AIO & LLMO for IR | When AI Mistakes You for Another Company — Name Changes, Restructuring, and Identifiers

When IR departments verify what generative AI says about their company, most of them start from the figures. Is revenue right? Are the dividends correct? Is the fiscal period off? There is, however, one more question in front of those. Is the “this company” that the answer points to really your company?

You changed your trade name. You sold a subsidiary. You merged. You transferred a business. A different company with the same reading exists. In these cases the AI's answer may not be “inaccurate about your company” but rather “not about your company at all.” Checking figures cannot answer that question, because the denominator is different.

This article sorts out the situations in which the entity shifts, confirms what securities codes, ISINs, Corporate Numbers and LEIs identify against the primary material of the bodies that assign them, clears up the misunderstanding surrounding sameAs, and goes through to the measurement procedure for checking whether you are being identified correctly.

One thing to say at the outset: this article does not claim that writing identifiers makes AI recognize a company correctly. Within the scope of this review, we could not confirm such a demonstration. What can be written stops at “do not confuse the code that identifies a legal entity with the code that identifies an issue.”

What this article covers

• Where the “entity shift” that figure-checking cannot catch arises (five situations)
• How a change of trade name, a corporate restructuring and a tender offer are positioned under the timely disclosure rules
• What a securities code, an ISIN, a Corporate Number and an LEI each identify (a comparison table confirmed against primary sources)
• The error of putting an identifier itself into sameAs, and where the specification puts it instead
• Concrete steps for measuring whether you are being identified correctly, and what must not be measured

Who this is for

IR staff at listed companies, and in particular those at companies that have been through, or have ahead of them, a change of trade name or a corporate restructuring. As secondary readers we have legal departments and corporate planning departments in mind.

Numbers worth holding onto

• Among the Decisions by Listed Companies for which the Tokyo Stock Exchange requires timely disclosure, a change in a trade name or a corporate name, corporate restructuring such as a merger, a transfer or acquisition of all or part of a business and a takeover bid each appear as separate items (Japan Exchange Group, as of July 10, 2026).
• An evaluation of 17,621 listed companies and 197,011 questions observed numerical hallucination even where the company name and the fiscal year were stated (arXiv:2504.00042, accepted at COLM 2025).

Frameworks for verifying AI answers in the IR domain are shared to a reasonable degree. Specify the fiscal period, state whether the figure is consolidated or non-consolidated, check the currency and the unit, align the accounting standard, and distinguish actuals from forecasts. Build a reference table carrying those attributes and match the AI's answers against it — that is the basic form.

This framework is a tool for pinning down “which figure”. It rests on a premise, though: that “whose figure” has already been settled.

1-1. Without a settled entity, the reference table does not work

Each row of a reference table takes a form like “our company, FY ending March 2026, consolidated, IFRS, revenue, millions of yen,” and the row holds together only where “our company” is uniquely determined.

Suppose a company has changed its trade name and the AI answers under the former name. If the former and current names point to the same legal entity, the name is old but the entity is right. If the former name comes with figures from the period when that name was in use, the entity is right but the point in time is old. If the figures belong to a different company with a name resembling the former one, the entity is wrong. These three cannot be told apart by checking figures, because all of them look like nothing more than “the numbers do not match.”

1-2. A problem on the time axis is a different thing from a problem on the entity axis

The same “100 billion yen” is different information for an investment decision depending on whether it is an initial forecast, a revised forecast, an actual, or a corrected figure. That is a matter of the time axis.

What this article deals with is the entity axis. Whose 100 billion yen is it? The surviving company before the merger, or the company after it? Does it include the business that was transferred out? A shift on the time axis can be resolved when a new disclosure updates it, whereas a shift on the entity axis is not resolved by an update, because the object being referred to is a different one to begin with.

1-3. Specifying the company name does not guarantee the accuracy of the answer

You might think it is enough to state the entity explicitly. Naming the company, however, is not a sufficient solution.

A research team at the Georgia Institute of Technology built 197,011 questions about the revenue of US listed companies and had several large language models answer them. The scope was the 43 years from 1980 to 2022, and 17,621 companies. The prompt was a fixed sentence — what was {company}'s revenue in {fiscal year}? — in which both the company name and the year are specified. Even so, answers differing from the correct value by more than 10% were observed (arXiv:2504.00042. A peer-reviewed paper accepted at COLM 2025; this article refers to the arXiv version).

Revenue came from Compustat and was converted into millions of US dollars with differences in accounting standards left unharmonized, and the API runs were carried out in February 2025. Within those conditions, what can be read is that stating the company name does not secure the accuracy of the figures. What this study verified, however, was the accuracy of revenue QA, not whether the model was able to identify the correct legal entity.

In a second study, using Japanese earnings reports (kessan tanshin) from 2019 to 2023, sentiment scores for the same disclosure text changed depending on whether the company name was stated or removed. The direction of the bias was not consistent across models (arXiv:2411.00420. 10,249 observations with GPT-4o). The label of a company name can change the evaluation given to identical text. This too was not a study that verified the accuracy of entity identification.

What the two together support goes as far as this: giving the company name does not mean a correct answer comes back. Whether the entity has been identified correctly cannot be judged from these two studies. That is precisely why it has to be measured separately (§8).

The general typology of errors and the basic design of their measurement are left to the existing articles in this lane. For the overall picture see How AI Describes Your Financials, and for the basic form of the measurement procedure see How to Verify What AI Says About Your Financials.

Here are five situations in which the outline of the entity moves. What follows is a sorting of types, not a ranking by frequency of occurrence.

