Most brands cannot answer that question today. They can report a Google ranking and a traffic figure, and neither describes what an AI assistant says when a buyer asks it who to hire. This guide sets out the measurement model, the prompt set discipline behind it, and the one metric almost every brand leaves off the dashboard.
What is brand visibility in AI search?
Brand visibility in AI search, often called AI share of voice, is the extent to which a brand is named, cited or recommended in answers generated by systems such as ChatGPT, Perplexity, Gemini and Google's AI Overviews. It measures presence inside a synthesised answer. It does not measure position on a results page.
That distinction is not academic. Traditional search returns a ranked list, so visibility could be expressed as a rank. A generative engine returns a single answer, assembled from sources it selects and paraphrases. There is no position three. A brand is either in the answer or it is not, and if it is in the answer it is either described correctly or it is not.
This is where most measurement stops too early. Visibility in AI search has two components that traditional search never had to separate: whether you appear, and whether what is said about you is true. A brand can hold a healthy mention rate and still be losing, because the engine is confidently describing it as something it is not.
The research community formalised the first half of the problem in 2024. The paper that introduced generative engine optimisation, presented at KDD by Aggarwal and colleagues, proposed a set of visibility metrics designed specifically for generative engines and showed that content optimised against them could lift visibility by up to 40 percent. The metrics matter more than the tactics. Without them there is nothing to optimise against.
Rank is no longer the unit of measurement
Two findings make the case better than any argument.
Pew Research Center tracked the browsing of 900 US adults across 68,879 Google searches in March 2025. Where an AI summary appeared, users clicked a search result in 8 percent of visits. Where none appeared, they clicked in 15 percent, nearly twice as often. Clicks on a link inside the AI summary itself happened in 1 percent of visits. Users also ended their session entirely on 26 percent of pages carrying an AI summary, against 16 percent of ordinary results pages.
Separately, Semrush's 2025 study of AI search found that when ChatGPT cites a webpage, that page sits in traditional organic positions 21 or lower for related queries almost 90 percent of the time. Ranking first is not the qualification it once was. Being quotable is.
Put the two together and the implication is blunt. The click is no longer the event worth counting, because in most AI mediated searches there is no click. The mention is the event. A dashboard still reporting keyword rankings and organic sessions as its visibility metrics is measuring a shrinking slice of what is happening.
There is a consolation, and it is a substantial one. The same Semrush study found that the average visitor arriving from an AI source converts at 4.4 times the rate of the average organic search visitor, because that visitor has already run the comparison before they arrive. Fewer clicks, better clicks. Which is exactly why the mention deserves a metric of its own.
Most AI visibility measurement fails because the prompt set keeps moving
Most brands measuring AI visibility today are doing it by asking. Someone types the company name into ChatGPT, reads the answer, and reports back. A month later someone types a slightly different question, gets a different answer, and the team concludes either that things are improving or that the model is unreliable. Neither conclusion is supported, because nothing was held constant.
Opinion polling solved this problem seventy years ago. A poll does not measure what people think. It measures what a defined sample says in response to a fixed question, worded the same way, asked at intervals. Change the wording and you change the answer. A pollster who rewrote the question every week would be laughed out of the room. That is how most AI visibility tracking is currently run.
A prompt set that changes every month is not a measurement instrument. It is an anecdote with a spreadsheet attached.
Three failures recur. The first is the moving prompt set: without a fixed, documented list, month on month comparison means nothing. The second is a missing competitor frame. A brand that knows it was named in four answers out of twenty knows nothing until it also knows a rival was named in sixteen. Share requires a denominator. The third is measuring presence and ignoring accuracy, which is the failure that costs the most and is dealt with below.
The fix is not a tool. Tools are useful, and the category is now crowded with them, but a dashboard sitting on top of an undisciplined prompt set produces confident nonsense at speed. The fix is measurement design.
Exhibit 1: The Answer Share Model
The Answer Share Model is the framework Vantage uses to measure a brand's standing in generative answers. It carries four metrics, deliberately few, because a board will act on four numbers and ignore forty.
- Presence rate. The percentage of prompts in the fixed set where the brand is named at all. This is the entry ticket. A brand absent from the answer is absent from the consideration set, whatever its Google ranking says.
- Answer share. Of all brand mentions across the whole set, the percentage that are yours. This is the number that deserves the name AI share of voice, and it is the one most often quoted without a denominator. Presence tells you that you exist. Share tells you whether you are winning.
- Description accuracy. Of the answers that mention the brand, the percentage describing it correctly against a single canonical entity description agreed in advance. This metric obliges the organisation to write that description down, which is a useful exercise on its own.
- Source concentration. The domains the engines actually cite when answering the category prompts, ranked by frequency. This shows where the consensus about your category is being manufactured, and whether you appear anywhere inside it.
Read together, the four answer a question a chief executive can hold in their head. When the market asks an AI which brands solve this problem, how often does it say ours, how often does it say a competitor's, does it get us right, and where is it getting its information?
