Skip to content
Learn

Measuring AI visibility and testing

In this lesson

Section titled “In this lesson”
  • Find AI referral traffic in Google Analytics and Microsoft Clarity, and explain what it misses.
  • Measure a brand in AI answers with five metrics: prompt coverage, mention rate, citation rate, answer share and sentiment.
  • Explain why one AI answer is not a measurement, and how many answers you need.
  • Compare four AI visibility tools: Profound, Otterly, Peec and Ahrefs Brand Radar.
  • Plan a test that shows whether an SEO or GEO change worked, with a control group.

On day 08 you measured the classic path: impression, click, visit. AI answers break this path. A person can read about your brand in ChatGPT, never click, and buy a week later. This lesson has two parts. First, you measure what you can see in AI answers. Second, you test changes, so that you know what caused a result.

The shape of the problem

Section titled “The shape of the problem”

AI visibility has three layers. Each layer needs a different kind of data:

Layer Question Data
Traffic Did AI answers send visitors? Your analytics: GA4, Clarity
Visibility Did AI answers show or cite us? Search engine reports, prompt tracking tools
Perception What do AI answers say about us? Prompt tracking tools, your own reading

The traffic layer is the most reliable, because it is your own data. But it is the smallest layer, because most AI exposure brings no click. The visibility and perception layers are larger, but you see them only through samples. Good reports use all three layers and say which layer each number comes from.

Zero-click attribution

Section titled “Zero-click attribution”

A zero-click search is a search that ends without a click to any website. In the SparkToro and Datos study of July 2024, 58.5% of US Google searches and 59.7% of EU searches were zero-click. The study used clickstream data from September 2022 to May 2024. On day 07 you saw that Pew found fewer clicks when an AI Overview appears.

Zero-click attribution is the work of giving value to this exposure without a click. You cannot see the person, so you measure signals around them:

  • Impressions and citations. Search Console now has generative AI performance reports. Since 31 August 2026, all sites have them. The reports show impressions in AI Overviews and AI Mode, by page, country, device and date. They show no clicks and no queries. Bing Webmaster Tools has a similar AI Performance report, covered below.
  • Brand demand. Exposure in answers can make more people search for your name later. Track branded queries and share of search, which you learned on day 08.
  • Direct questions. Ask new customers how they found you. A free-text field on a sign-up form is enough. This is the only way to hear “ChatGPT told me”, because no tool records it.

None of these signals proves cause on its own. Together they show a trend. Report them as signals, and do not add them up as if they were clicks.

LLM referral traffic

Section titled “LLM referral traffic”

LLM referral traffic is the visits that arrive from a link in an AI assistant, such as ChatGPT, Claude, Gemini or Perplexity. The roadmap also calls it AI citation referral traffic. It is the part of AI visibility that you can connect to sign-ups and revenue.

Three tools find it for you:

  • Google Analytics. On 13 May 2026, GA4 added an AI Assistant channel to its default channel group. When the referrer matches a known AI assistant, GA4 sets the medium to ai-assistant. Google names ChatGPT, Gemini, Deepseek, Copilot, Grok and Claude as examples. You learned on day 08 that GA4 puts Google’s AI Overviews and AI Mode in Organic Search. The AI Assistant channel excludes them.
  • Microsoft Clarity. Since August 2025, Clarity has an AI Platform channel group for organic visits from AI chat. A Paid AI Platform group counts ads inside AI chat.
  • UTM tags. OpenAI’s publisher FAQ says that ChatGPT adds utm_source=chatgpt.com to the links in its search results. So you can also filter by source.

Know the limits. First, Google AI features hide inside Organic Search. Search Console’s generative AI report is the only place to see their size, and it shows only impressions. Second, not every visit carries a referrer. Peec AI says it adds its own detection to GA4’s channel, because the default grouping “might miss” some referrals. Treat the AI Assistant channel as a minimum, not a full count.

Compare AI referrals with other channels on the same metrics: engagement rate, key events and revenue. Microsoft says that AI visitors land deep inside a site. Use Clarity recordings to see what they do on those landing pages.

