AI Visibility Metrics: The Complete Measurement Guide

Most AI visibility reports are a screenshot with a caption. Someone asked an assistant a question, liked the answer and pasted it into a slide. Answers vary from one run to the next, so a single capture proves almost nothing, and a board that learns this stops trusting the whole subject.

Anyone asked to explain AI search to a board will find this useful. The guide defines the metrics, flags where definitions disagree, shows what Google and Bing now publish themselves and lays out a routine that holds up under questioning. A list of prompts is the only prerequisite, and the guide shows how to build one.

Four questions every metric answers

Think of the field as four questions put to every answer an assistant gives. Did it name the brand? Did it cite the site? How did the brand fare against rivals in the same answer? And was every claim about the brand true?

Peec AI offers the cleanest first division: brand visibility means the company name is present in generated answers, while website citations mean your content is drawn on as a source. AirOps makes the same cut under the labels mention rate and citation rate. These move apart and call for different repairs, as brand visibility versus website citations shows.

The working set, in plain terms

Metric	How to count it	Question answered	Common trap
Mention rate	Answers that name you, out of all answers collected	Does the assistant know you exist?	Counting a name in a list of ten as a win
Citation rate	Answers that link your pages, out of all answers collected	Are your pages used as evidence?	Assuming a mention implies a link
Share of voice	Your mentions out of every tracked brand's mentions	How do you stand against rivals?	Letting the rival list change mid-quarter
Accuracy	Correct descriptions out of answers that mention you	Is the description right?	Never reading the answers
Sentiment	Split of positive, neutral and negative framing	How are you framed?	Trusting an automated label without a spot check
AI referral sessions	Visits in analytics from AI source domains	What traffic can you see?	Forgetting that some products send no referrer

The table hides a snag: vendors disagree about the denominator. AirOps writes share of voice as responses mentioning your brand divided by all responses for the prompt set, which most others would call mention rate. The share-of-all-brand-mentions version is a different number. Both versions are laid out in our share of voice walk-through, and the terms explainer fixes what mention, citation and link mean.

Why one capture is not a measurement

AirOps says roughly 30% of brands remain visible from one answer to the next, and that about 57% of brands which vanished from an answer returned within two more runs. No sample or method accompanies either figure, so read them as a plausible shape, not a constant. Testing it yourself is easy: put one question to the same assistant twice and compare the brand lists. The causes are covered in why answers change every run.

That variability dictates the routine. Run each prompt more than once per check, across two or more engines. Quote an average over a window, or a range, rather than a single day. Hold the list steady for a quarter and record each edit, with the date, engine and model setting. AirOps suggests a window of a week or more and attention to any move above five percentage points, though these are one vendor's working rules.

The prompt set decides the answer

A prompt that contains your brand name tests awareness and will flatter you. A generic prompt tests discovery and will humble you. A trustworthy set holds both, tagged. Draw the questions from real buyer language, following the prompt research method, and keep the balance fair using branded versus unbranded prompts.

What the platforms now report for free

Google: generative AI performance reports

Earlier guidance said AI feature traffic simply sat inside the ordinary Web results. That has changed. On 3 June 2026 Google announced dedicated Search Generative AI performance reports in Search Console, and by 31 August 2026 it had rolled them out to all websites worldwide. They show impressions within AI Overviews, AI Mode and generative features in Discover, plus pages, countries, devices and dates at hourly to monthly granularity. The post lists impressions as the metric, so clicks and citations are not separate figures there, and the data also remains inside the main performance report. Google's older documentation still says AI features count in the Search Console performance report. See how to report on AI Overviews.

Bing: the AI Performance report

Bing's AI Performance report entered public preview in February 2026 and covers Copilot, AI summaries in Bing and a few partner products. Its core numbers are total citations, the average of cited pages and the grounding queries behind them. Mid-June 2026 added Intents, Topics, Citation Share and Compare. Citation Share is the slice of a grounding query's citations that goes to your site rather than every other; Microsoft stresses that it is observational, with no rival domains, traffic share or quality score. A citation is not a click. Our Bing update covers the detail.

First-party traffic and server logs

Analytics logs a referral session only if the AI product sends a referrer. Server logs, by contrast, reveal which AI bots collect your pages. Cloudflare Radar data for July 2025 put the crawl-to-referral gap in stark terms, at about 38,000 crawls per referred visit for Anthropic, about 1,100 for OpenAI and about 195 for Perplexity. Treat a crawl as activity, never as traffic. The ratios are over a year old and moved sharply month to month, so check the live page before quoting them.

Our view: first-party reports before a paid tracker

Many teams buy a tracker first because it produces a dashboard immediately. We would start the other way. The Search Console and Bing reports are free, come from the platforms that generate the answers and carry no sampling opinion of a vendor. A spreadsheet with twenty buyer prompts, run three times a week on two engines, covers what those reports cannot: how rivals are described and whether the description is correct. Do that for a quarter before paying anyone.

