A project engineer at an operator in KL asks ChatGPT for a quick background on your company before a pre-qualification call. The answer names a supply base you closed, lists a vessel you sold, and calls your joint venture by the partner's old name. Nobody in your team saw it, and nobody can correct it in the room. That is the real cost of poor Malaysia energy logistics AI accuracy: wrong facts reach buyers before your sales team does.

This article explains why AI assistants get bunkering, terminal, marine and supply base operators wrong, what buyers in Malaysia actually ask, how to audit your own answers this week, and the specific fixes that give ChatGPT, Gemini, Perplexity and Google's AI features a current, quotable source. The steps work whether or not you hire anyone.

Why AI gets energy logistics facts wrong

AI assistants describe your company from the text they can find, and in energy logistics most of that text is old. Directories, tender notices, conference brochures and news reports outlive the assets they describe, and AI tools have no way to know which version is current unless you publish one.

There are two routes into an answer. The first is what a model learned in training, which is frozen at a point in time and blurry on companies that are not widely covered. The second is live retrieval, where the assistant searches the web and reads current pages.

OpenAI's crawler documentation describes the second route for ChatGPT. It says OAI-SearchBot "is for search" and that "sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers". GPTBot, which collects training data, is controlled separately, and OpenAI notes that "each setting is independent of the others". Many corporate sites in this sector sit behind strict firewalls or CDN bot rules, so the search crawler may never reach the page that would have corrected the answer.

Why the sector is especially prone to errors

Energy logistics firms change shape often. Vessels are chartered in and out, supply base contracts move between Kemaman, Labuan, Miri and Bintulu, terminals change hands and joint ventures are renamed when a partner exits. Each change leaves a trail of older pages that still describe the previous arrangement.

Ports and fuels are also shifting. MIDA's June 2025 article on bunkering notes Malaysia's first LNG bunkering operation at the Port of Tanjung Pelepas in February 2024 and biofuel bunkering at Port Klang in July 2024. When services move this fast, an AI answer built from 2022 sources will be out of date even if every source was accurate when written.

Why rivals sound more accurate

Google's guidance on its AI features says there are no special optimisations needed beyond normal indexing, and recommends "making sure that important content is available in textual form". It also says AI Overviews and AI Mode may use "query fan-out", running several related searches to build one answer. A competitor whose site states its ports, vessel classes and licences in plain text gives every one of those searches something to match. A rival with a glossy PDF capability statement gives them almost nothing.

What buyers in Malaysia actually ask AI

Buyers ask AI about energy logistics vendors by asset, location, licence and track record, mostly in English, with Bahasa Malaysia common among local operators and contractors. These are realistic prompts to test, grouped by stage.

Scoping the market

  • "Which companies provide offshore supply vessels out of Kemaman Supply Base?"
  • "Who are the main bunkering suppliers at Port Klang and Tanjung Pelepas?"
  • "Syarikat logistik minyak dan gas yang berlesen PETRONAS di Sarawak"

Checking a specific vendor

  • "What vessels does [your company] operate and where are they based?"
  • "Is [your company] a joint venture? Who are the partners?"
  • "Does [your company] supply LNG or biofuel bunkers in Malaysia?"

Comparing before a tender

  • "Compare [your company] and [competitor] for tank storage in Johor"
  • "Which Malaysian marine logistics firms have handled decommissioning projects?"

Run each with and without your company name. The unbranded prompts show whether you are named at all; the branded ones show whether what AI says about you is true.

How to check where you stand

You can audit AI accuracy in two or three working days with a spreadsheet and a few accounts. The aim is a baseline you can repeat, not a one-off screenshot. We describe the method we use in more detail on how we measure.

Step 1: Write down the truth first

Before you test anything, list the facts an AI answer should contain: legal entity names, joint venture names and partners, operating ports and bases, vessel types or terminal capacity categories, fuels handled, key certifications and your PETRONAS licence or registration status. Have operations and legal sign it off. Without this list, nobody can score an answer.

Step 2: Build a fixed prompt set

Pick 20 to 30 prompts from the patterns above, covering your main service lines and locations. Include a few in Bahasa Malaysia if your buyers use it. Save the exact wording, because small changes in phrasing produce different answers.

Step 3: Run each prompt several times

Test in ChatGPT with search on, Gemini, Perplexity, Microsoft Copilot and a Google search that triggers an AI Overview. Run each prompt at least three times on different days. Answers vary between runs and users, so one result is a sample, not a verdict.

Step 4: Score accuracy field by field

For every answer that mentions you, mark each fact as correct, wrong, outdated or missing. Use five fields: entity name, locations, assets, services and partners. A simple score out of five per answer is enough to see trends.

