A homeowner planning a kitchen remodel opens ChatGPT, asks which countertop material survives toddlers best, then asks for a few installers nearby. By the time anything gets typed into Google, the shortlist already exists. That shift creates a measurement gap for brands, because AI visibility is decided inside an answer the brand never sees, long before any click shows up in a report.
The scale is hard to ignore. OpenAI said in October 2025 that ChatGPT had reached 800 million weekly users, and Google reported in mid 2025 that AI Overviews served more than 2 billion people a month. A growing share of research now happens in a conversation rather than on a results page.
Where Traditional Reports Go Quiet
Search Console and analytics tools were built for a click-based web, so they capture only the edges of what happens inside AI answers. Google confirmed in 2025 that AI Mode data is counted inside the normal Search Console Performance totals, with no separate filter to isolate it. Analytics only records a visit when someone clicks a cited link, and many answers end without one. A reliable approach to AI visibility treats each answer as one sample rather than a final verdict, because the same prompt can return different brands and different sources from one run to the next.
That variability is the main reason a single screenshot of a ChatGPT answer proves very little. Measurement needs repetition, a fixed set of prompts, and a few supporting signals from data teams already own.
Four Layers Worth Tracking
No single metric captures how a brand performs in AI answers. Four layers, tracked together, give a far more honest picture.
| Layer | Question it answers | Main metric | Data source | How often |
| Presence | Does the brand appear in answers? | Mention rate across a fixed prompt set | Repeated prompt runs | Weekly |
| Citation | Is the site linked as a source? | Cited in N of M runs | Source lists in answers | Weekly |
| Traffic | Do answers send visitors? | Sessions from AI referrers | Analytics referral report | Monthly |
| Demand | Do answers spark new searches? | Branded query impressions | Search Console | Monthly |
The first two layers measure what happens inside the answer. The last two measure what spills out of it. Strong presence with no traffic or branded lift usually means the brand is mentioned but not recommended.
Presence starts with the right prompts
A useful prompt set holds 25 to 40 questions written the way real customers talk. Four types cover most businesses: category prompts (“best accounting software for freelancers”), problem prompts (“how do I stop clients paying invoices late”), comparison prompts between two named options, and local or niche prompts tied to a specific audience.
The best phrasing rarely comes from keyword tools. It comes from sales call notes, support tickets, and Search Console queries that run 7 words or longer, since those already read like spoken questions. Run each prompt at least 3 times per engine in every cycle, because one answer is an anecdote and three start to form a pattern.
Citation shows which pages do the work
A mention and a citation are different wins. A mention means the assistant named the brand. A citation means it linked to a page on the site as a source. Perplexity lists sources beside every answer, and ChatGPT shows them whenever it searches the web.
Record the exact URL that earns each citation. It is often a detailed guide or comparison page rather than the homepage, and that tells content teams which formats assistants trust. SEO Sets reports this as a rate, such as cited in 4 of the last 6 scans, which reflects how unstable single answers are.
Traffic arrives in smaller but sharper numbers
AI referrals are easy to spot once analytics is set up for them. Create a custom channel group named “AI Assistants” and match referrers with a pattern such as chatgpt\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai. ChatGPT also adds utm_source=chatgpt.com to many outbound links, which makes those visits simple to filter.
Volumes are usually modest. These visitors arrive with a narrower question, though, because the assistant has already handled the early comparison for them.
Demand shows up as a branded search echo
Plenty of people hear a brand name in an AI answer and then search for it on Google. Filter Search Console queries with a regex for the brand name and its common misspellings, then watch impressions month over month. Adding an “AI assistant” option to a “How did you hear about us?” form field captures the same signal directly from customers.
Turning Raw Runs Into a Weekly Number
A practical setup of 30 prompts, run 3 times each on 2 engines, produces 180 answers a week. Mention rate equals answers naming the brand divided by total answers. Track the same figure for 3 to 5 competitors in the same answers to get a share of voice.
Small swings mislead. A rise from 12 percent to 15 percent across 180 answers is only about 5 extra mentions, so a 4-week rolling average is a better guide than any single week. Fewer than 3 runs per prompt gives low confidence, and SEO Sets labels results by scan count for exactly this reason.
Context makes the weekly number useful. Log every content change, new page, or PR mention next to the date it went live. When mention rate climbs three weeks after a comparison guide is published, that timeline points to a likely cause. Keep the prompt set frozen for at least a quarter as well. Swapping prompts every week changes the test itself, and the trend line stops meaning anything. New prompts can go into a separate trial set until they earn a permanent spot.
Accuracy Counts as Much as Presence
Being named with outdated pricing or a discontinued product can cost more than not being named at all. Log every mention as correct, outdated, or wrong. Errors usually trace back to stale pages: an old pricing page, an about page that lists retired services, or third-party directory listings that never got updated. Fixing those sources, adding clear structured data, and keeping key facts consistent across the web gives assistants better material to draw from.
Questions Teams Ask When They Start Tracking
How many prompts are enough to start?
Twenty-five well-chosen prompts give a usable baseline. Expand to 40 once the first month of data shows which question types matter most for the business.
Which assistants should be tracked first?
Start with ChatGPT and Google’s AI features, since they reach the largest audiences, then add Perplexity, Gemini, or Claude based on where customers actually spend time.
Does ranking well on Google help with AI answers?
Often, yes. Many AI search features pull from live web results, so pages that rank well tend to be cited more. It isn’t guaranteed, which is why direct tracking still matters.
Final Thoughts
Customers now form opinions in conversations that leave no trace in standard reports. Measuring AI visibility through repeated prompts, citation rates, referral traffic, and branded demand turns that blind spot into numbers a team can act on, week after week.
Running 180 answers a week by hand is possible but tiring, and it tends to stop after the second month. Platforms like SEO Sets run these checks on a schedule and keep the history in one place, so trends replace guesswork. Start by writing 25 prompts this week, run them three times each, and use those results as the baseline every future change gets measured against.
