Definition
What is AI search visibility?
AI search visibility is the percentage of AI answers, across a defined set of prompts and engines, that mention a brand at least once. A brand named in 30 of 100 answers has a visibility of 30. The number tells you whether buyers who ask an AI engine for advice hear your name at all, before any question of rank or tone.
Formula
visibility = responses mentioning the entity ÷ in-scope responses × 100 The score matters because the answer is often where the search ends. In a study of browsing data from March 2025, Pew Research Center found that Google users who saw an AI summary clicked a traditional result on 8% of visits, against 15% when no summary appeared[1]. A brand left out of the answer loses part of the traffic that used to come from the results below it.
Every visibility score then rests on three choices, and changing any of them changes the number.
The prompt set comes first. Visibility is always measured on something, for example 120 questions a buyer might ask about project management software, written in the words buyers use. Swap a few prompts and the score moves, so the set stays fixed between two measurements.
The engine comes second. ChatGPT, Perplexity, Google AI Mode, Claude, Microsoft Copilot and Grok each search and write in their own way, and a brand can be named often on one and rarely on another. Report visibility per engine first, then blend if you need a single figure.
The third choice is what counts as a mention. Loopwork, Loopwork PM and loopwork.io are one brand, and counting them apart splits one score into three smaller ones. Merge the aliases into a single entity before you count.
Worked example: 18 answers out of 40 give Loopwork a visibility of 45
Sample data on fictional brands: 10 prompts on 4 engines, US / en.
Comparison
AI search visibility is 1 of 4 metrics, and each answers its own question
AI search visibility tells you whether a brand is named. Share of voice tells you how much of the conversation it holds, average position how early it comes up, and true reach whether the pages engines cite talk about it. Read together, the four explain movements that any one of them alone would misread.
| Metric | Question it answers | Formula | Unit | Use it when |
|---|---|---|---|---|
| Visibility | Is the brand named at all? | responses mentioning the entity ÷ in-scope responses × 100 | %, higher is better | You track presence over time |
| Share of voice | How much of the conversation does it hold? | entity mentions ÷ all entity mentions in scope | %, all brands add up to 100 | You compare the brand with named competitors |
| Average position | How early is it named? | mean rank of the entity where it appears | Rank, lower is better | Answers list several options and order matters |
| True reach | Do the cited pages mention it? | cited pages that mention the brand ÷ all cited pages | %, higher is better | You plan content and PR on other sites |
Take one sample answer to the prompt “best project management tool for creative agencies”. It recommends Cadence first, then Northbeam, then Loopwork, and names Cadence twice more while comparing features. The engine cites three pages, and none of them mention Loopwork.
On that answer, Loopwork is visible, since it is named. Its share of voice is one mention out of five, or 20%. Its position is 3, and its true reach is zero.
Each number calls for a different response. Full visibility with a low share of voice means the answers are crowded, and the work is to give engines more to say about the brand. A poor position points at how the brand compares on the criteria the prompt names. A true reach of zero says the pages the engine trusts on the topic leave the brand out, which is a job for PR and partnerships more than for your own site.
Research uses its own variants. The 2023 paper that introduced generative engine optimization scored the visibility of each cited source by the share of the answer's words in sentences that cite it, discounted by position[2]. Brand teams usually start with the simpler question of whether the brand is named, which is the definition used in the glossary of GEO terms[3].
Engines
The 6 AI engines answer from different sources, so visibility is read per engine
AI search visibility is read per engine because each engine searches the web its own way and shows its sources in its own format. Google describes AI Mode as running several related searches for one question, while ChatGPT search and Perplexity link to the pages behind each answer. One brand can be named by one engine and missed by another on the same prompt.
Google says AI Overviews and AI Mode may use a technique it calls query fan-out: several related searches across subtopics and data sources, run to build one response, and that this surfaces a wider and more varied set of supporting links than a classic search[4]. The pages those searches return are the pool the answer draws from, and an engine running different searches draws from a different pool.
ChatGPT search answers with links to its sources and a Sources button that lists them[5], and Perplexity gives every answer numbered citations that link to the original pages[6]. The format changes what a reader sees, but the mechanism is the same: the brands in the answer are, for the most part, the brands in the pages retrieved.
For creative agencies, Cadence is the safest pick: client approvals are built in and proofing works on PDFs and video. Northbeam suits smaller studios that want time tracking in the same tool. Loopwork is worth a look if your team already plans in boards and wants a lighter setup.
In this example, all three brands count as visible. Ask the same prompt on another engine and Loopwork may drop out while a fourth brand comes in. A blended score would show a small dip; a per-engine count shows exactly where the brand was lost, which tells the team which sources to work on. For a first check on one prompt, the AI visibility tool compares your brand with up to three competitors on six engines.
Method
AI search visibility is measured in 4 steps, and the first 3 need no software
AI search visibility is measured with a fixed set of prompts, a run on each engine, a count of the answers that name the brand and a way to repeat all of it unchanged. The first three steps can be done by hand on a small set. Past a few dozen prompts, the count is where manual tracking stops keeping up.
-
Step 1
Write a fixed prompt set
List the questions a buyer asks an AI engine before choosing a product like yours, in the words they use. Aim for 80 to 300 prompts per market and freeze the list once measurement starts.
-
Step 2
Run every prompt on each engine
Ask each prompt on every engine you care about, in the country and language of the market, and save the full answer. Run each prompt more than once, because the same engine can answer differently from one run to the next.
-
Step 3
Count the answers that name the brand
Mark each answer that names the brand or one of its aliases, then divide by the number of answers. Do it per engine before you blend anything.
-
Step 4
Automate the runs and the count
Repeat the same prompts on the same engines at a fixed interval, so a change in the score reflects a change in the answers. The AI Visibility API returns visibility, share of voice and average position per entity in one request. See the AI Visibility API.
Limits
AI search visibility misleads in 4 situations, starting with small prompt sets
AI search visibility is a ratio of answers, so it says nothing about how a brand is described or where it sits in the answer. A brand can raise its visibility while being named last, or named as the option to avoid. Read it next to share of voice, position and sentiment before acting on a change.
- Small prompt sets swing. With 20 prompts on one engine, a single answer moves the score by 5 points.
- Warnings count as mentions. An answer that tells agencies to avoid Northbeam still counts as a mention of Northbeam.
- One run is a sample of one. The same prompt can return a different answer on the next run, so a single run can pass noise off as a trend.
- Blended scores hide engines. An average across six engines can hold steady while one of them stops naming the brand.