Exhibit 2: The four prompt classes a measurement set must cover
Twenty five to forty prompts is usually enough for one market and one category. Below twenty, a single answer swings the percentage. Above fifty, the set becomes too expensive to run monthly, and cadence matters more than breadth. Draw the set from four classes.
- Category prompts. "Best brand strategy consultancies in Singapore." These decide whether the brand makes the consideration set at all. They are the most contested and the slowest to move.
- Problem prompts. "Patients do not trust our healthcare brand, who can help." This is how buyers actually talk to an assistant, in symptoms rather than category labels, and it is where specialists beat generalists.
- Comparison prompts. "Consultancy versus in-house team for a rebrand." Distinction questions carry unusually high citation value, because the engine needs a source willing to draw the line clearly.
- Entity prompts. "What is Vantage Branding." These test description accuracy directly, and they are the prompts most brands skip, because the answer is uncomfortable.
Fix the list. Date it. Run it against every engine that matters in the market, on the same day each month, from a clean session. Log the full answer text rather than a yes or no, because the language the engine chooses is itself the finding.
Description accuracy is the metric almost nobody tracks
Here is the uncomfortable part, and it is worth being specific rather than hypothetical.
When Vantage began running its own entity prompts, the engines returned a description that was confidently wrong. Vantage was characterised as an FMCG and packaging specialist. Vantage is a brand research, strategy and identity design consultancy, with unusual depth in healthcare, finance, government and cultural institutions. The mention rate looked respectable. The description was wrong, and a wrong description is worse than no mention, because it is a recommendation delivered to the wrong buyer.
That is not a model error. It is a consensus error. Engines assemble a description from what the web collectively says, and where the web is thin or contradictory they fill the gap with the nearest plausible pattern. The correction is not a prompt trick. It is publishing one consistent entity description everywhere the engines read, and supplying enough corroborated evidence that the correct version becomes the safest one to repeat. That is the argument set out in full in how to get your brand mentioned by ChatGPT.
Which is why description accuracy belongs on the dashboard beside presence and share. A brand that lifts its mention rate while the engines keep misdescribing it has succeeded only in being recommended more often to people who will never buy. The gap between what a brand intends to project and what the market receives is an old problem in a new medium, and it is the same gap a brand audit exists to expose.
Measuring across Southeast Asia has to be done market by market
A single English language prompt set run from Singapore tells a regional brand almost nothing about how it appears in Jakarta or Ho Chi Minh City. Generative engines answer differently in different languages, because the source material available in each language differs. Where the Indonesian language web is thin on a category, the engine leans harder on whatever it can find, and a brand with no Bahasa Indonesia corroboration is simply not in the answer.
This is not a marginal concern in this region. The e-Conomy SEA 2025 report from Google, Temasek and Bain describes AI as redefining the consumer journey from search to discovery, replacing linear search with an AI powered discovery process, and notes that Southeast Asian consumer interest in AI is globally leading. The region's digital economy passed 300 billion US dollars in gross merchandise value in 2025. Behaviour is shifting to AI answers faster here than in most of the world, across ten markets that do not share a language.
The consequence for measurement is that the prompt set must be replicated per market and per language, and the results must never be averaged. An average across six markets hides the one market where the brand is invisible, and that is usually the market that matters. This is the same discipline that separates the brands which travel from the ones that stall, as we argue in why Southeast Asian brands struggle to scale.
What AI visibility measurement costs, and what it replaces
Measurement is the cheapest part of the work. A designed prompt set, a baseline run across the major engines, a competitor frame and a quarterly readout sits at the entry of a brand engagement. Most Singapore branding programmes fall between S$5,000 and S$50,000, with enterprise work higher, and an AI visibility baseline typically sits at the lower end of that band when commissioned on its own. Commissioned alongside an audit or a positioning programme it adds little, because the entity work it depends on is already being done. The full picture is set out in our guide to branding costs in Singapore.
Singapore registered businesses should note that brand strategy work, including audit and positioning, may qualify for the Enterprise Development Grant administered by Enterprise Singapore, which provides up to 50 percent co-funding for qualifying projects.
The better question is what the measurement replaces. It replaces a recurring argument about whether the brand is showing up in AI, conducted by people typing questions into chatbots and drawing conclusions from single answers. That argument costs more than the measurement does, and it never resolves.
When to start measuring, and how often
Baseline before you act, not after. The most common sequencing error is to spend six months publishing content built for AI citation and only then begin measuring, at which point there is no before against which to read the after.
Run the set monthly. Generative engines update continuously, and a quarterly cadence will miss movement a month would have caught. Report quarterly, because a board cannot act on monthly noise. Re-cut the prompt set once a year and no more often, and keep the old set running in parallel for two cycles so the series does not break.
There is one case for not measuring yet, and it is a real one. If a brand has not agreed what it stands for, has no consistent description of itself across its own properties, and has published nothing an engine could quote, measurement will only confirm what it already suspects. Fix the brand first. Then measure it. The wider shift this sits inside is covered in how AI is changing branding.