AI Visibility Metrics: What to Measure and Why

What are AI visibility metrics?

AI visibility metrics are 11 measures in 3 layers: access metrics show whether AI crawlers fetch a site, presence metrics show whether AI answers name or link it, and outcome metrics show whether those answers send visits and conversions.

A metric here is one number, collected the same way each period, that answers one question about how AI search treats a site. The three layers run in order, because each depends on the one before it: an engine that cannot fetch a page cannot cite it, and an answer that never names a brand cannot send it a visit.

• Access metrics count requests from AI crawlers and from the agents that fetch a page when a user’s question needs it: AI crawler fetches and user-triggered fetches.
• Presence metrics describe what AI answers say: mention rate, citation rate, share of model, answer position and sentiment from prompt tracking, plus AI feature impressions from Google and Copilot citations from Microsoft.
• Outcome metrics follow the reader after the answer: AI referral sessions and the conversions those sessions produce.

Most published lists stop at the presence layer. Adding access and outcome turns a scorecard into a diagnosis: when presence falls, the access layer shows whether a crawler was blocked, and the outcome layer shows whether the drop cost any business.

Which decision does each AI visibility metric inform?

Each AI visibility metric informs one decision. Crawler fetches decide access rules, mention rate decides entity work, citation rate decides which pages to rewrite, share of model decides which competitor to target, and referral conversions decide budget.

A metric without a decision attached becomes a number that is reported and ignored. The matrix below pairs each of the 11 metrics with the question it answers, the action it drives and the place its data comes from. The decisions are this site’s method, not an industry standard.

Read down the Data source column and the pattern is plain: prompt tracking feeds 5 of the 11 metrics, and first-party tools feed the other 6.

Access: can AI crawlers fetch the site?

The access layer uses the site’s own server logs. An AI crawler is an automated client that an AI company sends to collect training data or build a search index; each identifies itself with a user agent, the name string attached to every request. AI crawler fetches count requests from agents such as GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot. A sudden fall to zero usually traces to a robots.txt edit, a CDN bot rule or a firewall change, and that is the decision this metric serves.

User-triggered fetches are different. ChatGPT-User, Perplexity-User and Claude-User fetch a page only when a person’s request needs it, so each fetch marks a page that a live answer was about to use. The pages that collect these fetches are the ones to keep accurate first.

Presence: do AI answers name or link the brand?

Presence is the layer most people mean by AI visibility. Five of its metrics come from prompt tracking: running a fixed set of prompts on chosen engines and recording what each answer says. Mention rate asks whether the brand is named; citation rate asks whether its site is linked; answer position asks how early the name appears; sentiment asks whether the description is favourable and correct.

The formulas, a calculator and a worked example for the 5 prompt-based AI visibility metrics sit on their own page, so this map lists them only as questions. Vendors often merge them into one composite; Semrush, for example, shows a 0–100 AI Visibility score. A composite hides which of the five moved, so read the parts before the total.

Two presence metrics come from the engines themselves: AI feature impressions from Google and Copilot citations from Microsoft. Competitive standing needs its own number, because a rising mention rate means less if competitors rose faster; that comparison is share of model.

Outcome: do AI answers send visits that convert?

The outcome layer lives in web analytics. AI referral sessions are visits whose session source is an AI engine; ChatGPT adds utm_source=chatgpt.com to the links in its answers, while other engines depend on the browser passing a referrer. AI referral conversions are the key events those sessions complete, such as a form submission or a purchase.

Outcome metrics answer the budget question. Presence without outcome can still be worth having, since many readers act on an answer without clicking, but a GEO programme that never moves referral conversions needs a reason written down for continuing.

Where does the data for each metric come from?

The data comes from 5 sources: prompt tracking, Search Console's Generative AI performance report, Bing Webmaster Tools AI Performance, GA4 referrals and server logs. Prompt tracking alone supplies mentions, share of model, position and sentiment.

Each source sees one slice of AI search. None of them sees all of it, which is why the matrix needs five.

Prompt tracking

Prompt tracking is the only source that reads the answers themselves. A prompt is the question typed into an AI engine; tracking means running the same prompts on the same engines on a fixed schedule and logging, for each answer, the brands named, the URLs linked, the order and the tone. It is also the only source that covers ChatGPT, Perplexity, Gemini and Claude answers at all. It can be done by hand in a sheet or with software, and the AI visibility tracking tools compared on engines, prompt limits and refresh rates show what the paid route adds. The prompt set, run conditions and sheet layout for AI visibility tracking are covered step by step on the method page.

Search Console’s Generative AI performance report

Google’s Generative AI performance report is the first-party source for Google’s AI features. Its scope, as documented by Google:

• Engines: Google AI Overviews and AI Mode.
• Metric: impressions only.
• Dimensions: page, country, device and date.
• Availability: all sites from 31 August 2026.

The report answers which pages Google shows in AI features and when. It does not show the prompts behind an impression, whether the brand was named in the answer text, or which competitors appeared beside it.