2-1. Change of trade name

The legal entity is unchanged and the name alone changes. Continuity of the entity is preserved, but two names are now circulating outside the company. Past news articles written under the former name, third-party databases still carrying it, URLs and file names containing it, past disclosure documents, prior-year pages on your own site. It takes time for information under the new name to accumulate.

The point to keep in mind is that information under the former name is not an error. It was correct at the time. That is precisely why deletion or correction does not deal with it. Unless it comes paired with the information “from when to when that name applied,” a description under the former name competes with the current description.

2-2. Merger

Several legal entities become one. The Securities Identification Code Committee is explicit about the treatment. Where listed companies merge, the securities code and sector of the surviving listed company are used. Where a newly established parent company is newly listed through a share transfer or the like, a new securities code and sector are assigned (Securities Identification Code Committee, FAQ).

In an absorption-type merger, then, the code on the side of the extinguished company falls out of use, and in a share transfer a new code comes into being. Information accumulated under the name and code of the side that disappeared does not attach itself automatically to the succeeding entity.

2-3. Company split

Part of a business is succeeded to by a different legal entity. Before and after the split, the range of business that a company under the same trade name accounts for changes. Consolidated revenue before and after the split covers a different scope under the same name. Because neither the name nor the legal entity changes, unlike a change of trade name or a merger, no signal that the entity has changed appears in the name.

2-4. Business transfer and changes in subsidiaries

You sold a business, or a consolidated subsidiary left the scope of consolidation.

What characterizes this situation is that your own site can become a supply source of incorrect information. The page introducing the transferred business, press releases about that business, the PDF of a medium-term management plan that includes it, a list of group companies that still includes the subsidiary. If these remain, then even where AI says “this is still one of the company's businesses,” the source it refers to is your own writing.

In that a candidate cause exists on your own side, this situation differs in character from the others. External information sources cannot be controlled, but taking stock of your own site is within your discretion.

2-5. Tender offer

A transaction capable of affecting control and the like. During the period of a tender offer, what is said about the target company includes events that have not been completed. “Was acquired,” “an acquisition has been proposed” and “an acquisition has been announced but is in process” are different states, but external accounts may describe them with the same vocabulary, and where AI summarizes that material, the distinction between states is easily lost.

On when each restructuring act takes legal effect, this article does not go further. The point at which effect arises differs with the type of act, the progress of the procedure and whether registration has been made, and we have not completed primary confirmation for this article. Please be sure to check this point with your legal department.

Among the corporate information for which the Tokyo Stock Exchange requires timely disclosure, the five situations of the previous section appear as separate items under Decisions by Listed Companies (as of July 10, 2026).

Situation in the previous section	Decision by Listed Companies as set out by TSE
Change of trade name	26. Change in a trade name or a corporate name
Merger / company split	10. Stock swap, stock transfer, merger, demerger
Business transfer	13. Transfer or acquisition of all or part of a business
Change in a subsidiary	17. Transfer or acquisition of shares or equity interest accompanied by a change in a subsidiary or other matters accompanied by a change in a subsidiary
Tender offer	11. Takeover bid or takeover bid for own shares

(Japan Exchange Group, Corporate Information Required for Timely Disclosure)

3-1. What can be read from this

A change of entity is positioned as an event to be disclosed in its own right, separately from a change in the figures. A change in a trade name or a corporate name in particular is set out as a stand-alone decision unconnected to the figures. The rules regard a change in “whose figures these are” as information in itself, and that is consistent with this article's approach of confirming identity of the entity before the accuracy of the figures.

3-2. What cannot be read from this

On the other hand, the following cannot be written.

There is no fact that the Tokyo Stock Exchange requires management of corporate perception on AI. Developing an IR system is an obligation under the Code of Corporate Conduct, and the state of that development has to be disclosed in the Corporate Governance Report (Japan Exchange Group). The specific content of the system, however, is left to each company's judgment.

The timely disclosure rules are also rules of the exchange rather than national law as such, and differ in character from the Financial Instruments and Exchange Act and the Fair Disclosure Rule. What this article can say goes as far as the question of how far entity confusion on AI is to be regarded as within the scope of management, inside the existing framework of an IR system.

3-3. A timely disclosure does not necessarily reach AI

Timely disclosures are made public in real time through TDnet. Public browsing runs for 31 days and search by company for 10 years, however, and past data and APIs are provided as paid information (Japan Exchange Group).

The fact of having disclosed, then, means that the information has been placed in a state where it can be retrieved mechanically. It does not extend to each AI service retrieving it, giving it priority over prior knowledge, and interpreting it correctly.

The exchange's disclosure data format (XBRL) and the markup of web pages (schema.org) are separate things. That distinction, and the retrievability of IR sites, are handled in a separate article in this lane, so we do not repeat them here. Please see Do AI Crawlers Read Your IR Site's JavaScript?.

Trade names are not unique within Japan. Several legal entities may hold the same trade name, and legal entities may exist whose readings are the same while their written forms differ.

This article does not name real companies as instances of entity confusion. It touches on the credit of specific companies, and we have not been able to confirm a published list of primary instances for this article. What follows is a sorting of types.

4-1. The structure of confusability

Entity confusion is more likely to arise with combinations such as the following.

Form of confusion	Content
Identical trade name	A different legal entity exists whose trade name matches exactly
Same reading	The written form differs but the reading is the same (kanji / katakana / hiragana / Latin letters)
Colliding abbreviations	The formal trade names differ, but the abbreviations used in the market overlap
Close English names	The Japanese trade names differ but the English forms resemble each other
Proximity within a group	The names of a holding company and an operating company, or of a parent and a main subsidiary, are close
Listed and unlisted mixed together	An unlisted company of the same name exists and its information is mixed in

“Proximity within a group” is easily overlooked. Under a holding company structure the arrangement tends to be that the figures investors want are at the consolidated level (the holding company) while the description of the business sits with the operating company, and these two entities can end up mixed within one answer.

