Which AI Visibility Metrics Matter?

Track visibility % (the share of answers that name you), share of voice against the brands AI names, and how often your own pages are cited and retrieved.

Share
Three drawings side by side: answers with a dot where your brand is named, a pie with your share of voice, and a grid of pages with the cited ones filled in.

The AI visibility metrics that matter are visibility % (the share of answers that name you), share of voice against the brands AI names, and how often your own pages are cited and retrieved.

AI visibility is how often AI engines such as ChatGPT, Gemini and Perplexity name your brand when buyers ask the questions you care about. The work of earning it is called AEO (answer engine optimization) or GEO (generative engine optimization), two names for the same work.

Why does AI search need its own metrics, not traffic and rankings?

Say a buyer asks ChatGPT for the best alternative to their current CRM for a small sales team. Your organic traffic cannot tell you whether the answer named you, because the buyer may not click any page it links.

Rankings cannot tell you either: an AI answer has no positions 1 to 10, and the same question may get a different answer the next time it is asked.

When you explain to colleagues why traffic and rankings are not enough, place AI visibility between performance marketing and brand marketing. Do not expect an exact dollar value per point, as you would from paid search. Do not treat it as unmeasurable either, as brand work often is. Keep your SEO numbers (rankings, organic traffic, Search Console impressions, branded search) next to your AI metrics.

What are the key KPIs to track for GEO?

The first four metrics below are the core set to track. The last three describe the quality of a mention. All seven replace organic traffic and rankings as the measure of content success in AI search.

Metric What it counts What it tells you
Visibility % Answers that name your brand, as a share of all answers to your visibility prompts (prompts that ask for a solution without naming you) Whether AI answers name you at all: the headline number
Share of voice Your brand mentions as a share of all brand mentions in those answers, yours and your competitors' How much of the conversation is yours
Your pages cited A page is cited when an answer shows it as a source. This counts answers that cite at least one of your own pages Whether engines use your pages as evidence
Your pages retrieved Answers where the engine pulled one of your pages into the pool it gathers before writing Whether engines find your pages at all
Recommended Answers that name you as the best fit, not only as one name in a list Whether answers point buyers to you
Position Where your name appears among the brands an answer names, for example second of five Your order in the answer, once your visibility is decent
Sentiment Whether answers describe you positively, neutrally or negatively Whether the words match your positioning

How do you calculate visibility % and share of voice in AI answers?

Visibility % is also called mention rate. To calculate it, divide the answers that name your brand by all answers to your visibility prompts, pooled across every prompt, run, engine and country you track.

Say you run 10 prompts 10 times each on one engine. If 20 of the 100 answers name you, your visibility % is 20%. You need repeated runs, because a single run of a prompt reads 0% or 100%.

We looked at ChatGPT data from the Zadoosh data platform, April to October 2026, B2B brands across software and other fields. Take prompts where the brand is named in some answers but not all. With 1 run, the typical miss against the prompt's 30-day share was 18 points, and with 10 runs it was 8 points. So plan on about ten runs per prompt, and read the result as a rough figure.

To calculate share of voice, count the answers that name each brand. In the 100 answers above, say two competitors are named in 40 and 20. That makes 80 brand mentions, so your 20 give you a share of voice of 25%. There is no standard definition yet, so check whether your tool counts one mention per answer or every time a name appears.

Two bars from the same 100 answers: 20 of 100 answers name you, so visibility % is 20%, while your 20 of 80 brand mentions give you a 25% share of voice.

Share of voice is not enough on its own, so read it next to visibility %. If answers rarely name any brand, you can hold a large share of very few mentions.

Compare yourself with the brands the answers actually name, not the rivals you assume matter. If rivals you expected never show up, your prompts may be too generic.

Does your brand's position in an AI answer matter?

It matters once your visibility is decent. But being named first in one answer says little, because the order changes from answer to answer.

So track your average position over many answers. Also check what your tool counts. If it orders only the brands you track, "third" means third of your tracked set, not third in the market.

No. Cited, mentioned and recommended are separate numbers because of how an engine builds an answer. It works out what the buyer wants and may run more than one search. Google calls this query fan-out.

From the pages that come back, it keeps a pool it judges useful, say 30. These are retrieved. It cites only the few with enough credible information, say 7 or 8. While it writes, it names the brands worth naming, and may single one out as the best fit.

A wireframe of one AI answer: your brand called the best fit is recommended, brand names in the text are mentioned, numbered links inside the answer are cited, and the other links in the sources list were retrieved but not cited.

So an engine can cite your page for a fact and still name other brands. And a mention can come from any page in the pool, or from what the model learned in training.

We looked at answers that showed sources, in data from the Zadoosh data platform, B2B brands across software and other fields. In ChatGPT, 8.1% of answers named the brand without citing one of its pages, and 5.2% cited one of its pages without naming it (April to October 2026). AI Mode went the other way: 5.8% and 17.5% (June to October 2026).

Across all engines in the same data, April to October 2026, third-party pages (listicles, review articles, etc.) were 80% to 86% of the pages AI answers cited. The brand's own pages were about 6%. The range depends on whether review sites count as third-party. So third-party pages that name you feed many of your mentions.

