AI engines describe your business from two stores: training memory, frozen when the model was trained, and live retrieval, which reads the web at question time. Google adds a third layer, the Knowledge Graph, which stores your business as an ID, not a name. All three break the same way: when your legal name, trading name, SSM number, address and category disagree across sources, the engine either merges you with someone else or hedges. We measure this with the Entity Agreement Rate: the share of 8 core identity facts that match across the 10 sources an engine is most likely to read about you. Fix retrieval-layer sources first (your site, Google Business Profile, directories), because training memory only updates at the next model release.

Introduction

Ask Google’s Knowledge Graph Search API about Stanford University and the reply does not lead with a name. It leads with identifiers: a Knowledge Graph machine ID (/m/06pwq), a Google Place ID, and a Wikidata ID (Q41506), all in one JSON-LD record (Google Cloud documentation). That is how a machine holds a brand. The name is a label attached to the ID, and a brand can carry several labels.

Malaysian businesses make this harder than most. A company registered as “KK Food Ventures Sdn. Bhd.” trades as “Kopi Kita”, lists “Kopi Kita Bangsar” on Google Maps, and shows a 2018-era registration number in the footer that no longer matches the 12-digit format SSM issues today. To a person those are obviously one business. To an entity system they are four candidate labels, and every one that cannot be tied back to the same ID weakens what an AI engine will say about you.

What is entity SEO?

Entity SEO is the work of making every source on the web describe your business with the same core facts, so that search engines and AI models resolve those sources to one entity. It targets identity (who you are, what category you sit in, where you operate), not keywords, and it covers your site, your schema, Google Business Profile, directories, SSM-linked records and third-party mentions.

The mechanism goes back to Metaweb, the company behind the Freebase knowledge base, which Google acquired in July 2010. Metaweb’s patent US 7,502,770 (“Knowledge Web”, granted 10 March 2009) describes assigning each named entity an ID so that facts collect around the ID rather than around a string of text. Google’s later patent US 8,738,643, “Learning synonymous object names from anchor texts” (filed 2 August 2007, granted 27 May 2014), describes how a system could collect alternative names for an entity from the anchor text of links pointing at the page its facts came from. Bill Slawski’s example from the patent: “International Business Machines Corporation”, “IBM” and “Big Blue” resolving to one ID (SEO by the Sea, June 2014).

A patent shows that Google explored a method, not that it runs today in that exact form. What it does show is the logic: aliases are acceptable when the evidence links them to one ID. Your Sdn Bhd name and your trading name can both be correct. The problem starts when nothing on the web connects them.

If your trading name and your registered name never appear together on a single page you control, you have given the entity system no evidence that they are the same business.

Where do AI engines get what they “know” about your brand?

From two places. Training memory is what the model absorbed from its training corpus and does not change until a new model ships. Live retrieval is what the engine searches and reads when the question is asked: Google AI Overviews and AI Mode retrieve from Google’s own index, and ChatGPT search, Perplexity and Claude run their own web searches. A brand fact that is wrong in training memory can outlive the fix on your website by a full model generation.

Layer What it holds When it updates What changes it Where you see it
Training memory Facts and associations absorbed from the training corpus Only at the next model release Consistent facts across many crawlable sources before the next training run Answers given with search switched off, or for questions the engine does not bother to search
Live retrieval Passages fetched from the web for this question Each time the source page is recrawled and re-ranked Your site, Google Business Profile, directories, press and forum pages that rank for your brand query Cited answers in AI Overviews, AI Mode, ChatGPT search, Perplexity, Claude
Google Knowledge Graph Entity records keyed by ID, with names, types, and links to Wikidata and Place IDs Continuously, as Google reconciles new evidence Corroborated facts, Business Profile data, Wikidata where it exists Knowledge panels, brand SERP features, entity-aware query rewriting

Caption: the three stores an AI answer about your brand can draw from, and how fast each one moves. Source: Ai Mode synthesis of Google Search Central’s “Optimizing your website for generative AI features on Google Search” (published 15 May 2026), Google’s Knowledge Graph Search API documentation, and Google’s knowledge panel help page, all retrieved 5 November 2026. Update speeds are qualitative; no engine publishes a recrawl or retraining schedule for individual brands.

This split is why we drew the Retrieval Line in our GEO vs AEO vs SEO explainer: above it you influence the answer this week, below it you influence the next model. For Google specifically, the retrieval layer is the organic index, and Google’s own May 2026 guide says AI Overviews and AI Mode are rooted in its core ranking and quality systems. How Google AI Mode chooses which sites to cite covers the Grounding Pool that retrieval draws from.

The practical consequence for entity work: retrieval is where a contradiction does the most visible damage, because the engine reads your conflicting sources side by side in one session and has to pick. In our State of AI Search in Malaysia 2026 pilot, the one AI Overview we captured (for “personal loan malaysia”, 21 July 2026) cited CIMB, Maybank and RinggitPlus: three entities with years of consistent, corroborated records. One AIO is a demonstration, not a rate, but it fits the pattern.