4-2. Why this surfaces with AI

In conventional search, several candidates were listed side by side and which company was meant was resolved on the user's side. A generative AI answer is consolidated into a single account. In the course of that consolidation, which candidate was taken disappears from the surface, and the user cannot even see that a choice was made.

A benchmark study of financial QA reports that even in the “high confidence” state where eight re-answers to the same question all agree, 15–23% were wrong (arXiv:2607.11414. A preprint released in July 2026; an evaluation of foundation models rather than of public AI search services). Consistency of answers does not mean the entity is correct.

4-3. The direction of response

The existence of a different company with the same name is not something you can change. What can be changed is whether your own descriptions are in a form that pins the entity down. Use the formal trade name including the corporate form, keep the English trade name consistent, put verifiable attributes such as the address and the year of establishment alongside it. These, however, are not measures by which “AI will then recognize you correctly”; they go as far as placing material that can be matched.

Pinning down the entity tends to turn into a discussion of identifier codes, but what has to be sorted out first is “what does that code identify?”

What securities codes and ISINs identify is the issue; what Corporate Numbers and LEIs identify is the corporation, the legal entity. Treat these two as one “company code” and a restructuring becomes impossible to sort out. Everything below was confirmed against the primary material of the bodies that assign them (confirmed on August 31, 2026).

5-1. Comparison table

Identifier	What it identifies	Structure	Assignment and management	Treatment on a name change or restructuring
Securities code (issue code)	The issue, such as a listed stock	4 characters. Historically digits alone (specific name codes for stocks run from “1300” to “9999”). Codes newly set from 2024 onward may contain letters	Securities Identification Code Committee	On a merger, the securities code and sector of the surviving listed company are used. Where a newly established parent company is newly listed through a share transfer or the like, a new one is assigned
ISIN	The issue of a security that is listed or publicly offered (international)	12 characters = 2-character country code + 9-character basic code (new securities code) + 1 check digit. ISO 6166	For equities, assigned by the body of the country to which the issuer belongs. In Japan, the Securities Identification Code Committee	Tied to the issue, so it follows the issue as that is extinguished or newly created
Corporate Number	The corporation (the entity for registration and tax purposes)	13 digits	Designated by the Commissioner of the National Tax Agency	Even where there is a change of name or address, a Corporate Number once designated is not changed
LEI	The legal entity participating in financial transactions	20 alphanumeric characters. ISO 17442	Issued by an LEI issuer (LOU) accredited by GLEIF, and managed by GLEIF	Has a mechanism that records name changes, mergers and the like as Legal Entity Events, with a field holding the succeeding entity

5-2. Securities codes — letters incorporated from 2024

There is a change that has to be taken in. Specific name codes for stocks set on or after January 1, 2024 use letters. The digits used are either or both of the second and the fourth counting from the front. The uppercase letters used number 19, excluding B, E, I, O, Q, V and Z (Securities Identification Code Committee). Codes set before that are not changed. Four-digit numeric codes and codes containing letters coexist in the market, so if you have systems or descriptions premised on four digits of numerals, that premise does not hold in general.

5-3. ISIN — the international code for an issue

An ISIN (International Securities Identification Number) is a 12-character code laid down in ISO 6166 and used as data for the settlement, custody and circulation of securities. For equities, the ISIN is assigned by the country to which the issuer belongs. ISINs for foreign equities are not assigned in Japan; the code agency of the country to which the issuer legally belongs assigns them (Securities Identification Code Committee, FAQ).

Here too what the ISIN identifies is the issue. One legal entity may hold several issues, and an issue may be extinguished while the legal entity survives. An ISIN is not a company ID.

5-4. Corporate Number — the issuing body itself writes that it does not prove identity

A Corporate Number is a 13-digit number held by corporations and the like, such as kabushiki kaisha, designated by the Commissioner of the National Tax Agency and free for anyone to use. What is published is the Basic 3 information, namely (i) the trade name or name, (ii) the address of the head office or principal place of business, and (iii) the Corporate Number. In addition, the phonetic reading of the name, English forms of the name and address (where these have been registered), and change history information for the name and address are provided (National Tax Agency, About Corporate Numbers).

In relation to a change of trade name, the following statement is decisive. Even where there is a change of name or address, a Corporate Number once designated is not changed. For corporations registered for incorporation, registering a change of trade name or a relocation of the head office at the Legal Affairs Bureau means that the information is passed from the Ministry of Justice to the National Tax Agency and the information on the publication site is updated (National Tax Agency, Frequently Asked Questions).

One point that is easily missed. Where an English form has been registered, a further registration procedure is required. A registered English form does not follow automatically, which means a former English trade name can remain as officially published information.

On the limits of identifiers, the National Tax Agency writes clearly itself.

The Corporate Number is designated on the basis of registration, tax filings and the like, for the purpose of using its function of identifying a particular corporation or organization to make administration more efficient and to reduce the administrative burden on companies; it does not necessarily prove that the corporation or organization exists. Nor does it prove that all corporations or organizations using the same Corporate Number are one and the same.

(National Tax Agency, Frequently Asked Questions. Translated from the Japanese)

The same page also records that a Corporate Number is designated even for a dissolved corporation as long as the registration record has not been closed, and that where the address has moved but no change procedure has been carried out, the pre-change address is what is published.

So an identifier is not a certificate of identity. The body that assigns it says as much itself. Any discussion of identifiers should start from here.