AI visibility measurement

Section titled “AI visibility measurement”

AI visibility measurement is the set of metrics that show how often and how well a brand appears inside AI answers. The method is prompt tracking: you choose a set of prompts, run them on AI systems on a schedule, and read the answers. Day 08 taught rank tracking for keywords. Prompt tracking is the same idea for prompts, with one big difference that the next section explains.

The core metrics are these:

Metric What it counts Example
Prompt coverage The share of your prompt set where the brand appears at least once Present for 18 of 40 prompts = 45%
Mention rate (Peec calls it visibility) Answers that name the brand, divided by all answers 120 of 400 answers = 30%
Citation rate Answers that link to your site as a source, divided by all answers 40 of 400 answers = 10%
Citation position Where your link appears in the source list First of five sources
Answer share Your mentions, divided by all brand mentions See the next section

A mention and a citation are different. A mention names your brand in the answer text. A citation links to one of your pages as a source. An AI system can mention you and cite a review site, or cite your page and not name you. Track both. Mentions show reputation. Citations show that your content is used, and only citations can send traffic.

Choose prompts like a buyer

Section titled “Choose prompts like a buyer”

The prompt set is the most important decision. A set that has your brand name in every prompt measures only brand recognition. Mix the prompt types:

  • Non-branded category prompts, for example “best project management tool for a small agency”.
  • Comparison prompts, for example “Asana vs Trello for agencies”.
  • Branded prompts, for example “is Trello good for agencies”, to check what the answers say about you.

Group prompts by topic and by funnel stage, as you grouped keywords on day 04. Peec AI classifies each prompt as branded or non-branded, and as informational, commercial or transactional.

One answer is not a measurement

Section titled “One answer is not a measurement”

AI answers change from run to run. In January 2026, SparkToro published a study by Rand Fishkin and Patrick O’Donnell. About 600 volunteers ran 12 prompts on ChatGPT, Claude and Google’s AI 2,961 times. ChatGPT and Google’s AI had less than a 1 in 100 chance to give the same list of brands twice. The same list in the same order came about once in 1,000 runs.

But how often a brand appears was more stable. One cancer hospital appeared in 69 of 71 ChatGPT answers, but was first in only 25. So the authors recommend a visibility percentage over many prompts and many runs. They call “ranking position in AI” unreliable. The study is not peer reviewed, and the authors say so.

This is why citation position and brand position are weak metrics. Use them as context, not as a target. The diagram shows the effect of sample size.

Example data. Each brand has a fixed true rate (orange line). Move the slider to collect more answers, and press 'Measure again'. Notice that with 5 answers the grey range is very wide, and that it gets narrow only near 100 answers.

First-party reports from search engines

Section titled “First-party reports from search engines”

Two search engines now give you citation data from their own systems, at no cost:

  • Bing Webmaster Tools AI Performance (public preview since 10 February 2026) shows citations in Microsoft Copilot, AI summaries in Bing and some partners. It reports total citations, average cited pages, page-level citations and grounding queries: the phrases that the AI used to retrieve your content. Grounding queries are a sample. On 16 June 2026, Bing added intents, topics, a compare view and citation share. Citation share is your part of all citations for a grounding query. Bing says that citations do not show “ranking, authority, or placement”.
  • Search Console generative AI reports show impressions in AI Overviews and AI Mode, as above.

These reports are real data, not samples of your own prompts. They do not cover ChatGPT, Claude or Perplexity. So you still need prompt tracking for those systems.

Competitive answer share and benchmarking

Section titled “Competitive answer share and benchmarking”

Competitive answer share is your brand’s part of all brand mentions in answers to the same prompts. Peec AI calls it share of voice. Peec’s example: your brand appears 4 times in 10 answers, and a competitor appears 12 times. Your share is 4 / (4 + 12) = 25%. Your mention rate is still 40%.

The two metrics answer different questions. Mention rate asks “do answers know us?”. Answer share asks “who wins the answers?”. If your mention rate rises but your share falls, competitors grew faster than you.

Competitor benchmarking compares your numbers with competitors on the same prompts, engines and dates. It tells you if a number is good. A 30% mention rate is weak in a category with three brands, and strong in a category with fifty. Benchmark the classic metrics from day 08 in the same report: rankings, share of search and backlinks. Bing’s citation share does not name competitors, so the tools below are the main source for this.