4. What is the Google Knowledge Graph, and how does a business get into it?

The Knowledge Graph is Google’s store of entities and the facts linking them. Its public API returns each entity with a machine ID (/m/ or /g/ format), schema.org types, and where available a Wikidata QID. Google says knowledge panels are generated automatically once enough information about an entity exists on the open web; you cannot apply for one. Local businesses manage their panel through Google Business Profile, not the knowledge panel claim process.

Freebase, the Metaweb database Google bought in 2010, was one of the main early sources behind the Knowledge Graph. Google shut Freebase down on 2 May 2016, and Wikidata became the open knowledge base most entity work now points at. That history matters because it tells you what kind of evidence the graph was built to ingest: structured, referenced, ID-linked records, not marketing copy.

For a Malaysian SME, there are three routes in, in the order most businesses actually get there:

  1. Google Business Profile. For anything with a storefront or service area, this is the panel you will see first. It is also the record most likely to carry your trading name rather than your Sdn Bhd name.
  2. Corroboration across the open web. Directories, news coverage, association listings and your own site repeating the same facts. Google’s help page says panels appear once enough information is available; corroboration is how “enough” is reached.
  3. Wikidata, where you qualify. Covered in the next section, and in detail in our upcoming Google Knowledge Panel and Wikidata guide for Malaysian businesses .

If you already have a panel, an official representative can claim it through Google’s “Get verified on Google” process and suggest edits. Claiming does not let you rewrite facts that the web contradicts.

Key takeaway: search your trading name and your Sdn Bhd name separately. If they produce two different panels (or a panel for one and nothing for the other), the graph has not joined them yet, and that is the first thing to fix.

5. Does your business need a Wikidata item?

Only if it qualifies, and most Malaysian SMEs do not yet. Wikidata accepts an item if it meets at least one of three criteria: it has a valid link to a Wikimedia page such as Wikipedia, it is a clearly identifiable entity that can be described using serious and publicly available references, or it fulfils a structural need for other items. Your own website and LinkedIn page do not count as independent references.

Wikidata items carry a permanent QID and typed statements (instance of, country, inception, official website, headquarters), and Google’s Knowledge Graph API returns wikidataQID as one of an entity’s identifiers. That is the real value: a stable external key. Treat it as identity plumbing: it helps systems resolve you, and Google’s May 2026 guide gives no indication that it lifts AI citations on its own.

The traps are well documented. Wikidata’s editing help says not to create items for promotional purposes, community norms treat writing about your own company as a conflict of interest, and items without independent references get nominated for deletion. A government registry entry can support basic facts like the legal name and inception date, but an item whose only references are the company’s own pages is the item that disappears in six months.

Who should do it now: listed companies (Bursa Malaysia filings are serious public references), businesses with coverage in The Star, The Edge, Malay Mail or Marketing Magazine, and founders with conference or press coverage. Who should wait: a three-year-old services firm with no independent coverage. For them, digital PR for AI citations comes first, and the Wikidata item follows the coverage.

6. Why do Sdn Bhd and trading-name mismatches break your entity?

Because Malaysian businesses routinely publish two or three official-looking names: the SSM-registered company name ending in Sdn. Bhd., one or more trading names, and a Google Maps label with a location suffix. Each is legitimate. Section 30(2)(b) of the Companies Act 2016 already requires your registered name and registration number on your website, so the fix is to publish the link between names, not to pick one.

SSM issued notices on 1 November 2023 reminding companies of the Section 30(2)(b) requirement to display their company name and registration number on their websites (MAICSA announcement). Most Malaysian sites comply in the footer and nowhere else. That footer line is the single most useful piece of entity evidence you own, because it is the one place your trading name and legal name sit on the same page. Most sites then waste it.

Mismatch pattern What the entity system sees Fix
Footer shows “KK Food Ventures Sdn. Bhd.” with no trading name nearby A company with no stated relationship to the brand on the rest of the page One line: “Kopi Kita is a trading name of KK Food Ventures Sdn. Bhd. (Company No. 201801012345 (1234567-X))”
Footer shows only the old-format number (1234567-X) A registration ID that does not match the 12-digit SSM record Show the 12-digit number with the old number in brackets, the format SSM uses
LinkedIn company page under the Sdn Bhd name, Instagram and GBP under the trading name Two organisations with overlapping staff Use the trading name as the page name on every profile and put the legal name in the “About” line
GBP name “Kopi Kita Bangsar – Best Coffee KL” A business whose name includes keywords, which also breaks GBP guidelines GBP name = the real-world signage name, nothing else
Schema name = trading name, no legalName No machine-readable link to the SSM entity Add legalName, alternateName and an identifier for the SSM number (Section 8)
Founding year 2018 on About page, 2020 on LinkedIn (date of rebrand) Two inception dates, so possibly two entities Use the incorporation year everywhere; mention the rebrand as an event

Caption: six recurring Malaysian name-mismatch patterns and the one-line fix for each. Source: illustrative example built on a fictional business (“Kopi Kita”, “KK Food Ventures Sdn. Bhd.”, registration numbers invented for the example). The patterns are those we check in every Ai Mode entity audit; we do not publish a frequency count because our audited sample is too small to state one honestly.