Bing Webmaster Tools AI Performance

Microsoft launched AI Performance in Bing Webmaster Tools as a public preview on 10 February 2026. It reports four things: Total Citations, Average Cited Pages, grounding queries and page-level citation activity. A grounding query is the phrase Microsoft’s systems used to retrieve content for an AI answer. Microsoft notes that the counts are reported “without indicating placement or presentation”, so a citation in the first sentence and one in a footnote count the same.

GA4 referrals from AI engines

GA4 attributes each session to a source. Visits from ChatGPT carry the UTM parameter utm_source=chatgpt.com, a tag added to the link itself, so they survive when the browser strips the referrer. Visits from Perplexity, Gemini, Copilot and Claude appear under their own domains only when a referrer is passed; otherwise they fall into direct traffic. Marking form submissions and purchases as key events turns those sessions into AI referral conversions.

Server logs

Server logs record every request to the site with its user agent, time and URL. Filtering them by agent name gives the access layer:

• Training and search crawlers: GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot.
• User-triggered agents: ChatGPT-User, Perplexity-User, Claude-User.

Logs show that a page was fetched, never whether an answer later used it. User-agent strings can also be copied by other bots, so the operators’ published IP lists confirm which requests are genuine.

What can first-party reports not measure?

First-party reports cannot measure unlinked brand mentions, competitor share, answer sentiment or engines outside Google and Microsoft. Search Console reports AI feature impressions, Bing reports Copilot citations, and GA4 reports clicks; everything said inside ChatGPT, Perplexity, Gemini or Claude answers needs prompt tracking.

The gaps follow from what each engine chooses to publish:

• Unlinked mentions never reach Search Console, Bing or GA4, because none of them reads answer text.
• Competitor share is absent from all three, since each reports only on the site that owns the account.
• Sentiment is absent too; no first-party report says whether an answer praised or warned against a brand.
• Other engines such as ChatGPT, Perplexity, Gemini and Claude publish no site-owner report, so their answers are visible only from outside.

The split works in both directions. Prompt tracking samples a small set of prompts, while Search Console counts impressions across every query Google served. Neither replaces the other.

Why do AI visibility metrics disagree with each other?

AI visibility metrics disagree because they count different events on different engines. Search Console counts Google impressions across every query, prompt tracking samples a fixed prompt set, and GA4 counts only the clicks, so one layer rises while another stays flat.

Three differences drive most conflicts. The engines differ: a gain in AI Mode says nothing about Perplexity. The method differs: Search Console is a census of Google’s own impressions, while prompt tracking is a sample chosen by the site owner. The event differs: an impression, a named mention and a click are three separate things, and each layer counts only one.

An illustration: Search Console impressions in AI Overviews climb over a month, while the tracked citation rate stays flat. Both numbers can be right. The new impressions may come from long, specific queries that the fixed prompt set never asks, and the prompt set may be weighted towards short commercial questions where nothing changed. The response is to read the pages gaining impressions, check which queries they answer, and decide whether those questions belong in the prompt set.

The reverse case also happens: tracked mentions rise while referral sessions stay flat, because the answers name the brand without linking it. That points back to citation rate, not to the analytics setup.

How often is each metric reviewed?

Server logs and prompt tracking are reviewed weekly; Search Console, Bing Webmaster Tools and GA4 are reviewed monthly. Weekly review catches blocked crawlers and lost citations early, while monthly review smooths the daily swings in impressions and referral sessions.

This is the cadence used here, not an industry standard; the Review cadence column in the matrix applies it to each metric. It matches the protocol of this site’s Experiment 01, which checks 50 prompts weekly on 6 engines and is still running, with no results reported yet.

The weekly checks are the fast-failing ones. A blocked crawler shows in the logs within days, and a lost citation shows in the next prompt run. The monthly reports move more slowly and swing from day to day, so a weekly reading of them mostly records noise. Record every reading with its date and keep the prompt set unchanged within a quarter, so each period compares like with like.

Each metric has a page that explains it in full. Which one comes next depends on the decision in front of you.

Which guide covers each AI visibility metric?

Four guides are live and three are in preparation. Each guide takes one row or source from the matrix and covers it in depth.

• AI visibility: 5 metrics and tracking

  Formulas, a calculator and a free sheet for mention rate, citation rate, share of model, answer position and sentiment.

• Share of model

  The competitive presence metric: how to pick the competitor set and report the share over time.

• AI visibility tracking tools

  Trackers compared on engines covered, prompt limits, citation tracking, refresh frequency and pricing model.

• GEO experiments

  The public test log, including Experiment 01's weekly prompt-tracking protocol.

• Guide in preparation

  ChatGPT referral traffic in GA4

  Separating AI-referred sessions and key events from other traffic.

• Guide in preparation

  Building a prompt tracking set

  Choosing prompts from real buyer questions and keeping them fixed.

• Guide in preparation

  AI Mode data in Search Console

  Reading the Generative AI performance report page by page.