
To measure AI visibility, pick a fixed set of questions your customers really ask, run them on a schedule across the AI assistants that matter, and log whether your brand is mentioned, whether your site is cited and which competitors appear instead. Then repeat the same protocol and read the trend, not any single run. This guide walks through each step, the metrics worth tracking, how to read a score, and how to connect it all to traffic and leads.
TL;DR
- AI visibility is how often, and how well, AI answers name or cite your brand. Rankings and clicks alone cannot show it.
- Track five things: mentions, citations, share of voice, position or prominence, and AI referral sessions.
- Use a frozen prompt library, fixed run conditions and a steady cadence. Log results per prompt and per engine.
- There is no universal “good” score. Compare yourself with your category and watch the direction over weeks.
- Treat every reading as a sample. AI answers vary from run to run, so trends beat snapshots.
Table of Contents
- What AI visibility is and why rankings don’t capture it
- The metrics worth tracking
- Step 1: Build a frozen prompt library
- Step 2: Run the prompts under fixed conditions
- Step 3: Set a baseline you can defend
- Step 4: Read the score without fooling yourself
- Step 5: Connect visibility to traffic and revenue
- Why your numbers will wobble
- A simple weekly routine
- Put measurement and publishing on one system
- FAQ
- Sources
What AI visibility is and why rankings don’t capture it
Traditional search visibility is about position in a list of links. AI visibility is different because AI systems synthesize information into direct answers, often with citations, instead of presenting links for users to click. A buyer can ask an assistant for the best option in your category, get a short answer naming three brands, and never see a results page.
That changes what you count. A page can rank well on Google and still be absent from the answer an assistant gives for the same topic. The reverse happens too: a brand can be named without its site ever being linked. So you need to measure two separate things: whether you are talked about, and whether you are used as a source.
The measurement also has to span several surfaces. ChatGPT, Perplexity, Gemini, Claude, Copilot and Google AI Overviews each retrieve and cite differently, so one blended number hides more than it shows. If you want the background on how to improve those numbers once you can see them, our guide to improving AI search visibility picks up where this one ends.

The metrics worth tracking
You do not need thirty metrics. Search Engine Land defines the core signals as mentions, citations, prompt-triggered visibility, the context of mentions and share of voice. Add AI referral traffic from your own analytics and you have a complete, workable set.
| Metric | What it tells you | How to calculate it |
|---|---|---|
| Mention rate | How often AI answers name your brand | Answers naming you ÷ answers checked |
| Citation rate | How often your site is linked as a source | Answers linking your domain ÷ answers checked |
| Share of voice | Your presence versus competitors | Your mentions ÷ all brand mentions in the same answers |
| Prominence | Whether you are first, a passing mention or a footnote | Position in the answer, or a simple 1 to 3 rating |
| Sentiment and context | Whether the description is accurate and favorable | Manual tag per answer: positive, neutral, negative, wrong |
| AI referral sessions | Real visits from assistants | Analytics sessions from AI referrers |
Mentions versus citations
A mention means the assistant names your brand in its text. A citation means it links to your site. Both matter, but they point to different fixes. Mentioned but not cited usually means the model knows you from elsewhere but does not find a page on your site that answers the question well. Cited but not mentioned means your content is feeding the answer while the brand stays invisible.
A simple worked example: you run 40 prompts on one engine. Your domain appears as a source link in 18 of them. Your citation rate is 18 ÷ 40, or 45%. If your brand name appears in the text of 22 answers, your mention rate is 55%. Keep both numbers.

Share of voice
Share of voice puts your number in context. If five brands are named across your prompt set and you account for 20 of 100 total mentions, your share is 20%. It is the best single metric for judging competitive position, because it moves when a competitor gains even if you stay flat.
Prominence and accuracy
Being named fifth in a list of ten is not the same as being the first recommendation. Record where you appear. Also read what the assistant says about you. A wrong price, an outdated product description or a missing feature is a content problem you can fix, and you will only find it by reading the answers, not by counting them.
Step 1: Build a frozen prompt library
Everything depends on the prompt set. Use real questions your buyers ask, in their words, not your internal product names.
- Collect 20 to 40 prompts. Pull them from sales calls, support tickets, search query reports and the questions customers ask before buying.
- Cover the funnel. Include broad category questions (“best tools for X”), comparison questions (“A versus B”), problem questions (“how do I fix Y”) and branded questions (“is your brand any good”).