Recommended is the goal, and being cited is usually the way in. Engines use the pages they cite to decide which brands to name.

Your own pages do not have to be cited first for you to be named. But they are the part you control, so if they are not cited yet, start there. Once they are, work on being recommended.

Do AI visibility tools count retrieved pages as citations?

Some do. A tool can store the pages an engine gathered but did not cite next to the cited ones, so its "cited" partly means retrieved. And a "citation rate" that counts how often your brand is named is your visibility % under another name. So check what your tool calls a citation, and track mentions, citations and retrieval as three numbers.

In ChatGPT you can check by hand: links inside the answer are the cited pages, and the Sources panel shows those plus other relevant links. The other links are pages it retrieved but did not cite. Most engines show a similar list. Server logs cannot show retrieval: a bot visit shows that a page was fetched, not that an answer used it.

How do you measure sentiment and recommendation quality in AI answers?

Sentiment belongs in a first report only when positioning is one of your goals. Most teams first want to know they show up, and care how AI describes them later. Once positioning matters, track sentiment weekly, next to visibility.

A premium product named in a list of the cheapest options still counts as a mention, so its visibility % goes up. But a mention with the wrong words does not drive demand, because visitors arrive expecting something else.

To spot a drift in how AI positions you, collect the phrases it uses about you across answers and compare them week by week. Three branded prompts do this job: "What is [brand]?", "[brand] pros and cons" and "Is [brand] good for [use case]?".

Once a month, and after any change, check the facts AI states about you: pricing, features, category and who you are for. Fix the pages that carry old facts. If a tool scores sentiment for you, spot-check a few scored answers each month.

How do you measure AI visibility by topic, by prompt and in AI Overviews?

Group your prompts by topic and read each group's visibility % on its own. One pooled number can hide a topic where you never appear.

Prompt coverage, the share of your prompts where you appear at least once, is a weak substitute for visibility %. It counts a prompt where you appear in 1 answer of 30 the same as one where you appear in all 30. As support, the share of days a page was seen cited, or the days since it was last cited, shows which pages still work.

If a new product has zero share of voice, mentions are the wrong first measure. Watch the pages you created or worked on instead: do they start being retrieved, then cited?

AI Overviews are the AI summaries Google places at the top of some search results, with links to the pages they draw on. Track them apart from AI Mode: AI Overviews works more like next-generation Google Search, and AI Mode more like ChatGPT.

Since August 2026, every site has a generative AI performance report in Search Console, showing how often links to its pages appeared in AI Overviews and AI Mode. It shows no queries, and it does not show whether your brand was named.

Should you track AI referral traffic and conversions, or only mentions?

Track both, but give them different jobs. AI referral traffic is the visits that reach your site from links in AI answers. It undercounts AI's real influence, because buyers rarely click the pages an answer cites. Leads who found you in ChatGPT often show up as Google, branded search or direct traffic.

So never count a mention as a visit. Treat referral traffic and its conversions as a directional signal, and your visibility against competitors as the early one.

To link AI to pipeline, add a "How did you hear about us?" field to your forms, and ask it on every demo call. Then watch whether the replies that name ChatGPT, Gemini or another engine rise over the quarter. Report that trend, not a precise ROI multiple. AI visibility is narrower than brand marketing, so a jump in revenue can still be attributed to it.

How do you set up AI visibility tracking: prompts, engines, app or API?

Visibility prompts come in four kinds: a category prompt ("best CRM software"), a category with context, an alternatives prompt and a full buyer situation. A prompt that names your brand, or sets you against a competitor, forces your name into the answer, so it stays out of visibility %. Informational prompts ("What is [category]?") and prompts that compare two competitors stay out too.

Track each engine on its own. A pooled number leans on the engine with the most answers, so a drop in one engine can hide behind another.

Run each prompt once a day per engine and per country, and decide on the trend over weeks. A weekly check gives each prompt only four or five answers a month. That is enough for a first audit, but to follow a trend you need about ten or more. On a tight budget, cut engines first, then prompts, and keep the daily runs.

Tools collect answers either from the engine's app, where buyers see them, or from its API, the interface developers use. In OpenAI's API, web search is a tool the request turns on, and the model still decides whether to search. Perplexity's own FAQ says its API and web results can differ. So prefer app answers, and use API answers only as a cheap supplement, checked against the app.

No tool sees what a real buyer saw, ours included. A buyer's answer can depend on their account and location, and ChatGPT may use saved memories when it rewrites a search. Tools run your chosen prompts as neutrally as they can, to show what the engines answer today.

Before you trust a tool, run a few prompts by hand, logged out, in ChatGPT and the Gemini app. Use only tools that show the raw answer text behind every number.

What should an AI visibility report include, and how do you present it to executives?

Give AI visibility its own section and KPI next to your SEO numbers, every month, with a short quarterly summary for the CFO. Set directional targets, such as rising visibility and a closing gap to your top competitors over the quarter, rather than a fixed percentage.