Entity and sentiment

Section titled “Entity and sentiment”

Being named is not enough. Entity and sentiment tracking checks two things in each answer:

  • Entity: does the answer connect your brand to the right topics, products and facts? You built these connections with structured data on day 05b and consistent profiles on day 06b.
  • Sentiment: is the tone positive, neutral or negative?

Peec AI scores sentiment from 0 to 100 and says that most scores are between 65 and 85. Its fact-checking view compares answers with the facts that you give it, and shows claims that contradict them. Wrong facts in answers are often the most useful finding. They show which source the AI trusts, and you can fix or correct that source.

Read a sample of the answers yourself too. A sentiment score is a model’s judgment of another model’s text. It can miss sarcasm, and a “neutral” score can hide an old price or a closed product.

AI visibility tools

Section titled “AI visibility tools”

Four tools from the roadmap, with facts from their own pages on 19 October 2026:

Tool What it does Engines Price (19 October 2026)
Profound Prompt tracking with citations, sentiment and competitors, plus agent analytics and AI marketing agents Up to 9, including ChatGPT, Perplexity, AI Mode, Gemini, Copilot, Claude Free 7-day trial (50 prompts); enterprise price on request
Otterly Daily prompt tracking, brand visibility, sentiment, citation sources and GEO audits ChatGPT, AI Overviews, Perplexity, Copilot; Claude, AI Mode, Gemini as add-ons Lite $29 a month (15 prompts); Standard $189 (100 prompts)
Peec AI Visibility, share of voice, sentiment and position against competitors; GA4 referral data Choose 3 models on standard plans; up to 13 on enterprise Starter 50 prompts; prices not shown in the page text
Ahrefs Brand Radar Custom prompt tracking, and an index of 481 million prompts modeled from Ahrefs keyword data AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini, Copilot Custom prompts from €47 a month; index from €179 a month

The big difference is where prompts come from. Profound, Otterly and Peec run the prompts that you choose. Brand Radar can also show any brand across its own large prompt index, with no setup. Ahrefs says the index works best “once your brand already appears in AI search”.

All four tools run prompts through their own accounts, not through your customers’ accounts. Answers for a real person can differ by place, history and settings. So all tool data is a sample. Choose a tool for its prompt count and engines, and keep the same setup for months. A new setup breaks your trend lines.

GEO and SEO testing

Section titled “GEO and SEO testing”

GEO and SEO testing is running a controlled experiment to check whether a change improved visibility. Without a test, you guess. Rankings, AI answers and demand all move every week, so a rise after a change does not prove that the change caused it.

Content and schema A/B tests

Section titled “Content and schema A/B tests”

A classic A/B test splits visitors into two groups. That does not work for SEO, because Googlebot is one visitor. You cannot show it two versions of a page. Showing Googlebot a different version than humans is cloaking, which you learned is spam on day 02.

So an SEO A/B test splits pages, not visitors. You take a group of similar pages, for example 1,000 product pages. You change half of them (the variant group) and keep the other half the same (the control group). Then you compare search traffic. SearchPilot builds the two groups so that they have similar traffic and seasonality. It then forecasts what the variant pages would have done without the change, from the control group’s live data. Demand, seasonal changes and algorithm updates affect both groups. So the difference between them is the effect of the change.

You can test content changes, such as new titles, or schema changes, such as adding Product markup. The test needs many similar pages with enough traffic. A small blog cannot run one.

Google gives these rules for tests that use separate URLs:

  • Do not cloak.
  • Put rel="canonical" on the test URLs, pointing to the original.
  • Use 302 redirects, not 301.
  • End the test as soon as you have an answer.
Method

Example data. Set demand to +20% and the true effect to 0%. With before/after, notice the fake +20% result. Choose the split test: the control pages remove the demand change. Then try other values.

Before/after prompt tests

Section titled “Before/after prompt tests”

A before/after prompt test runs the same prompts before and after a change, and compares the answers. It is the usual test for GEO, because you cannot split AI answers into groups. The diagram shows its weakness: anything else that changed between the two dates mixes into the result. Model updates happen often. Bing also warns that citation patterns change with “model changes, freshness signals” and other causes.