- Tag each prompt. Label by topic, intent and product line so you can slice results later.
- Freeze it. Do not reword prompts between runs. Small wording changes shift answers, so a changing library destroys your trendline. Add new prompts only at planned intervals, and report them separately from the core set.
A good rule: keep a stable core of prompts that never changes, and add a handful of new ones each quarter to follow how your market asks questions. Our team’s approach to question research, which starts from what customers and competitors actually rank for, is covered in building an AI search optimization strategy.
Step 2: Run the prompts under fixed conditions
Consistency is what makes the data credible. Decide the conditions once and write them down:
- Engines: the platforms your buyers use, for example ChatGPT, Perplexity, Gemini and Google AI Overviews.
- Region and language: same country and language every run.
- Cadence: weekly is a sensible default for a small team. Daily is better for volatile categories.
- Account state: use a clean session where you can, so personal history does not skew the answer.
- Repeats: run each prompt more than once per cycle, because a single answer is a sample, not the truth.
Then log one row per prompt per engine per run. Useful columns: date, engine, prompt, brand mentioned (yes or no), site cited (yes or no), cited URLs, position, sentiment, competitors named. A spreadsheet is enough to start. As the library grows, a tracking tool saves hours, which is why we built AI visibility tracking into the platform.
Step 3: Set a baseline you can defend
Your first run is a snapshot, not a baseline. We recommend running the full library for several weeks under the same conditions, then taking the average as the number you will be judged against. Four weeks is a reasonable minimum, and up to eight gives a steadier picture if your volume is small.
For each engine, record:
- mention rate and citation rate,
- share of voice against your three or four closest competitors,
- the pages that get cited most often (and the ones that never do),
- any inaccurate statements, with the prompt that triggered them.
Keep results separate per engine. Averaging ChatGPT with AI Overviews hides the fact that you might be strong in one and absent in the other.
If you want a fast starting point before building the full protocol, the free AI visibility checker shows whether ChatGPT can see your website at all.
Step 4: Read the score without fooling yourself
The simplest score is answers naming you divided by answers checked. People then ask the obvious question: is that number good?
There is no universal threshold. A score that is excellent for a three-month-old product can be weak for a category leader. Judge your number against three references:
- Your category. Compare against the top brand in your set and where you sit in the cluster of competitors.
- Your own history. Direction matters more than level. A rise from 8% to 15% over a quarter is meaningful progress.
- Your intent mix. Branded prompts should score high, since people asking about you by name should get an answer about you. Broad category prompts are the hard ones, and the real battleground.
A few practical interpretations:
- High mention, low citation: publish clearer, more specific pages that answer the exact question and mark them up properly.
- High citation, low mention: strengthen brand signals, such as consistent naming, an about page and third-party references.
- Low everywhere on branded prompts: check that assistants can read your site and that your key pages describe what you do in plain language.
- Strong on one engine, weak on another: study which sources that engine cites and match them.
When the weak spots are about content coverage, the playbooks in our guides on getting cited by ChatGPT and ranking in AI Overviews give the next moves.
Step 5: Connect visibility to traffic and revenue
Mention rates are a leading indicator. Your boss or client will want to know about visits, leads and sales. Bridge the gap in your analytics:
- Tag AI referrers in GA4. Create a channel group or custom dimension that groups sessions from AI assistants, for example referrals from chatgpt.com, perplexity.ai, gemini.google.com and claude.ai.
- Report on landing pages. See which pages AI visitors land on, and compare them to the pages your citation logs show assistants quoting.
- Track conversions. Attribute sign-ups, demo requests or purchases to AI-sourced sessions.
- Add a self-reported field. Add “How did you hear about us?” with an option for AI assistants. Many AI-influenced buyers never click a link, so this catches what analytics misses.
- Feed your CRM. Pass the source into your CRM so you can follow AI-sourced leads through to closed deals.
Expect the traffic numbers to look small next to organic search at first. AI referrals are typically a fraction of total visits, and many assistant users read the answer and then search your brand name directly. Branded search growth and direct traffic are secondary signals worth watching next to the referral line.
Why your numbers will wobble
AI answers are not fixed rankings. A 2026 academic paper on quantifying uncertainty in AI visibility argues that citation visibility metrics behave like random variables, because generative systems sample their output. The same query sent twice to the same platform can return a different set of cited sources, so one run cannot be read as your true citation rate.