5-5. LEI — self-declaration at the base, complemented by outside information

An LEI (Legal Entity Identifier) is a 20-character alphanumeric code identifying legal entities that participate in financial transactions. Based on ISO 17442, it is managed by GLEIF (the Global Legal Entity Identifier Foundation). GLEIF is a not-for-profit foundation with the G20 and the Financial Stability Board (FSB) behind it, supervised by the Regulatory Oversight Committee (ROC), and LEIs are issued by LEI issuers (LOUs) accredited by GLEIF. In structure, characters 1 to 4 are the issuer prefix, 5 to 18 are the entity-specific part and 19 to 20 are check digits, the entity-specific part being designed to carry no meaning (GLEIF). GLEIF explains that the LEI answers “who is who” (Level 1 data) and “who owns whom” (Level 2 data) (GLEIF).

What deserves attention in relation to restructuring is Legal Entity Events. Under the ROC's definition these are significant changes in an entity's reference data, life cycle or organizational structure, and according to GLEIF there are currently 20 types. They include changes in ownership, mergers and acquisitions, changes of name or address, and the ending of an entity (GLEIF). The data format has items such as the type of event, the legal effective date, the recorded date and the type of validation documents (GLEIF LEI-CDF 3.1).

For all the structure the LEI has, the ROC's document contains a statement to this effect. Without a declaration by the affected entity itself, an LEI issuer cannot necessarily know that an event requiring an update of reference data has occurred (LEI ROC).

The same document also sets out complementary means of detecting changes, through challenges from third parties (data challenges) and external data feeds. The design therefore has self-declaration at the base, with outside information complementing it. An identifier is not a device that becomes correct by itself; it is something maintained.

5-6. So what can be said, and what cannot

What can be said comes down to a single point.

The code that identifies a legal entity and the code that identifies an issue are different things. In a restructuring, these two move separately.

What cannot be said should be made equally clear.

We cannot write that “stating an LEI or a Corporate Number makes AI recognize the company correctly.”

Within the scope of this review, we could not confirm such a demonstration. Nor have we confirmed a study that isolated and measured whether stating an identifier contributes to resolving the entity.

When people try to indicate the entity in structured data, there is an error seen frequently: putting an identifier string such as a securities code or an LEI into sameAs. This runs against the specification.

6-1. In the specification, what sameAs expects is a URL

The definition of sameAs in schema.org is as follows. The URL of a reference web page that unambiguously indicates the item's identity — for example, the item's Wikipedia page, its Wikidata entry, or the URL of its official website. The type expected as the value is URL, and it applies to Thing (schema.org).

sameAs, in other words, is a property for indicating that a reference page describing this entity exists elsewhere. It is not the place to put an identifier itself.

6-2. Identifiers have separate properties provided for them

schema.org provides separate properties for identifiers.

Property	Definition	Applies to
leiCode	An organization identifier that uniquely identifies a legal entity, as defined in ISO 17442	Organization
tickerSymbol	The exchange-traded instrument associated with a Corporation. Expressed with the exchange and the issue name separated by a space; use of the ISO 15022 market identifier code (MIC) is recommended for the exchange part	Corporation
identifier	A general-purpose identifier. Expressed as text, a URL, or a PropertyValue (a pair of propertyID and value)	Thing
legalName	The official name of the organization, such as the registered company name	Organization
alternateName	An alias	Thing

(schema.org leiCode, schema.org sameAs)

That tickerSymbol is a property of Corporation rather than Organization is a point to check at implementation.

6-3. An example

Where the entity is to be written out in machine-readable form, a way of writing it that follows the specification looks like this.

{
  "@context": "https://schema.org",
  "@type": "Corporation",
  "name": "(current trade name)",
  "legalName": "(formal trade name including the corporate form)",
  "alternateName": ["(English trade name)", "(commonly used abbreviation)"],
  "url": "https://example.com/",
  "leiCode": "(20-character LEI; where one has been obtained)",
  "tickerSymbol": "(the exchange's MIC) (issue code)",
  "identifier": {
    "@type": "PropertyValue",
    "propertyID": "https://www.houjin-bangou.nta.go.jp/",
    "value": "(13-digit Corporate Number)"
  },
  "sameAs": [
    "https://(URL of your company profile page)",
    "https://(URL of your IR site)"
  ]
}

Points to be explicit about.

• The above is a way of writing that follows the schema.org specification, not a claim that “writing it this way works on AI.” Effect is a domain to be measured and confirmed.
• Do not write identifiers you have not obtained. An LEI is stated where one has been obtained. Every value has to be one that your company actually holds and uses.

6-4. How to handle a former trade name

At a company that has been through a change of trade name, the question arises whether the former name should go into alternateName. The specification allows writing it as an “alias,” but a former trade name is not a current alias; it is a past name. Listed without distinction, it leaves room for the former name to be treated as though it were the current appellation.

This article takes the view that, rather than placing the former trade name as a machine-readable attribute, stating on a history page, in visible text, from when to when each trade name applied leaves less room for misunderstanding. This is not a constraint in the specification; it is this article's judgment.

The first condition to satisfy in putting structured data in order is that the content marked up is also visible to the reader of the page.

7-1. What Google asks for

In its general guidelines for structured data, Google asks that content not visible to readers of the page not be marked up. Structured data is required to be an accurate representation of the content of the page, and misrepresenting ownership, affiliation or primary purpose is likewise prohibited (Google, General structured data guidelines). It states further that structured data must not be added about information that is not visible to the user, even where the information is accurate (Google, How structured data works).

Translated back into descriptions of the entity, it comes to this. The trade name, the English trade name and the identifiers written into JSON-LD have to be displayed on screen as well. Writing the correct trade name in the structured data alone while the former name remains on screen is not a desirable state under the guidelines either.

