AI gets facts about a company wrong because it knows the company from public data, which is often duplicated, contradictory or empty. Claude, Microsoft Copilot, ChatGPT and Gemini know a company in two ways: what the model learned from web data up to its cutoff date, and the pages it finds when it answers. Both ways read the same set of identity data: the company's website and the markup on it, Wikipedia, Wikidata and Google's Knowledge Panel. A study at ACL 2023 found that models get little-known subjects wrong far more often. GPT-3 answered only 19% of questions correctly in the least-known group. We surveyed the companies in the SET50 on 13 Nov 2026. All 50 have a Wikidata item that records their ticker symbol. But for 5 companies the ticker sits on more than one item, including items for a state enterprise and a subsidiary. Another 9 only got an item in April 2026, and it has no website, logo or Wikipedia link. Only 22 companies have complete basic data. Of the 35 home pages we could read, 14 have Organization markup, and only 2 link to Wikipedia or Wikidata. This article has the survey results, the order to fix things in, a table of Wikidata properties and a markup template you can use right away.
TRAINIAC ran the survey on 13 Nov 2026. The method is in the survey section. We do not name the companies. Information from Google, Microsoft, OpenAI, Anthropic and Wikimedia was checked on their official pages on the same day.
Where AI learns about a company
| AI | Knowledge built into the model | Looks up new information from |
|---|---|---|
| Claude | Anthropic says its models learn from publicly available web data along with other datasets. The current models have a cutoff of June 2026 | Brave Search and the Claude-SearchBot index |
| Microsoft Copilot | Microsoft does not state a cutoff date in the documents we checked | Bing. Microsoft says Copilot uses the same crawling and ranking base as Bing |
| ChatGPT | OpenAI says the main data comes from publicly available information on the internet. The main model, GPT-6 Astra, has a cutoff of 30 Apr 2026 | The OAI-SearchBot index and partner search providers |
| Gemini | Not checked in this round | The Google Search index |
This has two consequences.
- Wrong data from before the cutoff date stays in the model until a new version comes out. What can be fixed quickly is the information AI finds when it answers. Anthropic says web search lets Claude answer with information from after its cutoff date, and OpenAI says ChatGPT searches the web on its own when a question needs current information.
- Little-known companies lose out the most. A study by Mallen et al. (ACL 2023) built a set of 14,000 questions from Wikidata and measured popularity by Wikipedia page views. In the 4,000 least-known questions, GPT-3 answered 19% correctly, and making the model bigger barely helped. But systems that retrieve information to support their answers did much better. Most mid-sized Thai companies are in this group. So public data that can be found, and that matches everywhere, matters more to these companies than to famous brands.
Google says the information in Knowledge Panels comes from many sources on the web and is generated automatically. A 2020 Google blog post says Wikipedia is a frequently cited source but not the only one, and that public knowledge bases such as Wikidata help Google's systems a great deal. Bing says it uses markup to supplement and verify its own sources, and that naming organisations clearly and consistently helps them be used as sources and cited more accurately. No provider says that editing Wikidata or markup will change AI answers right away. What you can do is make the data these systems read correct and consistent.
Survey: the identity of SET50 companies on Wikidata and their home pages
On 13 Nov 2026 we used the SET50 list and the website each company gives on its company profile page at the Stock Exchange of Thailand (SET). This is the same set as the robots.txt survey in our GEO guide (in Thai). We then checked two things.
- Wikidata: through the Wikidata Query Service, we pulled every item that lists the Stock Exchange of Thailand with a ticker symbol, and matched them against the 50 tickers. We took each item's creation date from its edit history.
- Home pages: we requested the home pages of the 48 companies that have a website, with a normal browser user agent and without running JavaScript. A page counted as readable if it returned 200, had more than 300 characters of text and was not an intro page. We then read any JSON-LD of an organisation type.
| What we checked | Result | Notes |
|---|---|---|
| Has a Wikidata item that lists the ticker symbol | 50 of 50 | The ticker is stored as a qualifier of the stock exchange property |
| The same ticker is on more than one item | 5 of 50 | 2 are duplicate items for the same company · for 3, the ticker is also on another organisation's item |
| Item only created in April 2026 | 9 of 50 | None of the 9 has a website, logo or Wikipedia link |
| Lists an official website (P856) | 30 of 50 | 27 match the website given to SET · 3 differ |
| Has a logo (P154) | 27 of 50 | |
| Has a Thai name | 49 of 50 | |
| Has a Thai · English Wikipedia article | 40 · 28 of 50 | 10 have neither |
| All four: website, logo, Thai name and Thai Wikipedia | 22 of 50 | |
| Home page readable (of 48 sites) | 35 of 48 | 8 sites refused the request or did not respond · 5 were an intro page or a blank page |
| Home page has organisation-type markup | 14 of 35 | 11 include a logo |
| Markup has sameAs | 8 of 35 | 1 gives sameAs as text, not a URL |
| sameAs links to Wikipedia or Wikidata | 2 of 35 |
Five kinds of error we found in large companies
Each one makes systems that read Wikidata or web pages see one company as several, or two companies as one.