The SSM side has one more wrinkle. SSM’s name-comparison rules ignore words such as “Sdn”, “Bhd”, “Group”, “Holdings” and a trailing “Malaysia” when deciding whether two company names are the same. Entity systems do the opposite: “Kopi Kita Group” and “Kopi Kita” are different strings until something proves otherwise. A name SSM treats as identical can still look like a second company to an AI engine.

7. How do you measure entity consistency? The Entity Agreement Rate

The Entity Agreement Rate is the share of your 8 core identity facts that match across the 10 sources an AI engine is most likely to read about you. Score each fact on each source as match, mismatch or absent, then divide matches by filled cells. It turns “is our brand consistent?” into a number you can re-check every quarter.

The 8 facts:

  1. Trading name
  2. Registered company name (Sdn. Bhd.)
  3. SSM registration number (12-digit)
  4. Address (or service area for non-storefront businesses)
  5. Phone or WhatsApp number
  6. Primary category (“specialty coffee roaster”, not “F&B solutions provider”)
  7. Founder or principal person
  8. Founding year

The 10 sources: pick them, do not guess them. Run your brand name through Google (incognito, Malaysia) and take the first 7 non-social results that mention you, then add your Google Business Profile, your LinkedIn company page, and the first source ChatGPT search or Perplexity cites when you ask “What is [brand]?”. Those are the documents retrieval actually reaches.

The formula: Entity Agreement Rate = cells that match your canonical record ÷ cells where the source states that fact at all. Absent cells are tracked separately, as the coverage gap.

Source Trading name Legal name SSM no. Address Phone Category Founder Year
Own website ✓ ✓ ✗ old format ✓ ✓ ✓ ✓ ✓
Google Business Profile ✗ keyword suffix absent absent ✓ ✓ ✗ “Cafe” absent absent
LinkedIn ✗ legal name used ✓ absent ✗ old address absent ✓ ✓ ✗ 2020
Directory A ✓ absent absent ✗ old address ✓ ✓ absent absent
News article ✓ absent absent ✓ absent ✓ ✓ ✓
Remaining 5 sources 13 match, 7 mismatch, 20 absent across the 40 cells
Total (80 cells) 33 match, 14 mismatch, 33 absent

Caption: worked Entity Agreement Rate for a fictional business. 33 ÷ 47 filled cells = 70%. Coverage: 47 of 80 cells filled (59%). Source: illustrative model, not client data. Figures are constructed to show the arithmetic.

Reading it: 70% agreement means roughly 3 in 10 facts an engine reads about this business contradict another source. In the rows shown, the trading name and the address each disagree on two sources. Start with the trading name, because it is the label every other fact attaches to; the old address on LinkedIn and a directory comes second.

Our working rule, which is a practitioner threshold and not a measured benchmark: fix any fact that disagrees on 3 or more of the 10 sources before you publish anything new, because new content gets read alongside those contradictions.

Send us your brand name and we will score your Entity Agreement Rate across all 80 cells, free, and mark the three fixes that move it furthest.

How do you fix entity contradictions? 6 steps, fastest layer first

Fix in the order engines re-read: your own site and schema (next crawl), Google Business Profile (usually days), the directories and profiles in your 10-source set, then third-party coverage. Training memory comes last because you cannot touch it directly; you can only make the corpus agree before the next model trains on it.

  1. Write the canonical record. One internal document with the 8 facts as they should appear everywhere, including the exact trading-name spelling and the 12-digit SSM number. Every later step copies from it.
  2. Put both names on one page. The footer line from Section 6, plus the same sentence on your About page, which is the page engines most often fetch for “who is [brand]”.
  3. Ship Organization schema that links the names. Google documents legalName, alternateName and sameAs for Organization markup and says some properties help it disambiguate your organisation from others. Google’s May 2026 guide is equally clear that no special schema is required for generative AI features, so treat this as disambiguation evidence, not an AI ranking lever.
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Kopi Kita",
  "legalName": "KK Food Ventures Sdn. Bhd.",
  "alternateName": ["Kopi Kita Bangsar"],
  "identifier": {
    "@type": "PropertyValue",
    "propertyID": "SSM Company No.",
    "value": "201801012345"
  },
  "foundingDate": "2018",
  "url": "https://www.example.com.my/",
  "sameAs": [
    "https://www.linkedin.com/company/example",
    "https://www.instagram.com/example",
    "https://www.wikidata.org/wiki/Q00000000"
  ]
}