7-2. What can be confirmed about AI features

Google states that the condition for appearing as a reference link in AI Overviews or AI Mode is that the page is indexed and can be shown with a snippet, and that there are no additional technical requirements. It states explicitly that no special schema.org markup exists for generative AI features, that machine-readable files such as llms.txt or Markdown versions need not be prepared, and that Google Search does not use them (Google AI features). Accordingly, within the scope of this review, we could not confirm any “special entity markup for AI.”

At the same time, “there are no AI-specific requirements” and “there is no AI-specific processing” are different things. Google explains that AI Overviews and AI Mode may use query fan-out (a technique that issues several related searches across subtopics and data sources), and that the two may use different models and technologies. The absence of additional requirements does not permit the conclusion that the same processing as ordinary search is being carried out.

7-3. Taking stock of what has to be made to match

Bringing “match what is displayed on screen” down to practice makes it the work of finding out where the trade name is written. It tends to be scattered across the following places.

Place	What is easily overlooked
The copyright line in the footer	Written straight into the template, so updates are missed
The company profile page	The formal trade name and the common name are mixed
The IR site header and title tag	Updates lag where the site is on a separate domain or a separate system
Prior-year news releases	Still carry the trade name of the time. Without an explicit date they read as present tense
The cover, header and footer of PDF materials	Where the body is an image, checking the wording in bulk is difficult
English pages	Update timing diverges from the Japanese pages
The list of group companies	Companies already transferred or sold remain on it
Separate domains such as recruiting and product sites	Outside the IR department's control, so they drop out of the stocktake

This stocktake is not carried out as an AI measure; it is a matter of consistency in how disclosed information is described. That the material AI reads is put in order as a result is the accurate order in which to see it.

Everything sorted out up to here exists for this section. There is no way to know whether the entity is being identified correctly other than by checking. The identifiers are in place so it must be fine; we disclosed so it must have got through; we fixed the site so it must be resolved — each of these is a hypothesis, and a hypothesis is something to be measured and verified.

8-1. Add an “entity column” to the reference table

For checking figures, you prepare a reference table with attributes such as the fiscal period, consolidated or non-consolidated, the accounting standard, the currency and unit, and the release date. Judging the entity requires different columns there.

Column	Content
Current trade name	The formal form including the corporate form
Former trade names and periods of use	From when to when each name applied (start and end dates)
English trade name	The registered or published form. Including whether abbreviations are used
Abbreviations	The appellations actually used in the market and in reporting
Corporate Number	13 digits
Securities code	And the ISIN where needed
LEI	Where one has been obtained
Main consolidated subsidiaries and dates of change	When each entered consolidation and when it left
Businesses transferred or sold and effective dates	What left the company, and when
Confusable legal entities	Legal entities with the same or a similar name that could be confused with you (as far as you are aware)

This is a reference table for judgment, not a description meant to be read by AI as it stands. With this table in place, you can finally say objectively that “the AI's answer has the entity wrong.”

8-2. Five ways of asking

(1) Ask about the same company under several appellations

The current trade name, the former one, the abbreviation, the English trade name, the issue code. Put each in the subject position of the same question and see whether the answers point to the same entity. If the answer changes with the appellation, resolution of the entity depends on the appellation. Whether asking under the former trade name returns current information, information from the period when that name was used, or an entirely different company, are three separate states.

(2) Ask about businesses that have been sold or transferred

Ask “tell me about your ◯◯ business” with a business you have already let go in the subject position. If it is still spoken of as one of your current businesses, the separation has not been reflected. Following the information sources the answer refers to can lead back to leftover pages on your own site.

(3) Check for mixing with same or similar names

Where a legal entity confusable with you exists, run questions that include the attributes distinguishing the two (address, year of establishment, listed market, main business) separately from questions that do not. If information about the other company mixes in the moment the attributes are removed, the name alone is not settling the entity.

(4) Look at the attribution of the URLs cited

Where sources are presented with the answer, classify whether each URL is your own, an old site or old domain, or a third party's. Even on your own domain, an un-updated prior-year page or the page of a transferred business makes it an error caused by your own writing. Where the answer depends on external sources, check the possibility that the entity description in that source is out of date.

(5) Take stock of how the trade name is written across your own site

Work through the list in the previous section. There is one thing to look at. Does the form machines read match the form people read? The new trade name on screen, the former one in the structured data, another form on the cover of a PDF — this state is not unusual.

8-3. Decide the judgment categories in advance

Treating answers as a binary of right and wrong makes it impossible to separate causes. For the entity, dividing them as follows leads to a response.

Judgment	Content	Where the main cause sits
Entity matches	The entity the answer points to is your company	—
Wrong entity	A different legal entity is described as your company	External information sources / confusion of names
Merged	Information about you and about another entity is mixed together	Name proximity within the group / same-name legal entities
Separation failure	A transferred or split-off scope is still spoken of as yours	Leftover pages on your own site
Name out of date	The entity is right but the name is a past one	Surviving former-trade-name information
Unidentifiable	Which entity is meant cannot be determined	Ambiguity in the description

Of these, separation failure is the type that is easiest to move on, because the candidate cause can be checked on your own side.

8-4. Repeat under conditions that carry no context over

Asking questions in sequence within one conversation lets the immediately preceding exchange help resolve the entity. Continue with “about that company you mentioned just now” and the answer comes back with the entity already fixed.

What investors actually do, however, is in most cases a single question from a blank state. To measure standing perception, you need to repeat the same question under conditions that carry no context over.

And one answer is no more than a single sample. Generative AI answers vary from run to run. A single answer obtained from a single model with a single prompt cannot be regarded as standing perception. Decide the number of repetitions and the conditions in advance.