- Duplicate items. One company has two items with separate data. One has the website and Wikipedia links, the other is almost empty. A system that picks up the empty one gets nothing.
- The ticker symbol is on the wrong organisation. We found three kinds: a state enterprise with a name similar to the listed company, a subsidiary in the same group, and the original company before it was restructured into a holding company. All three mix up the data of two organisations.
- Empty items. Only a name and a ticker, with no website, logo or any link that confirms which company it is.
- The website does not match the one given to SET. We found old domains that redirect to a new site, websites of sub-brands, and domains that no longer respond.
- Home page markup points nowhere. 6 of the 14 sites with markup have no sameAs at all, and one more gives sameAs as text instead of a URL. schema.org defines sameAs as the URL of a page that confirms the identity, such as Wikipedia, Wikidata or an official website.
Fix things in this order
1. Check your company's Wikidata
Search wikidata.org for the company's name in both Thai and English. See how many items there are, and whether each one is really about your company. Wikidata allows you to edit the item about yourself or an organisation you represent, but every statement must have a verifiable reference. You can use the company's own website as a reference only for facts about the company itself. Do not remove statements that others have sourced. If you disagree, discuss it on the talk page.
- Duplicate items: Wikidata can merge two items when they really are about the same thing, using the merge gadget. It moves the data and turns the old item into a redirect. The help page stresses that you must be sure they are the same thing before you merge.
- Ticker on another organisation's item: move the ticker to the correct item, with a reference to the company profile page on SET.
- No item: anyone can create an item on Wikidata, but the item must meet the notability criteria: it has a Wikipedia article, or it can be described using reliable public references. Wikidata advises against creating an item about yourself if you are not sure it meets the criteria.
- If you pay someone to edit: since 2022 Wikidata has had a policy that people who are paid must disclose the organisation that pays them.
| Property | Thai label in Wikidata | What to enter |
|---|---|---|
| P31 | เป็น | The type, for example public company |
| P856 | เว็บไซต์ทางการ | The same website you gave SET and use in your markup |
| P414 + P249 | ตลาดหลักทรัพย์ + ชื่อย่อหลักทรัพย์ | The Stock Exchange of Thailand, with the ticker as a qualifier |
| P1448 | ชื่อทางการ | The legal entity name as registered |
| P13857 | (no Thai label yet) | The company registration number. This property points to the data warehouse of the Department of Business Development (กรมพัฒนาธุรกิจการค้า) as its source |
| P1278 | (no Thai label yet) | The LEI code, if you have one |
| P154 | โลโก้ | The logo file on Wikimedia Commons |
| P571 · P159 | วันที่ก่อตั้ง · ที่ตั้งสำนักงานใหญ่ | The founding date, and the province or city |
| P749 · P355 | บริษัทแม่ · หน่วยงานลูก | Use these to tell companies in the same group apart |
| Labels and descriptions | In Thai and English | A description that sets the company apart from companies with similar names. Wikidata does not allow two items to have the same label and description |
2. Add Organization markup that matches the page
Google says Organization markup helps Google tell your organisation apart from others. No property is required. Google recommends putting it on the home page or one about-us page, and using the most specific type, such as Corporation for a public company. Google's general rules are that markup must reflect the real content of the page, that you must not mark up content readers cannot see, and that you must not impersonate another organisation. Bing says the same: markup must match the visible content.
{
"@context": "https://schema.org",
"@type": "Corporation",
"name": "Example Group",
"alternateName": "ตัวอย่างกรุ๊ป",
"legalName": "Example Group Public Company Limited",
"url": "https://www.example.co.th/",
"logo": "https://www.example.co.th/logo-512.png",
"foundingDate": "1995-03-01",
"iso6523Code": "0199:YOUR-COMPANY-LEI",
"address": {
"@type": "PostalAddress",
"streetAddress": "Number, Road",
"addressLocality": "District",
"addressRegion": "Bangkok",
"postalCode": "10xxx",
"addressCountry": "TH"
},
"sameAs": [
"https://www.wikidata.org/wiki/Q000000",
"https://th.wikipedia.org/wiki/Article_title",
"https://www.linkedin.com/company/company-name",
"https://www.facebook.com/page-name"
]
}
- name and alternateName: Google asks you to use the same name as the site name you have set. If the company uses both a Thai and an English name, put the other one in alternateName.
