Most businesses check their AI visibility the wrong way. They type their own company name into ChatGPT, get a flattering paragraph back, and conclude they are fine. That proves nothing — nobody who already knows your name is a new customer.
The question that matters is different. When a stranger asks an assistant the question your service answers, how often does your name come out, and how often does a rival’s come out instead? That ratio is your AI share of voice.
Malaysia makes this urgent rather than academic. DataReportal’s Digital 2026: Malaysia report counted 35.4 million internet users at the end of 2025, at 98.0% penetration. The whole buying public is online, and a growing slice now asks a model instead of scrolling ten blue links.
This guide covers how to define the category, build the prompt set, score the answers and read the result — plus what to do when a rival owns the answer you want. For wider context, see our SEO services or the ZenWeb home page. The video below is a useful primer first.
Source video: How to Track Your Brand in AI (ChatGPT, Gemini) | AI Visibility Dashboard Explained
Quick Answer: AI share of voice is your brand’s mentions divided by all brand mentions across a fixed set of category prompts, as a percentage. If a 40-prompt set returns 90 brand mentions and 18 are yours, your AI share of voice is 20%.
The metric borrows its logic from old media share of voice, where an advertiser measured its slice of total category advertising. The difference is that you cannot buy this one — nobody sells placement inside a ChatGPT answer. Your share is a by-product of how quotable your pages are and how consistently you are described elsewhere.
Three numbers make up the picture, and they are not the same thing:
An honest measurement records all three. Reporting only the first is how thin work is made to look impressive — test for it when you evaluate an AI SEO agency in Malaysia.
Quick Answer: A mention count rises whenever answers get longer, so it can climb while your position falls. Share of voice divides by the total, so it stays honest when engines list eight brands instead of three.
Picture two months of the same 40-prompt set. In month one you are named 12 times out of 40 total mentions — a 30% share. In month two you are named 15 times, which looks like progress, but the engines have grown chattier and the set now returns 75 mentions. Your share is 20%. The count went up; your position went down.
That is why the denominator matters more than most reports admit. It also protects you from the reverse error: a month when your mentions dip but two rivals vanish is a month you won.
A mention count tells you that you exist. A share tells you whether you are winning.
The same discipline applies in ordinary search, where nobody accepts “we rank for more keywords” without asking which ones and against whom. Treat SEO competitor analysis as the template, applied to answers instead of results pages.
Not sure where you sit in your category yet?
A first measurement gives you the baseline everything else is judged against. See what an AI visibility audit covers →
Quick Answer: Use 30 to 50 prompts, written the way a buyer actually types, and freeze them. The prompt set defines your category — change it mid-programme and every comparison against last month becomes meaningless.
The prompt set is the most consequential decision here. Too narrow and you report a flattering share of a category nobody searches. Too broad and you drown among national brands you never compete with.
A workable set covers four clusters:
Write them in the phrasing your market uses, Malaysian city names included. Then freeze the set. Frozen prompts make the second measurement comparable to the first, and they are the asset that powers ongoing AI citation tracking.
Quick Answer: Record four things per answer: which brands were named, in what order, which sources were cited, and whether your mention was positive, neutral or negative. That is enough to compute AI share of voice and explain it.
Scoring is clerical work, and that is a feature. The moment it needs judgement, two people scoring the same answer disagree and the trend line becomes noise. Keep the fields mechanical:
Run each prompt in a clean session with no history, or the model answers based on who you are rather than what you asked. Repeat each prompt two or three times per engine and take the majority result. If the volume makes you wince, that is the honest argument for a monthly AI SEO retainer.
Quick Answer: The same prompt set produces very different shares on different engines, because each one weighs live web results and stored knowledge differently. A single blended number hides the gap that tells you what to fix.
| Engine | Median share | % | Brands named per answer |
|---|---|---|---|
| Perplexity | 21% | 4.8 | |
| Google AI features | 17% | 5.6 | |
| Microsoft Copilot | 14% | 4.1 | |
| ChatGPT | 11% | 3.4 | |
| Gemini | 9% | 3.9 |
Source: ZenWeb client tracking, Malaysia, 2024–2026. Licence.
The pattern is consistent enough to plan around. Engines leaning on live retrieval reward fresh, well-structured pages. Engines leaning on stored knowledge reward businesses described consistently across the web for years. A brand strong on one and weak on the other has a corroboration problem, not a content problem.
Quick Answer: Most mentions trace back to a page the brand does not own. Logging cited sources alongside brand names turns AI share of voice from a scoreboard into a work list — it names the pages you need to appear on.
| Source type | Share of citations | Brand controls it? |
|---|---|---|
| Brand’s own website | 31% | Fully |
| Listings and directories | 22% | Partly |
| Editorial and trade press | 18% | No |
| Forums and communities | 16% | No |
| Video and social | 13% | Partly |
Source: ZenWeb client sample, n=500+, 2024–2026. Licence.
Read the right-hand column carefully. Two-thirds of what feeds an answer sits on a page you do not own — which is why business directory citations and consistent off-site descriptions are now a ranking activity, not an admin chore.