What to keep on record is five things: the wording of the question, the date and time it was run, the body of the answer, the source URLs presented, and the judgment. Without the wording of the question and the date and time in particular, you later become unable to judge whether something has been resolved or is within the range of variation. Being able to measure again under the same conditions after fixing your own site is the purpose of the record.

8-5. What to use as indicators

As indicators to track from the standpoint of the entity, the following can be considered (these are this article's design proposals, and are neither optimal values derived from research nor indicators established in the industry).

Indicator	Definition
Entity match rate	The proportion of all answers in which the entity pointed to was your company
Former-trade-name mention rate	The proportion in which a former trade name was used in the present tense in the body of the answer
Transferred-business residue rate	The proportion in which a business you have let go was spoken of as still yours
Own-domain citation rate	The proportion of the sources presented that are accounted for by your own domain
Leftover-page origin rate	The proportion of own-domain citations accounted for by un-updated prior-year pages

We cannot write “do this and confusion falls by X%.” Within the scope of this review, we could not confirm published data showing such an effect. Indicators do not promise a margin of improvement; they are a tool for watching a state continuously.

8-6. State the limits of the measurement up front

Finally, what this measurement cannot do, made explicit.

• The specific mechanisms by which each AI service identifies companies — within the scope of this review, we could not confirm them. Learned knowledge, general web search, partner data, search caches and material uploaded by users may be combined, but the breakdown is not known.
• Answers vary from run to run. Always allow for the possibility that an apparent improvement is within the range of variation.
• Because the information sources referred to can change with language and region, results in Japanese cannot be generalized to other languages.
• We have not confirmed a correction or appeal process for companies common to AI providers. Finding an error does not mean there is an established procedure for having the provider fix it. What can be done in practice is to put your own descriptions in order and measure again after a period.

Note that you must not attempt to correct an AI answer using material non-public information. Corrections have to be limited to re-presenting already published material. This point is dealt with in detail in the parent article of this lane.

In preparing this article, the following items could not be confirmed. We make them explicit so as not to write about things we have not confirmed as though they existed.

Item	Status
Published statistics showing how often companies are confused in AI answers	Within the scope of this review, we could not confirm them. We give no frequency figures
A demonstration that stating identifiers on a site or in disclosures improves entity identification in AI answers	As above
The specific mechanisms of company entity resolution at each AI service	Within the scope of this review, we could not confirm them
Measurements of the period for which AI continues to use a former trade name after a change	Within the scope of this review, we could not confirm measurements
The point at which each restructuring act takes effect under the Companies Act and the Financial Instruments and Exchange Act	We have not completed primary confirmation for this article, so we do not go into it
A correction or appeal process for companies, and processing deadlines, common to AI providers	We could not confirm such a process
Causation in which entity confusion moved a share price or the cost of capital	We could not confirm it

These do not mean “they do not exist.” They mean that within the scope of this review, we could not confirm them.

Q1. If we change our trade name, does the Corporate Number change too?

It does not. The National Tax Agency states explicitly that even where the name or address changes, a Corporate Number once designated is not changed. For corporations registered for incorporation, registering the change of trade name at the Legal Affairs Bureau means the information is passed to the National Tax Agency via the Ministry of Justice and the Corporate Number Publication Site is updated. No procedure with the National Tax Agency's Corporate Number Administration Office is required. Where an English form has been registered on the publication site, however, a further registration procedure is required, and a registered English form does not follow automatically.

Q2. What happens to the securities code after a merger?

The Securities Identification Code Committee states that where listed companies merge, the securities code and sector of the surviving listed company are used, and that where a newly established parent company is newly listed through a share transfer or the like, a new securities code and sector are assigned. Codes that fell out of use through delisting or the like from July 1993 onward are not reused, although codes deleted before that may have been assigned to a different company.

Q3. If we obtain an LEI, will AI start recognizing our company correctly?

Within the scope of this review, we could not confirm grounds for writing that. An LEI is a 20-character code based on ISO 17442 that identifies legal entities participating in financial transactions, and it is managed by GLEIF. The Legal Entity Events framework for recording restructuring is in place as well. We have not, however, confirmed a study that isolated and measured whether this contributes to entity identification in AI answers. The LEI is a mechanism for financial transactions and regulatory reporting.

Q4. How should we use securities codes, ISINs and Corporate Numbers differently?

They identify different things. What securities codes and ISINs identify is the issue, that is, the security being traded. What Corporate Numbers and LEIs identify is the corporation, the legal entity. One legal entity may have several issues, and an issue may be extinguished while the legal entity survives. In a restructuring these two families move separately, so trying to bundle them under the single concept of a "company ID" makes the situation impossible to describe.

Q5. Can we put a securities code or an LEI into sameAs?

That runs against the specification. What schema.org's sameAs expects as its value is a URL, pointing to a reference web page that unambiguously indicates the item's identity: a Wikipedia page, a Wikidata entry, an official website. It is not the place to put an identifier string. For identifiers there are leiCode (the legal entity identifier defined in ISO 17442), tickerSymbol (the exchange-traded instrument associated with a Corporation) and identifier (a general-purpose identifier that can be expressed as a PropertyValue).

Q6. Does the Tokyo Stock Exchange require management of corporate perception on AI?

It does not. Developing an IR system is an obligation under the Code of Corporate Conduct, and the state of that system has to be disclosed in the Corporate Governance Report. The content of the system, however, is left to each company's judgment, and measuring or managing perception of your company on AI is not mandated. The timely disclosure rules are also exchange rules rather than national law. What this article deals with is the range of the question of how far this issue is treated as within scope inside the existing framework of an IR system.

Q7. There is a different company with the same name as ours. What can we do?