This has practical consequences:
- Never report a single run. Report an average over repeats and a trend over weeks.
- Expect noise from model updates. Engines change models, interfaces and citation behavior without notice, so a drop can come from the engine rather than your site.
- Mind wording, region and account state. Different phrasing, locations or logged-in histories produce different answers. This is why fixed conditions matter.
- Do not over-claim causation. A rise after you published new content is encouraging, but it does not prove the content caused it. Note the action, the date and the change, and look for the pattern across several cycles.
- Watch for ghost citations. Sometimes an assistant uses your information without naming or linking you. Your logs will undercount that, so pair them with branded search and self-reported data.
None of this makes measurement pointless. It means you read trends, keep the method steady and write your reports with ranges instead of false precision.
Our view: measure to decide what to publish next
We built Tepine on a simple belief: measurement only matters if it changes what you publish. A visibility score that sits in a dashboard does nothing. A score that tells you which question you are missing, so a new article goes live tomorrow, compounds.
That is why our platform plans content from proven search demand, from competitors and from your own products, and publishes daily. It extracts your brand voice from your existing site content, so the articles sound like your business rather than generic AI text. It adds structured schema for Google and AI assistants, and it tracks AI visibility metrics so you can see the effect over time.
For freelancers, small businesses and agencies, the problem is rarely a lack of ideas. It is the time to write, publish and check results every week. Our own claims are a 38% mention rate across AI platforms and an average position of #1.8 when referenced. We treat those as the direction we are building toward and measure ourselves with the same method described above.
Tepine
A simple weekly routine
Once the system is in place, it should take under an hour a week for a small prompt set:
- Run the frozen prompt library on each engine, repeating each prompt a few times.
- Log mentions, citations, position, sentiment and competitors.
- Update three charts: mention rate by engine, citation rate by engine and share of voice against competitors.
- List the prompts where a competitor was named and you were not. Those are your content gaps.
- Pick two or three gaps and publish a page for each that answers the question directly.
- Check GA4 for AI referral sessions and note any content or site changes made that week.
Every month, review the trendlines rather than the weekly dots. Every quarter, add a few new prompts, retire any that no longer reflect how customers ask, and write a short report: what moved, what you did, what you will do next.
Put measurement and publishing on one system
The hardest part of AI visibility work is keeping up: finding the gaps, writing pages that fill them, publishing consistently and then checking the effect. Tepine handles that loop. It plans topics from competitor-informed keyword research, writes in your brand voice, holds drafts in a review queue if you want to approve them, publishes to common CMS platforms, and tracks AI visibility alongside.
The Starter plan is $79 per month, with a free trial before you subscribe. See the pricing page for what is included, or look at how automatic daily publishing works.

If you are still deciding whether automated content can support search performance, our article on whether AI content generation works for SEO goes through it in detail.
FAQ
What is a good AI visibility score?
There is no single number that counts as good. It depends on your industry, your competitors and how long you have been in the market. Compare your score with the category leader, check it separately for branded and broad prompts, and pay close attention to whether it is rising over weeks.
How many prompts do I need to track?
Start with 20 to 40 real customer questions that cover the whole buying journey. That is enough to see patterns without making each cycle unmanageable. Keep a stable core set and add a few new prompts each quarter so your trendline stays comparable.
How often should I measure AI visibility?
Weekly runs with several repeats per prompt suit most small teams, and you should read the results as monthly trends. A single day’s reading is mostly noise because assistants vary their answers. Volatile categories may justify daily runs, while slow-moving ones can report quarterly.
Can I measure AI visibility for free?
Yes. You can run your prompts by hand and log results in a spreadsheet, and a free AI visibility checker gives a quick first read on whether assistants can see your site. Manual tracking gets time-consuming as your prompt list and engine count grow, which is when automated tracking pays off.
Does higher AI visibility mean more revenue?
Not automatically. Visibility shows you appear in answers, but it does not prove a user saw the answer or bought because of it. Connect it to GA4 sessions, conversions and CRM data, and add a self-reported source field to see the real business effect.
Sources
- What brands need to know about AI visibility (TechCrunch)
- How to track visibility across AI platforms (Search Engine Land)
- Quantifying Uncertainty in AI Visibility (arXiv)