Step 5: Log the cited sources

Where the tool shows sources, record every URL. You will usually find the same few culprits: an old directory profile, a conference speaker bio, a dated news item, a partner's page that still lists the previous JV name. Those URLs become your correction list.

Step 6: Check crawl access

Ask IT to confirm that robots.txt and any firewall or CDN rules allow OAI-SearchBot and Googlebot to reach your main pages. If your services are only described in a gated portal or a downloadable PDF, note that too.

How to fix it

The fixes are mostly about publishing one current version of the truth and making the rest of the web agree with it. Start with your own site, then work outwards.

Publish a dated facts page

Create one HTML page, linked from the main navigation, that states your current facts in short sentences: who you are, which legal entities trade under the brand, where you operate, what you operate, what you handle and which licences and certifications you hold. Put a visible "last updated" date at the top and change it whenever a fact changes. This page is what you want ChatGPT and Perplexity to cite when someone asks "what does [company] do?".

Write one page per service line and location

Bunkering at Port Klang, tank storage in Johor, OSV charter from Kemaman and marine services in Labuan are different questions, so they deserve different pages. Each should name the port or base, the service, the vessel or asset types and the customers you serve, without claims you cannot back. These pages match the specific prompts buyers type far better than a single "Services" page.

Make your licence status clear and accurate

PETRONAS distinguishes between a licence, which allows a company to supply goods and services to both upstream and downstream, and a registration, which covers downstream only. Its licensing page also explains that SWEC is "the list of work categories in upstream and downstream". State which you hold and for which work categories, in the same terms PETRONAS uses, and keep it current at renewal. Do not imply a licence or category you do not hold; AI tools will repeat the claim and buyers will check.

Fix joint venture and entity naming

Write one plain sentence for each joint venture: its legal name, both partners and what it does. Use the same sentence on your site, in press releases, on LinkedIn and in conference bios, and ask partners to match it. When a JV is restructured or renamed, publish a short dated note saying so, with both the old and new names, so AI tools can connect them.

Add Organization structured data

Google's Organization structured data documentation explains how to describe a company's legal name, logo, address, contact points and profiles. Add it to your home page and facts page, list your subsidiaries where appropriate and make sure every value matches the visible text. Google's AI features guidance specifically recommends structured data that "matches the visible text on the page".

Correct the third-party sources AI already reads

Use the URLs from your audit to drive this work.

Source type Why AI tools use it What to do
Industry and business directories Structured, easy to parse, often old Claim, update or ask for removal of stale profiles
Port and terminal partner pages Seen as independent confirmation Ask partners to update your name, services and contact
Trade press and news Authoritative and frequently cited Issue dated releases for asset, port or JV changes when disclosable
Conference and association pages Long-lived speaker and exhibitor bios Send current boilerplate after every event
LinkedIn and Google Business Profile Frequently crawled, often the first profile found Align descriptions word for word with the facts page

Keep asset changes flowing

When you take delivery of a vessel, start a new bunkering service or exit a base, publish it. A short, factual release and an update to the facts page on the same day gives search-based AI a fresh source. Silence leaves the old version standing.

What to measure

Measure accuracy first, then visibility. An AI answer that names you with the wrong ports is worse than no mention at all, because it travels with your brand attached. We set this up as part of AI Visibility Intelligence.

Track these monthly on the same prompt set:

  • Accuracy score: correct fields out of five per answer, averaged per engine.
  • Mention rate: share of unbranded prompts where your company is named.
  • Citation share: share of cited sources that are your own pages, and which third-party pages appear most.
  • Stale source count: how many outdated URLs still appear in citations.
  • Description quality: whether the answer describes your current services, not just your history.

Expect noise. Results vary between runs, engines and users, so look at trends across three or more months rather than reacting to a single bad answer. Re-run the audit after any major asset or JV change.

Common mistakes

  • Treating the capability statement PDF as the source of truth. AI tools struggle to quote it, and it is often years old.
  • Correcting your own site but leaving old directory profiles and partner pages untouched.
  • Using three versions of a joint venture name across the site, press and LinkedIn.
  • Blocking all AI crawlers by default, then wondering why ChatGPT's search answers cite a directory instead of you.
  • Claiming licences, work categories or capacities you cannot document, which AI tools then repeat to buyers who will check.
  • Measuring once, finding a good answer and assuming the problem is solved.

How Shubi helps

Shubi is an AI visibility company based in Malaysia. For energy logistics firms we build the prompt set, score accuracy across the main AI engines, trace wrong answers back to their sources and help your team publish facts AI can quote, through AI Visibility Intelligence and content and entity work. We do not promise mentions or rankings, only a clear method and honest reporting. Message Shubi on WhatsApp