Why monitoring only ChatGPT is not enough in Malaysia

Short answer: ChatGPT is one AI referral source out of four major ones. Claude, Gemini and Perplexity together account for a substantial share of AI-driven traffic, and two of them rarely label themselves in your analytics. A ChatGPT-only check leaves most of that invisible.

The HiGoodie AI Search Traffic Report found that ChatGPT's share of measurable business-to-business AI referrals fell from 89% to 63% over eight months, with Claude at 18.5%, Gemini at 10.6% and Perplexity at 7.3%. On those figures, roughly 37% of measurable AI referrals arrive from engines a ChatGPT-only check never looks at. We can't link that report here, so treat it as one vendor's dataset, drawn from business-to-business traffic rather than from Malaysian traffic specifically.

This is written for owners and practice managers of Klang Valley professional services firms with 2 to 50 staff: clinics, dental and veterinary practices, law and accounting firms, tuition centres, property agencies. Disclosure: it comes from Geo One, a Generative Engine Optimisation (GEO) and Answer Engine Optimisation (AEO) agency based in Kuala Lumpur.

Why doesn't your analytics show the other engines?

Because two of them largely don't tag themselves. Industry monitoring suggests Claude and Gemini commonly strip referrer information, so a visitor who arrived after reading an AI answer lands in your analytics as direct traffic, or as nothing identifiable at all. Perplexity behaves differently: it shows its sources openly and links them, which is why it tends to produce more visible referrals than its share of overall usage would suggest.

There is one weak proxy worth watching while you wait for better attribution. Direct traffic normally lands on your home page, because the home page is what people type or bookmark. If direct sessions start arriving instead on a deep page, one service page, a fee page, a single practitioner's bio, something pointed them at that exact URL, and an unlabelled AI answer is one of the few candidates. Treat it as a hint that tells you where to go and test manually. It is not a measurement, and you should not report it as one.

Your traffic report is not a measurement of AI visibility. It's a measurement of the engines that happen to label themselves. If you want to know whether Claude names your firm when someone asks for a conveyancing lawyer in Bangsar, the only reliable method is to ask the question and record the answer.

Why does Gemini deserve its own line in a Malaysian report?

Because of where it already sits in this market. Malaysia's National AI Office, set up under the Ministry of Digital, has been scaling AI use inside government, including a rollout of Google Workspace's Gemini Suite to 445,000 public officers, as reported in Chambers and Partners' 2025 Malaysia guide. If you sell professional services to people who work in those departments, and most Klang Valley accounting firms, clinics and law practices do, Gemini is already open on their desk, in the same tab set as their email, with no separate app to install. That is a different adoption path from ChatGPT's, and it does not show up in global usage averages.

Two things follow for a Malaysian firm. Test in Bahasa Malaysia as well as English, because the prompt language changes the answer. Keep your mobile pages fast and readable without JavaScript, because in a phone-first market most of the asking happens on a handset, not a desktop. We go into the engine-by-engine detail in our note on how Gemini and Perplexity do two different jobs.

How do Claude and Perplexity behave differently?

Claude is a research tool used at a desk, often inside longer conversations, and it frequently answers without displaying a source list. That makes it the hardest of the four to monitor from the outside, and the easiest to ignore by accident. The check has to be manual and repeated: fresh chat, buyer-style question, record who gets named. Our walkthrough on what Claude says about your business sets out the five-question method.

Claude also reaches your site with its own crawler, under its own user agent, which is separate from OpenAI's. A robots.txt rule copied from a template can block one, both or neither. Before you edit that file, read the crawler and user agent pages in Anthropic's documentation and in the OpenAI platform documentation, and read them on the day you make the change, because the published user agent names are updated from time to time.

Perplexity is the citation-first engine. It shows its working, which makes it the best early warning system of the four: if a competitor's page is cited in a Perplexity answer about your category, you can see exactly which page and go and read it. Scale matters here too. Index.dev reports that Perplexity handled around 780 million queries in May 2025, up from roughly 230 million in mid-2024, and estimates 1.2 to 1.5 billion queries a month by mid-2026. That last figure is a projection, not a published platform number.

What does a multi-engine visibility check actually find?

In our most recent self-scan, run in a single day across ChatGPT, Claude, DeepSeek, Gemini and Perplexity, we checked 110 answers. Geo One was named in 9 of them.

The interesting part is where. We were named in the answer to "recommended AEO agency for small business in Kuala Lumpur". We were not named in the answer to "best generative engine optimisation agency in Kuala Lumpur". Same business, same day, same engines, two closely related questions, two different outcomes. A single question on a single engine would have told us either that things were fine or that things were broken, and both readings would have been wrong.

That is the practical case for running a list rather than a spot check. Presence in AI answers is not a property of your business; it is a property of each individual question. Two synonyms for the same service can sit on opposite sides of the line.

We publish our own weak number on purpose. A provider who shows you their own scan, including the prompts they lose, is a provider you can check. One who only shows you a score cannot be checked at all.

What should a Malaysian SME actually monitor?

Four things, on a fixed list of buyer questions rather than your company name.

1. The same prompts on every engine. Ten to twenty questions a real buyer would type, run across ChatGPT, Claude, Gemini and Perplexity. Same wording each time, or the comparison between months means nothing.
2. Named or not named, and in what position. Record it as a count, not an impression. Nine out of 110 is a number you can move. "We're doing better on AI" is not.
3. Accuracy, separately from presence. Being named with the wrong clinic hours or an old branch address is a different problem from not being named, and it has a different fix.
4. Language. English and Bahasa Malaysia at minimum, Mandarin if your buyers use it. Our note on tracking Bahasa Malaysia prompts explains why the answers diverge.

Monthly is a reasonable rhythm for a stable business. Weekly while you're actively changing schema, content or your Google Business Profile, because changes take time to propagate into AI answers and you need to know when they land.