The existence of a same-name legal entity is not something you can change. What you can change is your own descriptions. Using the formal trade name including the corporate form, keeping the English trade name consistent, and putting verifiable attributes such as the address, year of establishment and listed market alongside it are possible responses. These are not measures that "make AI recognize you correctly," though; they go as far as placing material that can be matched. Confirm the effect through measurement that asks under several names.

Q8. If we name the company in the question, will the entity not be resolved correctly?

Naming the company does not guarantee the answer is accurate. In an evaluation covering 17,621 US listed companies and 197,011 questions, answers differing from the correct value by more than 10% were observed even with a fixed prompt specifying both the company name and the fiscal year. In a study using Japanese earnings reports, sentiment scores for the same disclosure text also changed depending on whether the company name was present. However, neither was a study of entity identification accuracy as such. Whether the entity was identified correctly cannot be judged from these studies, and there is no way to confirm it other than measuring for yourself.

Q9. How often does this confusion actually happen?

Within the scope of this review, we could not confirm reliable published statistics showing the frequency. This article therefore gives no frequency figures. For reference, a benchmark study of financial QA reports that even in "high confidence" states, where eight re-answers to the same question all agreed, 15–23% were wrong (a preprint released in July 2026; an evaluation of foundation models, not of public AI search services). What is happening in your own case can be known no other way than by measuring it yourself.

Q10. Where should we start?

By adding entity columns to the reference table. The current trade name, former trade names and the periods they were used, the English trade name, abbreviations, the Corporate Number, the securities code, and transferred businesses with their dates. Without this baseline you cannot judge an answer to be "confusing us with someone else" even when you look at it. Once the baseline exists, put the same question under the current trade name, the former one, the abbreviation and the English trade name, and check whether the answers point to the same entity. In parallel, take stock of how the trade name is written across your own site. These two can be started independently.

Starting the verification of AI answers from checking the figures is natural. But checking figures presupposes that the entity is settled. You changed your trade name, you merged, you split a business, you let a business go, a company with the same name exists — in these cases it is the presupposition that has broken down.

• A change of entity is set out as a stand-alone item under the timely disclosure rules as well.
• Securities codes and ISINs identify the issue; Corporate Numbers and LEIs identify the corporation, the legal entity. In a restructuring the two families move separately.
• An identifier is not a certificate of identity. On the Corporate Number, the National Tax Agency states explicitly that it does not necessarily prove that a corporation or organization exists, and that it does not prove that all bodies using the same Corporate Number are one and the same.
• Even in a mechanism as finely built as the LEI, capturing change rests at base on declaration by the entity itself, with outside information complementing it. An identifier is something maintained.
• What schema.org's sameAs expects is a URL, not an identifier string.
• The trade name written into structured data has to match what is displayed on screen.
• And there is no way to know whether you are being identified correctly other than by measuring.

That last point is the thrust of this article. Because the identifiers are in place, because you disclosed, because you fixed the site — each of these is a hypothesis. Ask under the current trade name, the former one, the abbreviation, the English trade name and the issue code, and check repeatedly, under conditions that carry no context over, whether the answers point to the same entity. Whether confusion is occurring can be known no way other than as the result of that.

Grasping perception of your company on AI continuously, as content rather than as volume of exposure. “Whose figures are those?” is the question that sits at the start of it.

The sources supporting this article's claims are set out in three tiers according to their character. All were confirmed at the primary level on August 31, 2026.

Tier 1: Primary and official sources (numbering bodies, the exchange, specification authors, search providers)