Make the test stronger:

  1. Run each prompt many times before the change, as the first diagram showed.
  2. Wait until the AI systems have fetched the changed page. Check your server logs for their crawlers, which you learned on day 07c.
  3. Add control prompts on topics that you did not change. If they move too, the cause is not your change.
  4. Record the engine, the country and the dates with each result.

Before/after tests give evidence, not proof. Split tests give stronger proof, but they need a large site. Use split tests for big decisions where you can, and before/after tests for the rest.

Key terms

Section titled “Key terms”
  • Zero-click search: a search that ends without a click to any website.
  • Zero-click attribution: giving value to exposure in search or AI answers that brings no click.
  • LLM referral traffic: visits that arrive from a link in an AI assistant.
  • AI Assistant channel: the GA4 default channel for visits from AI assistants such as ChatGPT; it excludes Google’s AI features.
  • Prompt tracking: running a fixed set of prompts on AI systems on a schedule and recording the answers.
  • Prompt coverage: the share of tracked prompts in which a brand appears at least once.
  • Mention rate: the share of AI answers that name a brand.
  • Citation rate: the share of AI answers that link to a site as a source.
  • Grounding query: in Bing’s AI Performance report, a phrase that the AI used to retrieve your content.
  • Competitive answer share: a brand’s mentions divided by all brand mentions in the same answers; also called share of voice.
  • Competitor benchmarking: comparing your metrics with competitors on the same prompts, engines and dates.
  • SEO A/B test: a test that changes a variant group of pages and compares it with a control group of similar pages.
  • Before/after prompt test: running the same prompts before and after a change and comparing the answers.

Exercise

Section titled “Exercise”

Exercise · about 15 min

Run a small prompt tracking test by hand

You need a browser and two free AI assistants, for example ChatGPT and Perplexity. Pick a category that you know, for example running shoes or note-taking apps.

  1. Write five prompts that a buyer could ask. Do not put a brand name in four of them. Put one brand name, your “brand”, in the fifth.
  2. Open a new chat for each run. Run each prompt three times on each assistant. This gives 30 answers.
  3. For your brand and two competitors, count the answers that name each brand.
  4. Count the answers that link to each brand’s own website.
  5. Calculate the mention rate, the citation rate and the answer share for your brand.
  6. Write one sentence: which number would you trust least, and why?
  7. Optional: if you have a GA4 property, open Reports, Acquisition, Traffic acquisition. Find the AI Assistant row.
Show a possible solution

Your numbers depend on the category. A typical result for a note-taking app:

  • Mention rate: the brand appears in 14 of 30 answers, so 47%. Competitor 1 appears in 22 answers, competitor 2 in 9.
  • Citation rate: only 4 of 30 answers link to the brand’s own site, so 13%. Most links go to review sites and Reddit. That is a clue about which sources the AI trusts.
  • Answer share: 14 / (14 + 22 + 9) = 31%.
  • Least trusted: any number for one prompt, because three runs is too few. The brand can appear in all three runs today and in none tomorrow. Thirty answers in total give a rough picture, but a change of a few points means nothing.
  • GA4: a small site often sees only a few AI Assistant sessions a month. Google’s AI Overviews and AI Mode visits are in Organic Search, not in this row.

Check yourself

Section titled “Check yourself”
A report says: 'GA4 shows 300 AI Assistant sessions, so AI search brings us 300 visits a month.' What is wrong?
You ran one prompt once on ChatGPT, and your brand was first in the list. Your manager wants to report 'position 1 in ChatGPT'. What should you say?
This quarter your mention rate rose from 30% to 35%, but your answer share fell from 25% to 20%. What happened?
You add FAQ text to 500 of 1,000 similar product pages. Why keep the other 500 pages unchanged?
After you rewrite a guide, your brand appears in more Perplexity answers. Which extra check makes the before/after test most convincing?

Sources

Section titled “Sources”

Next

Section titled “Next”

Day 09 leaves measurement and covers local and international SEO, special kinds of search, site migrations and SEO audits.