The identifier with an SSM propertyID is a convention, not a Google-documented field for Malaysia. Only add the Wikidata URL if an item genuinely exists. 4. Correct Google Business Profile. Signage name only, the most specific primary category, and the same phone number as your site. In our July 2026 pilot, 5 of the 8 commercial queries returned a Local Pack, so for local businesses this is often the first record a searcher sees. Engine-specific steps are in our Perplexity citation guide and Claude citation guide. 5. Work through the 10-source table. Edit what you control (LinkedIn, Facebook, Instagram, directory listings you claimed). Email the publisher for what you do not. Re-score after 30 days. 6. Add corroboration, not volume. Two independent sources stating the correct facts beat ten more pages on your own site. This is where forums like Lowyat and Reddit and digital PR come in. Google’s own guide warns that seeking inauthentic mentions is not as helpful as it might seem, so the mentions have to be real.

When is entity work not worth paying for?

When there is nothing yet to reconcile. If a Google search for your exact brand name returns fewer than 5 results that are about you, your problem is coverage, not consistency, and an entity audit will mostly report “absent”. Spend that budget on a correct Google Business Profile and two or three pieces of independent coverage first.

Three other cases where we would tell you not to buy this yet:

  • You are mid-rebrand. Scoring before the new name is live measures a record you are about to discard. Wait until the SSM name change, site and GBP are all updated, then score. (After a name change under Sections 28 or 29 of the Companies Act 2016, the former name has to stay displayed beneath the new one for 12 months, which is useful entity evidence in itself.)
  • Someone is selling you a Wikidata item as a visibility product. If your business does not meet the notability criteria in Section 5, a created item is likely to be deleted, and paying for it buys nothing durable.
  • You want schema to fix it on its own. Markup that says “legalName: X” on a site where the visible footer says “Y” is a contradiction inside your own page.

10. Conclusion

AI engines decide what your brand is from training memory, live retrieval and, on Google, a graph that stores you as an ID with several names attached. All three reward the same thing: the same 8 facts, stated the same way, across the sources they read. For Malaysian businesses the most common failure we see is a correct Sdn Bhd name and a correct trading name that never appear together.

Score your Entity Agreement Rate, fix the column with the most mismatches, and re-score in 30 days. The retrieval layer will show the change within weeks. The training layer will show it at the next model release, provided the web agreed with you before that run.

Want the 80-cell score for your brand and your closest competitor side by side? We will run both and send the table, no call required.

11. Frequently Asked Questions

What is entity SEO? Entity SEO is the practice of making every source about your business state the same core identity facts, so search engines and AI models resolve them to one entity. It covers names, registration details, address, category, founder and founding year across your site, schema, Google Business Profile, directories and press. The goal is a single unambiguous record, not more keyword rankings.

How is the Google Knowledge Graph different from a knowledge panel? The Knowledge Graph is Google’s underlying store of entities, each held under a machine ID with linked facts and identifiers such as a Wikidata QID. A knowledge panel is one visible display of that record in search results. A business can exist in the graph without a panel, and Google generates panels automatically once enough information about the entity is available.

What is the Entity Agreement Rate? The Entity Agreement Rate is the share of 8 core identity facts (trading name, legal name, SSM number, address, phone, category, founder, founding year) that match across the 10 sources an AI engine is most likely to read about you. Divide matching cells by filled cells. Track empty cells separately as your coverage gap.

Should my website show my Sdn Bhd name or my brand name? Both, on the same line. Section 30(2)(b) of the Companies Act 2016 requires your registered company name and registration number on your website, and a sentence such as “Brand is a trading name of Company Sdn. Bhd. (Company No. 12-digit number)” also gives entity systems direct evidence that the two names describe one business.

Can I create a Wikidata item for my own company? You can, but it may not survive. Wikidata requires serious, publicly available references, treats self-promotional items as unwelcome, and regards editing about your own organisation as a conflict of interest. Items referenced only by the company’s own website and social pages are routinely nominated for deletion. Earn independent coverage first.

How long does it take for AI engines to reflect corrected brand facts? It depends on the layer. Retrieval-based answers (AI Overviews, AI Mode, ChatGPT search, Perplexity) can change once the corrected pages are recrawled, often within weeks. Training memory changes only when a new model is trained, which no engine schedules per brand. Re-score your Entity Agreement Rate 30 days after fixes.


Nnabuike Precious

Nnabuike Precious is Lead SEO Strategist. He specialises in technical SEO, crawl optimisation, and revenue attribution. He has worked on campaigns for brands including Grab SEA, SOCAR, Trevo, Bateriku, and Kaspersky.

Author posts