- logo: at least 112x112 pixels. Google uses it to choose the logo in search results and the Knowledge Panel.
- iso6523Code: Google recommends giving the LEI with the prefix 0199: instead of the leiCode property.
- sameAs: enter the URL of the correct Wikidata item, after you have fixed it in step 1, not a duplicate item.
3. Get verified on your Knowledge Panel
Google lets an official representative click "Claim this knowledge panel" and verify through an account linked to the official website, such as Search Console or YouTube. But Google says not every panel can be verified yet. Once verified, you can suggest changes with links to evidence. But the description in the panel comes from other sources and cannot be edited directly. You have to fix it at the source, which takes you back to steps 1 and 4. For shops or branches that customers travel to, use Google Business Profile instead. See our article on AI for shops and branches.
4. Wikipedia: do not write it yourself
Thai Wikipedia considers a company notable enough for an article when it has significant coverage in several reliable secondary sources that are independent of each other. Press releases, paid articles and the company's own website do not count. Nor does the number of staff, revenue or the company's age make it meet the criteria.
Thai Wikipedia's conflict of interest guideline bans editing for your own benefit. If you are involved, disclose it, and propose changes on the talk page instead of editing the article directly. The Wikimedia Foundation's Terms of Use require anyone who is paid to edit to disclose their employer, client and every beneficiary. The Thai guideline also warns that the media have reported on companies that secretly edited Wikipedia and damaged their reputation. And once an article is up, nobody owns it: other people can add things the company does not like.
What a company can do is earn real coverage from independent sources, such as news the media write themselves or interviews, and let volunteer editors write the article. If an article already exists but has wrong facts, propose a fix on the talk page with references.
5. Use the same name and data everywhere
The Thai name, English name, legal entity name, website and address must match on the company's website, the company profile page on SET, Wikidata, LinkedIn, Facebook and other directories. Bing says that naming organisations, people, products and places clearly and consistently helps them be cited more accurately, and that you should remove outdated information.
How TRAINIAC helps with this
TRAINIAC provides GEO services for businesses, including checking and fixing a company's identity following the steps in this article. We follow Wikidata and Wikipedia policies, which includes disclosure when we edit on a client's behalf. If you want your corporate communications team to do it themselves, there is a one-day GEO course. Send us the details through our contact page. Before you hire anyone, also read how to hire for GEO and write the TOR.
Sources
- Organization structured data, Google Search Central (last updated 8 Sep 2026)
- General structured data guidelines, Google Search Central (last updated 10 Jul 2026)
- About knowledge panels · Get verified on Google · Submit feedback on content about you, Google Search Help
- A reintroduction to our Knowledge Graph and knowledge panels, Google, 20 May 2020
- Organizing the world's information: where does it all come from?, Google, 3 Dec 2020
- Marking up your site with structured data and Bing Webmaster Guidelines, Microsoft
- Wikidata:Notability · Wikidata:Autobiography · Wikidata:Disclosure of paid editing · Help:Merge · Help:Items, Wikidata
- Wikipedia:Notability (organizations and companies) (วิกิพีเดีย:ความโดดเด่น (องค์การและบริษัท)) and Wikipedia:Conflict of interest (วิกิพีเดีย:ผลประโยชน์ทับซ้อน), Thai Wikipedia
- Terms of Use, Wikimedia Foundation (section: Paid Contributions Without Disclosure)
- sameAs, schema.org
- Transparency Hub and Models overview, Anthropic
- Web search tool, Claude Docs
- How ChatGPT and our foundation models are developed and Searching the web with ChatGPT, OpenAI Help Center
- Models, OpenAI (the cutoff date of each model)
- When Not to Trust Language Models, Mallen et al., ACL 2023
- SET50, the Stock Exchange of Thailand, and Wikidata Query Service (survey data)
Read next
For an overview of where each AI searches the web, see the GEO guide for Thai businesses (in Thai). For what not to do to get AI to mention your company, such as faking reviews or hiding text that gives AI instructions, see GEO that breaks the rules and the law.