages.
For what the market data says about where visitors go, see our AI search statistics. For the engine-specific side, read how to rank in Perplexity and Google AI Mode optimization. If you would rather have it run for you, AI SEO pricing shows what is included. Our AI visibility hub ties the pages together.
What most teams do today
- Ask one assistant one question and call it a result
- Trust a single blended visibility score
- Read Search Console clicks as the whole story
- Count mentions and never check what is said
- Change the prompts every month so nothing compares
What we build
- A fixed 20-prompt panel with named prompt types
- Three to five logged-out runs per prompt per engine
- Four separate lines: presence, citation, accuracy, referral
- Evidence saved as answer text, so changes can be shown
- A plain list of what the numbers cannot say

