AIO & LLMO for IR | Which Version of Your Disclosure Is AI Describing? — Forecasts, Actuals, and Corrective Disclosure

When companies verify what generative AI says about them, most designs look at whether the figure is right. Fiscal period, accounting standard, currency, unit, scope of consolidation. But the disclosure of a listed company carries a second axis alongside the value. The same "100 billion yen in revenue" is an entirely different piece of information for an investment decision depending on whether it is an initial forecast, a revised forecast, an actual, or a figure that has passed through corrective disclosure.

The Tokyo Stock Exchange lists amendments to performance estimates and differences in estimates and earnings values, and dividend estimate or amendment to dividend estimate, as categories of timely disclosure that stand on their own (TSE, "Corporate Information required for Timely Disclosure", as of July 10, 2026). On the rules side, what is distinguished is not the value itself but which status that value carries. Even so, when it comes to verifying AI answers, a framework for managing which status and which version a figure belongs to is something that, within the scope of this review, we could not confirm.

This article makes one claim. Give the reference table a column for status and a column for version, not only a column for value. Disclosure gets overwritten. Forecasts are revised, replaced by actuals, and sometimes rewritten retroactively by corrective disclosure. The premise itself — that a published primary source is the correct answer — decays with time.

• Why the wrong status is a serious error even when the figure matches, and what in the rules supports that
• An organization of disclosure status into four buckets (initial forecast / revised forecast / actual / restated)
• What to do when corrective disclosure rewrites the expected value in the reference table itself
• How to decide precedence when the annual securities report, the earnings report, timely disclosure, earnings presentation materials, and the body text of the IR site disagree
• The five columns to add to the reference table (status / supersedes / superseded_by / effective_from / precedence) and an implementation example in CSV
• How to make disclosure events the trigger for re-measurement, and how to set the checking timings
• What the IR site can and cannot do to show which version is current

IR practitioners at listed companies who have already begun verifying AI answers, or are about to begin. Secondarily, legal and accounting staff involved in decisions about corrective disclosure.

This article is a sequel to How to Verify What AI Says About Your Financials. The basic design of the reference table, why repeated and stateless measurement is necessary, what to record, how to define the metrics, and the shape of the report to management are all in that article. This article takes all of that as given and inherits it, and adds only the axes of status and version. For the problem framing itself — how AI describes your financial information — the four types of error, and the boundary set by fair disclosure, see the parent article How AI Describes Your Financials.

Disclaimer: This article provides an organization of publicly available information as a reference for IR practice at listed companies. It is not intended for use in legal or accounting judgments, in judgments about whether disclosure is required, or in individual investment decisions. On the application of the rules, please confirm with the relevant stock exchange, the supervisory authorities, and your own legal counsel and accounting auditor.

1-1. Existing verification is built on the axis of value

When you implement verification of AI answers, the reference table starts out holding values. Revenue, operating profit, dividend per share, headcount. Then you notice that values alone cannot settle the judgment, and columns get added for fiscal period, accounting standard, currency, unit, and scope of consolidation. Up to here, the design exists to prevent accidents in which the same number carries a different meaning.

This axis is necessary, but it is not sufficient. There is a path that still ends in a wrong answer even when fiscal period, accounting standard, currency, and unit all match.

1-2. The same number becomes a different piece of information

Make it concrete. Suppose that for one company's consolidated revenue for the fiscal year ending March 2026, the following four figures all appear in materials the company itself published.

Now suppose AI answers that "consolidated revenue for the fiscal year ending March 2026 was 100 billion yen." Unless it is settled which of the four that 100 billion yen refers to, no judgment of right or wrong can be made. If it is answering with the initial forecast, the information reaching investors is an outlook more than a year old. If it is answering with the revised forecast, a projection is still circulating after the actual has been published. If it is answering with the figure from before the corrective disclosure, a number the company itself republished as erroneous is still being spoken.

In every one of these cases, the number itself is a value the company once published, not a fabrication. Ask for a source and a real company document may well come back. Verification of the traditional kind — does the value match, does the source exist — passes straight over this path.

1-3. Two things to add: status and version

This article uses two terms distinctly.

• Status: the bucket a figure falls into — whether it is a forecast, an actual, or something that has passed through corrective disclosure
• Version: within the same status, which publication in sequence it is. A revised forecast can be issued more than once, and corrective disclosure is not necessarily a one-time event

Status and version are independent. "The second revised forecast" and "the actual after corrective disclosure" both sit in the same column if you look only at the value. They can be told apart only when you hold the document and the date on which it was published.

It is worth noting that the output on the AI side can move with the input conditions, and this has been observed with Japanese data as well. One study reports that for earnings reports disclosed through TDnet from 2019 to 2023, sentiment assessments change depending on whether the company name is stated or removed (Nakagawa, Hirano, Fujimoto, Evaluating Company-specific Biases. An IEEE version exists. 10,249 observations with GPT-4o. The direction of the bias is not consistent across models). That study did not measure mistaken disclosure versions, but it is one example showing that model output can change with the input conditions. In the verification described here, to keep the target of judgment unambiguous, which version is meant is pinned down on the question side.