Quick Answer: Enquiries from AI referrals rise with share of voice, but not linearly. Moving from a low band to a mid band changes little; crossing into the top band, where you are named first, is where enquiry volume moves.
| Share of voice band | Micro (1–10 staff) | Small (11–50) | Mid (51–200) |
|---|---|---|---|
| Under 5% | 0–2 | 1–3 | 2–5 |
| 5–14% | 2–5 | 4–9 | 7–14 |
| 15–29% | 5–11 | 10–22 | 18–38 |
| 30%+ | 9–19 | 20–41 | 36–72 |
Source: ZenWeb operational data, Malaysian SME campaigns, 2024–2026. Licence.
Two things follow. Early on, judge the programme on share movement, not enquiry counts, because enquiries lag. And the target is not “be mentioned” but “be mentioned first” — the substance behind AI answer optimisation.
Quick Answer: Expect roughly a point of AI share of voice per month once corroboration work starts, with the first two months flat. Programmes that only publish on-site content stall around month three; those that also build off-site mentions keep climbing.
| Programme | M1 | M2 | M3 | M4 | M5 | M6 |
|---|---|---|---|---|---|---|
| On-site content only | 6% | 6% | 8% | 9% | 9% | 10% |
| Content plus off-site work | 6% | 7% | 10% | 13% | 15% | 18% |
Source: ZenWeb client tracking across 12 industries, 2024–2026. Licence.
Flat opening months are normal, not failure. Engines re-crawl slowly and stored knowledge updates slower still. Judging a programme at week six is the commonest reason good work gets cancelled — settle it in writing before any generative engine optimisation services engagement begins.
Ready to put these benchmarks to work?
We measure your baseline first, then build the loop that moves it. Compare AI SEO services →
Quick Answer: Monthly is right for most Malaysian SMEs. Weekly produces noise you will misread as trend; quarterly is too slow to catch a rival’s move. Treat under three points of change as normal variation.
Answers vary between runs even when nothing has changed, so one measurement is a snapshot with an error bar. Three rules keep the reading honest:
In-house, the discipline breaks under workload rather than skill — the case for an outside SEO consultant owning the cadence. Sites with hundreds of category pages need different sampling, which is enterprise SEO in Malaysia territory; agencies working through white label SEO standardise the set across clients for the same reason.
Quick Answer: Work backwards from the sources the answer cites, not the rival’s website. If three pages keep feeding the answer, your job is to belong on those pages, or on equivalents the engine trusts.
The instinct is to study the rival’s site and copy it. That rarely works, because their site is only about a third of what the engine reads. The cited-sources column is the better map. A sensible sequence:
The order matters. Businesses that start at step four and never reach step one publish a lot and move very little. It is the clearest difference between AI SEO vs traditional SEO: classic SEO was often won on your own domain, while answers are won across the wider web. The on-page half is covered in our guides to showing up in ChatGPT answers and ranking in Google AI Overviews.
Quick Answer: Most broken AI share of voice reports fail on method, not effort. Logged-in sessions, drifting prompts, branded queries and a single engine are the errors that turn months of work into an unusable number.
The failures worth checking before you trust any report, in-house or from a supplier:
Any supplier who cannot show the raw scoring sheet is asking you to take the number on trust. Ask to see it. The same applies to the answer-formatting side, which an answer engine optimisation agency should evidence page by page — alongside the basics in our guide to getting mentioned in Perplexity.
AI share of voice is the closest thing AI search has to a market-share figure. It is harder to fake than a mention count, it names your rivals, and it points at the pages deciding the answer.
The method is simple: a frozen prompt set, five engines, four scored fields, monthly repetition, and the discipline to record what your rivals are doing too. The hard part is doing it the same way for six months.
If you want the baseline measured properly and the work planned around real numbers, our SEO services team runs this loop for Malaysian businesses.
It depends on how crowded the category is. In a narrow local category with four or five real players, 20% and above is strong. In a broad national category naming eight brands per answer, anything above 10% is competitive. The better test is direction: are you gaining on your named rivals?
Thirty to fifty is the practical range for most SMEs. Below thirty, one unusual answer swings the percentage too much. Above fifty, the workload grows faster than the accuracy. Split them across category, problem, comparison and qualifier prompts, then keep the set unchanged.
Yes. A spreadsheet with one row per prompt-and-engine pair and columns for brands named, order, sources cited and sentiment does the job. For a 40-prompt set across five engines, manual measurement takes a focused day each month.
Plan for six months. The first two usually look flat while engines re-crawl and stored knowledge catches up. From month three, programmes combining on-site work with off-site corroboration typically gain about two points a month; on-site-only programmes tend to flatten around month four.
Want to know where you stand in AI answers?
Book a free 30-minute strategy session — we’ll measure your baseline against your three closest rivals, show you the sources driving their mentions, and give you a 90-day plan to close the gap.
Complete the form and our team will contact you to discuss your goals. Let’s grow your business.
Online
