What 34,794 AI answers reveal about how Asian B2B brands actually show up in ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews

Every AI-visibility vendor publishes a benchmark now. Ahrefs studied 75,000 brands. Semrush analysed 126 million prompts. Both are excellent, and both have the same blind spot: they are overwhelmingly built on US and European brands competing in US and European categories.

Nobody has published the number that matters to a mid-market company in Singapore, Hong Kong, Bangkok or Ho Chi Minh City: when a buyer in my category asks an AI assistant for a recommendation, how often does my name come out?

So we measured it. This is the first edition of the Asia AI Visibility Index, built from 148 B2B brands, 587 clean scans and 34,794 individual prompt-runs across five AI systems between December 2025 and July 2026.

The headline is uncomfortable, and the most useful finding actively works against what our own industry sells.

The cohort

Market composition: 116 brands in confirmed Asian markets (Singapore 47, Hong Kong 24, Thailand 11, Indonesia 10, Vietnam 8, Japan 4, India 4, Malaysia 4, Taiwan 2, China 1, Philippines 1), 7 in Australia, and 25 whose operating market we could not confirm with certainty from first-party signals. Every brand is anonymised throughout. No brand name, domain or per-brand figure appears anywhere in this report.

Two metrics do the work:

• Appearance rate is the share of category prompts where the brand is named in the answer body. This is being recommended.

• Citation rate is the share of prompts where the brand's own domain is used as a source. This is being read.

They are not the same thing, and the gap between them turns out to be one of the more important findings here.

Finding 1: the typical Asian B2B brand is effectively invisible

The median brand in our cohort appeared in 1.71% of the prompts its own buyers would plausibly ask.

Not 17%. 1.71%. Roughly one answer in sixty.

Worse, 29.7% of brands scored exactly zero. Not low, not marginal: across every prompt, across all five AI systems, they were never named once. A little over a quarter (25.7%) were never cited as a source either.

The distance between the median (1.71%) and the mean (7.44%) is the whole story in two numbers. A small group of brands is winning heavily and dragging the average up, while the typical brand sits near zero. AI answer space is not evenly distributed, and it is not close.

For the record, this pattern holds against the competitors too. Where we could compare a brand against its named category rivals in the same scan, 70% of brands trailed the leader in their own category (21 of 30 comparable sets), by a median gap of 15.3 percentage points.

Finding 2: your market matters more than you would like

Same methodology, same five models, wildly different outcomes by market.

Read the small samples with appropriate caution: India, Malaysia and Japan have four brands each, and one unusual brand moves those medians a long way. Indonesia, Hong Kong and Singapore are the cuts we would defend.

The genuinely surprising result is Singapore. It is our largest market sample at 47 brands, it is the most digitally mature economy in the region, and its median brand appears in 2.00% of its own category prompts. Hong Kong, with half the sample, more than triples it.

Our read: Singapore's category prompts are disproportionately contested by global players. When a buyer asks about payments, logistics or SaaS "in Singapore," the models answer with Stripe, DHL and the global category leaders, because those brands dominate the English-language corpus these systems were trained on. A Singaporean challenger is not competing with the company down the road. It is competing with whoever owns the global head term.

Thailand and Japan show the opposite failure. Over half of Thai brands and three quarters of Japanese brands never appeared at all. But note the Thai citation figure: 1.20% median citation against 0.00% median appearance. Thai brands are being read as sources while not being recommended as options. Their content is feeding answers that name somebody else.

Finding 3: manufacturing is the most invisible sector at scale, and marketing agencies are the most invisible of all

Manufacturing and industrial is our largest single industry sample at 34 brands, and it is close to the bottom: a median appearance rate of 0.67%, with 41.2% of brands at absolute zero. This is the sector that defines mid-market Asia, the contract manufacturers, component suppliers, textile mills and machinery builders across the Pearl River Delta, Vietnam and Thailand. These are real companies with real revenue and decades of trading history, and AI assistants largely do not know they exist.

The reason is structural. Their buyers found them through trade shows, Alibaba listings, industry directories and personal referral. Almost none of that leaves the kind of public, text-based, independently-authored trail that a language model learns from. A factory with 800 staff and a 30-year record can have a smaller footprint in an AI model's training data than a two-person SaaS startup with an active blog.

And then there is the result we debated whether to publish. All four marketing and agency brands in the cohort scored exactly zero appearance. Four is a tiny sample and we are not going to over-read it. But the profession that sells visibility was, in this dataset, the single least visible category we measured. We include ourselves in that.

Finding 4: "AI visibility" is not one number, and the platforms disagree

Here is the same set of brands, measured across five systems.

Look at what happens when you sort by the other column.

Gemini ranks second on appearance and last on citation. Perplexity ranks last on appearance and second on citation. These are close to mirror images. Gemini will name your brand and not source you. Perplexity will source you and not name you.

This matters commercially, because almost every AI-visibility product on the market rolls these into a single blended "visibility score." That number is close to meaningless. A brand with a Gemini problem and a brand with a Perplexity problem can score identically and need completely opposite work: one needs authority and mentions elsewhere on the web, the other needs its own pages to be retrievable and quotable.

It also lines up with what Semrush found at far greater scale: ChatGPT surfaces roughly 15 sources per response against Gemini's 3. Different systems have structurally different appetites for citation, so the same content strategy produces different results depending on where your buyers actually ask.

One methodological note in the interests of not overselling our own numbers: Claude is measured through a direct API with web search enabled, while the other four are measured as a user would encounter them. Claude's higher figures partly reflect that cleaner measurement path, so we would treat the ordering of the citation column as the finding here, not Claude's absolute lead.

Finding 5: brands do move, but you become a source before you become a recommendation

Twenty-eight brands in the cohort had four or more clean scans, giving a median observation window of 103 days.