#	Source	Use in this article
1	Japan Exchange Group, “Corporate Information Required for Timely Disclosure” (as of July 10, 2026) https://www.jpx.co.jp/equities/listing/disclosure/info/	That the Decisions by Listed Companies set out a change in a trade name or a corporate name (26), corporate restructuring such as a merger (10), a takeover bid or takeover bid for own shares (11), a transfer or acquisition of all or part of a business (13) and matters accompanied by a change in a subsidiary (17) as separate items (§3)
2	Japan Exchange Group, “Investor Relations” https://www.jpx.co.jp/equities/listing/investor-relations/index.html	That developing an IR system is an obligation under the Code of Corporate Conduct and that the specific content of the system is left to each company's judgment (§3-2). Exchange rules, not national law
3	Japan Exchange Group, “TDnet” https://www.jpx.co.jp/english/equities/listing/disclosure/tdnet/	The retrieval conditions of 31 days for public browsing and 10 years for search by company (§3-3)
4	Securities Identification Code Committee, “Securities Codes will Include Letters” https://www.jpx.co.jp/sicc/code-pr/index.html	Specific name codes for stocks are four numeric digits (1300–9999). From codes set on or after January 1, 2024, 19 uppercase letters (excluding B, E, I, O, Q, V and Z) are used in the second and fourth digits. Existing codes are not changed (§5-1, §5-2)
5	Securities Identification Code Committee, “FAQ” https://www.jpx.co.jp/sicc/securities-code/02.html	An ISIN is 12 characters = 2-character country code + 9-character basic code + 1 check digit (ISO 6166). On a merger the securities code and sector of the surviving listed company are used, and where a newly established parent company is newly listed through a share transfer or the like a new one is assigned. Codes that fell out of use from July 1993 onward are not reused (§2-2, §5-1, §5-2, §5-3)
6	National Tax Agency, “About Corporate Numbers” https://www.houjin-bangou.nta.go.jp/setsumei/index.html	The Corporate Number is 13 digits. The Basic 3 information published (trade name or name / address of head office or principal place of business / Corporate Number), and change history information for the name and address (§5-4)
7	National Tax Agency, “Procedure where the name or head office address has changed” https://www.houjin-bangou.nta.go.jp/shitsumon/shosai.html?selQaId=00045	A Corporate Number is not changed even where the name or address changes. For corporations registered for incorporation, registration at the Legal Affairs Bureau updates the publication site. Where an English form has been registered, a further registration procedure is required (§5-4, FAQ Q1)
8	National Tax Agency, “Does a Corporate Number prove existence?” https://www.houjin-bangou.nta.go.jp/shitsumon/shosai.html?selQaId=00055	Core to this article. A Corporate Number does not prove the existence of a corporation or organization, nor does it prove that bodies using the same Corporate Number are one and the same. It is designated even for a dissolved corporation as long as the registration record has not been closed (§5-4)
9	GLEIF, “The Legal Entity Identifier (LEI)” https://www.gleif.org/en/organizational-identity/lei-vlei/the-legal-entity-identifier-lei	The LEI is 20 characters. Based on ISO 17442, managed by GLEIF with the G20 and the FSB behind it and supervised by the ROC. Level 1 “who is who” / Level 2 “who owns whom” (§5-5)
10	GLEIF, “ISO 17442: The LEI Code Structure” https://www.gleif.org/en/about-lei/iso-17442-the-lei-code-structure	Characters 1 to 4 are the issuer prefix, 5 to 18 the entity-specific part (designed to carry no meaning), 19 to 20 the check digits (§5-1, §5-5)
11	GLEIF, “Level 1 Data: LEI-CDF Format 3.1” https://www.gleif.org/en/lei-data/access-and-use-lei-data/level-1-data-lei-cdf-3-1-format	The data items of Legal Entity Events (type of event, legal effective date, recorded date, type of validation documents) and the succeeding entity (SuccessorEntity) (§5-5)
12	LEI Regulatory Oversight Committee (October 30, 2018) https://www.leiroc.org/publications/gls/roc_20181030-1.pdf	Core to this article. Without a declaration by the affected entity itself, an LEI issuer cannot necessarily know that an event requiring an update of reference data has occurred (§5-5)
13	schema.org, “sameAs” https://schema.org/sameAs	What sameAs expects as its value is a URL, pointing to a reference web page that unambiguously indicates the item's identity. It applies to Thing (§6-1, FAQ Q5)
14	schema.org, “leiCode” https://schema.org/leiCode	The property for stating the legal entity identifier defined in ISO 17442 (§6-2)
15	Google Search Central, “General structured data guidelines” https://developers.google.com/search/docs/appearance/structured-data/sd-policies	Do not mark up content that is not visible to readers of the page. Structured data has to be an accurate representation of the content of the page (§7-1, premise of FAQ Q6)
16	Google Search Central, “How structured data works” https://developers.google.com/search/docs/appearance/structured-data/intro-structured-data	Do not add structured data about information that is not visible to the user, even where the information is accurate (§7-1)
17	Google Search Central, “AI features and your website” https://developers.google.com/search/docs/appearance/ai-features	There are no additional technical requirements for reference links in AI features, and no special schema.org markup exists for generative AI features. Machine-readable files such as llms.txt are not used by Google Search (§7-2)

Tier 2: Peer-reviewed research

#	Source	Use in this article and reservations
18	Shah, Ye, Jaskowski, Xu, Chava (Georgia Institute of Technology), “Beyond the Reported Cutoff: Where Large Language Models Fall Short on Financial Knowledge” https://arxiv.org/abs/2504.00042	17,621 listed companies, 197,011 questions. Answers differing from the correct value by more than 10% observed even with a fixed prompt stating the company name and the fiscal year (§1-3, FAQ Q8). A peer-reviewed paper accepted at COLM 2025. This article refers to the arXiv version. Reservations are placed on the experimental conditions rather than on peer-review status: the scope is US listed companies; revenue is Compustat-derived, converted into millions of US dollars with accounting-standard differences unharmonized; API runs were in February 2025
19	Nakagawa, Hirano, Fujimoto, “Evaluating Company-specific Biases in Financial Sentiment Analysis using Large Language Models” https://arxiv.org/html/2411.00420v1	Japanese earnings reports on TDnet (2019–2023). Sentiment scores changed depending on whether the company name was stated or removed, and the direction of the bias was inconsistent across models. 10,249 observations with GPT-4o (§1-3, FAQ Q8). An IEEE version exists. This article refers to the arXiv version

Tier 3: Preprints and material from interested parties

#	Source	Use in this article and reservations
20	Benchmark on high-confidence wrong answers in financial QA https://arxiv.org/abs/2607.11414	Even in the “high confidence” state where eight re-answers to the same question all agree, 15–23% were wrong (§4-2, FAQ Q9). A preprint released in July 2026. An evaluation of foundation models, not of public AI search services
21	GLEIF Blog, “Metric of the Month – Legal Entity Events” https://www.gleif.org/en/newsroom/blog/transforming-data-into-opportunities-metric-of-the-month-legal-entity-events	Legal Entity Events currently have 20 types, including changes in ownership, mergers and acquisitions, changes of name or address, and the ending of an entity (§5-5). GLEIF is the body that manages the LEI and is also in a position of promoting its adoption. Treated as material from an interested party

This article's policy on sources

• The character counts, structures and specification dates of identifiers were confirmed against the primary material of the numbering bodies or the specification authors. Figures derived from research are sourced to the paper in question. No statement rests on a secondary commentary article.
• Published statistics showing how often companies are confused in AI answers could not be confirmed, so no frequency figures are given anywhere.
• A demonstration that stating identifiers improves entity identification in AI answers could not be confirmed, so no such claim is made.
• The point at which effect arises under the Companies Act and the Financial Instruments and Exchange Act has not been confirmed at the primary level, so this article does not go into it.