Report section What it shows
Visibility % and share of voice by engine, ranked against competitors Where you stand, and whether the gap to the leaders is closing
Citations by channel: your own site, community pages (Reddit threads, forum posts, etc.) and third-party pages (listicles, review articles, etc.) Which kinds of pages the engines draw on, so you can see why the number moves
The pages you created or worked on: retrieved and cited The earliest clear sign that your work is paying off
Sentiment and the phrases AI uses about you How AI positions you, if that is one of your goals

Start with a baseline: your starting numbers, recorded before you change anything. For a first audit, choose your buyer prompts, engines and countries, and collect 4 to 5 answers per prompt, engine and country over at least a few days.

In every report, say plainly that each figure is a sample of your tracked prompts. Citation counts are a floor, because they exist only on those prompts.

Which AI visibility metrics should you stop trusting?

Stop trusting any single check. The most common mistake is to run a prompt once, often logged in, and treat that answer as the truth.

We looked at ChatGPT data from the Zadoosh data platform, April to October 2026, B2B brands across software and other fields. Two runs of the same prompt on the same day shared 53% of the brands named, 31% of the cited sites and 22% of the cited pages. Brands stayed steadier than cited sources, so read citations as rates over many runs too.

Stop trusting daily numbers and raw counts. A single day can carry too few answers to mean anything. A citation count rises and falls with the number of answers collected, so compare rates, such as the share of answers that cite your pages.

If the numbers look wrong, check the prompts before the metrics. A common mistake is prompts that are too simple, or too informational for a buyer to ask.

When two tools disagree, compare what each one does differently: prompts, engines, country, number of runs and how it counts a mention. Then pick one and read its trend.

An AI visibility score is a single number that sums up how visible a brand is in AI answers. Use it as a signal, not as the goal itself. Its value is in what it shows: where you get onto a buyer's shortlist, which pages earn you that place, and where you are missing.

How do you prove a change in AI visibility is real and not noise?

You can claim a change worked when the move is bigger than the noise, you gain on your competitors, and the pages you worked on start being retrieved and cited. For a client, back it up with a few prompts run by hand, logged out, so the proof does not rest on the tool alone.

How big does a weekly change in AI visibility need to be?

Say your visibility % moves from 22% to 25% in one week. On its own, that move is probably noise.

We looked at ChatGPT data from the Zadoosh data platform, March to October 2026, B2B brands across software and other fields. Most week-to-week moves in a prompt set's visibility were noise, and the median move was about 3 points.

A weekly change was at least 80% likely to be real from about 11 points across all prompt sets. On sets with 100 to 299 answers a week, the bar was about 7 points. Under 100 answers a week, no weekly change could be told from noise.

Comparing one 4-week block with the next, changes of 3 to 6 points were already likely real. Engines where you collect fewer answers need a bigger move.

So act on a weekly change of about 10 points or more when the set has 100 or more answers a week, and read smaller moves over a month. As a guide, 15 prompts run once a day on one engine give just over 100 answers a week.

A three-way check for a weekly change in AI visibility: under 100 answers a week no change can be told from noise, with 100 or more act on about 10 points or more, and read smaller moves over a month.

What is AI visibility momentum, and how do you measure it?

AI visibility momentum is the direction of your visibility against your top three competitors on the same prompts, read month by month over about three months. Momentum is positive when your number rises and the gap to the leader closes.

Say you go from 18% to 24% over three months, while the leader stays at 30%. The gap closed from 12 points to 6, so your momentum is positive. If the leader had climbed to 38%, the gap would have grown to 14 points. You would be losing ground, even though your number rose.

This is also why no score is good or bad on its own. A number means something only next to the brands that compete for the same answers.

How do you connect your actions to changes in AI visibility?

Keep a dated list of your actions, such as pages you created or changed on your own site, a third-party article or a community thread you joined. Then check whether each page starts being retrieved, then cited, for your prompts, and whether mentions follow. If nothing you did is retrieved or cited, the change in your AI results came from somewhere else. After a rewrite, also read the rewritten pages' rankings and Search Console impressions to judge the change in search.

For a before-and-after report, compare the share of cited pages from each channel before and after. The channel whose share grew is the likely cause. If you added prompts in between, also report the number for the original prompt set, so you compare like with like.

Frequently asked questions

What does 50% share of voice mean? Half of all brand mentions in your tracked answers are yours. It does not mean you appear in half the answers, which is what visibility % measures. If the answers name only you and one rival, 50% means you two appear equally often.

What is a good share of voice percentage? There is no universal good share, because it depends on how many brands the answers name. Compare yours with your top three competitors on the same prompts, and look for a rising share over about three months.

What is an AI visibility report? An AI visibility report shows how often AI answers name you compared with your competitors, and which kinds of pages they cite. Build it from the same prompts every month, so one month compares with the next.

Which AI visibility metrics can you check for free? By hand, you can get a sense of visibility % and citations for a few prompts. Run each one several times, logged out, and count the answers that name you and the pages they link. For daily numbers across engines, the Zadoosh AEO platform's free trial tracks 25 prompts for 7 days, with no card, as of November 2026. It covers ChatGPT, Gemini, Perplexity, Copilot, AI Overviews and AI Mode, in the countries you choose.