• 23 improved, 4 declined, 1 stayed flat.

• Median appearance rate change: +2.16 percentage points.

• Median citation rate change: +7.34 percentage points.

Two things stand out.

First, citation moved roughly 3.4 times more than appearance. Brands become sources that models read well before they become names that models recommend. If you are measuring only whether you get mentioned, you will conclude nothing is working for months after something has, in fact, started working. Citation rate is the leading indicator; appearance rate is the lagging one.

Second, and more encouraging: of the 8 brands that started at exactly zero appearance, 7 finished above zero. Getting off the floor is a reliably solvable problem. Seven of eight is not a coin flip.

The ceiling is another matter. The median brand gained 2.16 points. Nobody went from invisible to dominant in a quarter, and any vendor promising that is selling something we did not observe.

Finding 6: publishing volume does not predict visibility gains, at all

This is the finding we least expected and least wanted.

Of those 28 tracked brands, 22 had content published during the measurement window. The median brand among them received 142 blog articles. That is a serious, sustained content programme by any standard.

We then asked the obvious question: did the brands that got more content improve more?

No. Not even slightly.

A correlation of 0.155 is barely distinguishable from noise. A correlation of -0.020 is noise. Within a cohort ranging from 20 to over 300 published articles, the number of articles a brand published told us essentially nothing about how much its AI visibility improved.

Content is not irrelevant. Brands with a content programme did outperform those without. But once we correct for the fact that the with-content group was simply observed for longer (a median 109.5 days against 55), the advantage narrows sharply:

That is a real edge, but it is 1.8x, not the 4x the raw uncorrected numbers suggested. And with only six brands in the comparison group, we would not want anyone building a budget on it. We are publishing it precisely because the uncorrected version is the one a vendor would have shown you.

The conclusion we draw is narrow and firm: past some modest threshold, publishing more does not buy more. Something other than volume determines whether a model learns to recommend you.

Ahrefs reached a compatible conclusion from a completely different direction and a much larger sample. Across 75,000 brands, they found that a site's total page count correlated only weakly with AI visibility, while third-party brand mentions correlated far more strongly, and YouTube mentions strongest of all at roughly 0.737. Two independent datasets, two different methodologies, same shape of answer: models weigh what the web says about you far more heavily than what you say about yourself.

What we would actually do with this

Stop buying volume. If you are commissioning content by the article count, you are optimising the one variable we could not connect to outcomes. The brands in our data publishing 300 articles did not reliably beat the ones publishing 60.

Track citation rate, not just appearance rate. Citation moves first and moves roughly 3.4x more. A programme judged only on mentions will be cancelled a quarter before it starts working.

Find out which system your buyers actually use, then optimise for that one. Gemini and Perplexity are near-mirror-images in our data. A single blended visibility score will hide which problem you have.

If you are in manufacturing or industrial, the opportunity is unusually large. 41.2% of that sector scored zero. When four in ten of your competitors are completely absent, the bar to becoming the named default in your category is far lower than it will ever be again.

If you are in Singapore, check who you are actually losing to. In our data the competition is frequently a global brand that owns the head term, not a local rival. That is a different content problem and needs a different answer.

Method, and what this study cannot tell you

We would rather state the limits than have them found for us.

How the data was gathered. Each brand was scanned with a set of category prompts written to mirror how a real buyer asks (product-category questions, comparison questions, recommendation requests), then run across five AI systems. We recorded whether the brand was named in the answer and whether its domain was used as a source. Scans where more than 25% of prompts failed to return were discarded, not averaged in; 10 scans were dropped this way. Scrape health causes swings that look like visibility change and are not, and this is the main reason we report medians rather than means throughout.

This is observational, not experimental. No brand was randomly assigned to a treatment. Brands that received content programmes are Simaia clients, which means they self-selected, they were commercially motivated, and they differ from the comparison group in ways beyond content. The six-brand comparison group is far too small to carry weight. Treat Finding 6's negative result (volume does not correlate) as the robust part, and the positive claim (content beats no content) as suggestive only.

Sample sizes vary enormously by cut. The Singapore (47) and Hong Kong (24) market cuts and the manufacturing (34) industry cut are reasonably solid. India, Malaysia, Japan and Marketing & Agency have four brands each. We have shown every n so you can discount accordingly. Buckets with fewer than four brands were not reported at all.

Share of Voice is deliberately absent. We track it, but it is computed against each brand's named competitor set, and that set grows as an engagement matures. A brand's Share of Voice therefore falls mechanically as we add competitors to track, which makes it useless for a longitudinal study. Publishing it would have produced a tidy, entirely artificial finding.

The cohort is not a random sample of Asian business. These are companies that came to a GEO agency, which skews toward firms that already suspected they had a visibility problem. If anything, that biases the baseline downward. A random sample of Asian mid-market B2B would likely look somewhat better than 1.71%, though we would be surprised if it looked dramatically better.

Everything here is anonymised and stays that way. Brands are identified internally by one-way hash. No client is named, and no figure that could identify a single brand is published.

Download this report here

Next edition

We intend to run this quarterly. Two things we want to add: a test of whether third-party mentions (the signal Ahrefs found strongest) move our cohort the way their data predicts, and a proper matched comparison so the content question can be answered with something better than six control brands.

If you want your own category measured against this benchmark, that is what we do.

References

1. Ahrefs, Q1 2026 AI Search Benchmark Report, 75,000-brand correlation study. ahrefs.com

2. Semrush, 2026 AI Visibility Index, 126 million US AI search prompts, January to April 2026. semrush.com

3. Simaia AI Visibility Archive, 148 brands / 587 scans / 34,794 prompt-runs, December 2025 to July 2026. First-party data.