The cost of skipping this order is real. A tracker bought early locks in somebody else's prompt set and denominator, and the team learns the vendor's definitions instead of its own question. A category with dozens of competitors and several engines to cover is different, because manual runs stop being feasible. In that case buy, but demand the prompt list, the run counts and the formulas in writing. The spreadsheet playbook and seven trial questions support the choice.

Tying visibility to the business

Being named is an early signal, not a result, so connect it to figures you already hold. Offer an AI answer in your "how did you hear about us" question and ask on discovery calls. Branded search and direct visits often rise with awareness; in a 75,000-brand sample, Ahrefs found branded search volume tracking AI Overview visibility at 0.392, a correlation and not a cause. Set pipeline on deals where the buyer names an AI source beside the rest, allowing for the small sample.

Clicks alone understate presence. In Pew Research Center's 2025 browsing study, users clicked a result link in 8% of visits with a Google AI summary and 15% without one. Alternatives to counting clicks are in zero-click search, and the Pew trigger data shows which queries produce summaries.

One week of arithmetic, with invented data

These figures are fabricated to show the sums and are not a benchmark. Imagine 30 prompts on 2 engines, each run once in the week, which gives 60 answers. The brand is named in 21 of them, so the mention rate is 35 in 100. Its pages are linked in 9, so the citation rate is 15 in 100. Across the 60 answers every tracked brand is named 140 times in total, and this brand accounts for 21 of those, so its share of all mentions is 15 in 100. Reading the 21 answers that name it shows 5 quoting a price from a retired plan, so accuracy is 16 of 21.

The distance between 35 and 15 says assistants know the brand but seldom use its pages as evidence. The five wrong answers become a repair list: correct the pricing page, check directories that repeat the retired plan, and rerun in a fortnight. Each metric prompts its own action, so they are never merged into one score.

Where the numbers go wrong

Logged-in sessions bend results, so use clean ones when the tool permits. Changing prompts partway through a quarter makes a trend meaningless. Where you sit inside an answer counts, first or last. A mention that states a wrong fact hurts. Crawls are not readers, and Bing citations are not clicks. Tools sample in their own ways, so track one tool over time instead of comparing two. And benchmark tiers from measurement vendors, such as AirOps' early, competitive and leading bands, are one seller's rules of thumb; our benchmark claim check explains why.

Reading a drop without panicking

A fall in any of these numbers has more explanations than a fall in rankings. The model behind an assistant may have changed. The assistant may have started drawing on different sources. A competitor may have published something stronger. Your own prompt wording may have drifted. Or the movement may be nothing more than the run-to-run variation described earlier.

Check in that order: first confirm the prompts and engines were unchanged, then rerun the affected prompts a few times, then read the actual answers to see what replaced you, then look at whether a page you rely on changed or went missing. Only after that decide whether the drop deserves a response. Bing itself notes that citation trends can reflect model changes, freshness signals and shifts in user demand, which is a reminder that the cause is often outside your control.

Questions a board tends to ask

Expect three. Is this growing? Answer with the window average against the prior period, plus the range. Are we ahead of competitors? Answer with share of voice on the fixed rival list, and say the list. Does it make money? Answer honestly: you can show self-reported sources, branded search movement and a handful of deals, and you cannot yet prove cause. Saying where the evidence stops does more for your credibility than a confident number that the first follow-up question undoes.

Two quarters, planned

Quarter one: fortnight one builds the prompt set, fortnight two takes a baseline, and the remaining eight weeks hold weekly runs plus a monthly summary while the three worst accuracy errors get fixed. At the end you can give mention rate, citation rate and share of voice as ranges. Quarter two: introduce the self-reported source field, revisit the prompt set once and try a small test, refreshing five pages and comparing their citation rate with five left alone. Five pages produce a signal and not proof, so state cause cautiously.

Who owns the report

Name one owner. In a small team that is the content or SEO lead, in a larger one often marketing operations. The owner keeps the prompt set, runs the checks, writes the summary and sends accuracy errors to the people who can fix them: product marketing for pricing and positioning, web for technical faults, PR for third-party descriptions. Without an owner the report becomes a quarterly scramble. The metrics by funnel stage piece shows where these numbers sit beside your other measures.

A summary a board can read

Lead with mention rate and share of voice as a four-week average against the prior period. Add three dated example answers, a list of wrong claims with the action on each, a line on source health from Search Console, Bing, crawl logs and referral sessions, and two next actions with one test. State the prompt set, engines, run count and window at the top. A modest report with those facts is more credible than a confident one without them. Content Agents helps with the content side of the loop; the measuring remains with your analytics and your judgment.