2-1. The rules already distinguish status

The Tokyo Stock Exchange sets out the corporate information required for timely disclosure by category. In the primary source as of July 10, 2026, for information on listed companies, 39 items of decisions and 29 items of facts which occurred are enumerated, and separately from these there are three kinds of earnings information (earnings report, earnings report for Q2 (interim), and quarterly earnings reports for Q1 and Q3), two kinds of amendments to performance estimates or dividend estimates, and eight items of other information (TSE, "Corporate Information required for Timely Disclosure").

What is decisive here is that amendments to performance estimates, differences in estimates and earnings values, and dividend estimate or amendment to dividend estimate, are placed as categories set apart from earnings information. Issuing a forecast, amending a forecast, and disclosing an actual are each treated as a separate act of disclosure under the rules.

This is what supports, on the rules side, the claim that a wrong status is a serious error even when the figure matches. On the rules side too, a change of status is treated as an object of disclosure in its own right.

It is worth adding that the Exchange publishes the perspectives it applies in examining disclosed information to secure the appropriateness of disclosure. There are five: whether the timing of disclosure is appropriate; whether the content is untrue; whether information important to investment decisions is missing; whether the disclosure would give rise to a misunderstanding relevant to investment decisions; and whether it is otherwise lacking in appropriateness (TSE, "Examinations pertaining to Disclosure of Corporate Information". Securities Listing Regulations, Rule 415, Paragraph 1 and Rule 3, Paragraph 2). These are perspectives on the disclosure a company makes, not perspectives on AI answers.

2-2. The four statuses

This article divides the status of a disclosed figure into the following four, to the extent needed for verifying AI answers.

These four buckets are this article group's own organization, not a classification established in the industry. Nor do they map one-to-one onto the disclosure categories in the rules. The purpose is to bring the distinction down to a granularity that can be judged mechanically as a single column in the reference table. If they do not fit your own disclosure practice, add buckets or remove them. What matters is not the names of the buckets but holding a column for status separately from the column for value.

2-3. Changes other than value that have the same structure

The problem of status and version is not confined to forecasts and actuals. The following changes have the same structure: the previously published value was not wrong, and yet the correct answer now is a different value.

• Before and after adjustment for a stock split: dividend per share, earnings per share, and share price levels all change across a split. Stock split is set out as a decision by a listed company
• Change of accounting period: a change in the end date of the business year is also set out as a decision. Across such a change, the very period that the phrase "fiscal 2026" points to changes
• Change in the definition of an internally defined KPI: membership numbers, utilization rates, order backlog. Change the definition and the same-named metric becomes a different value. Because it does not correspond to a disclosure category in the rules, it can be tracked only internally
• Change in the disclosure categories themselves: the primary source notes that adoption of a resolution to adjust debt obligations concerning financial claims on the basis of the Early Business Recovery Act will be added to facts which occurred from December 11, 2026, and that disclosure of minority shareholder approval rates and the like will apply from the date of the annual general shareholders meeting relating to business years ending on or after December 1, 2026. The framework of disclosure has versions too

The last point bears directly on how the reference table is operated. If the reference table holds which category in the rules a question is tied to, a revision on the rules side makes that tie stale. The primary source for the rules needs a confirmation date of its own.

3-1. Fixing the terms here

This article draws a strict line between two similar acts. From here on, neither is written as a bare term.

Under the rules these two are entirely different things. The former is a procedure inside the disclosure regime; the latter is a claim made to a private service. How far it is actually possible to ask an AI provider to correct or remove an AI answer is covered in How Far AI Answers Can Be Corrected. This article deals only with how the company manages the versions of its own disclosure.

3-2. The premise that a published primary source is the correct answer breaks down

When verification of AI answers is designed, one premise is placed there tacitly: that the primary sources the company published are correct. The expected_value in the reference table is transcribed from the earnings report or the annual securities report. The source field takes the URL and the page of that document.

Corrective disclosure breaks this premise. If the figure held in the reference table is among those covered by an amended report or a corrective disclosure, the expected value in the reference table itself has to be revisited.

Three things happen at that point.

1. The expected_value in the reference table has to be rewritten. If it is not, an AI answer giving the correct post-corrective-disclosure figure will be judged a mismatch
2. The materials from before the corrective disclosure can still exist on the web. The company's own site, news coverage, data vendors, knowledge already learned by AI, search indexes. Where they remain cannot be grasped completely from the company's side
3. The reading of past measurement records changes. A result recorded as a match before the corrective disclosure is, seen now, a match against the pre-corrective-disclosure value

The third is easy to miss. Measurement records must not be overwritten. Records from before the corrective disclosure stay as judgments that were correct at the time. That past judgments change because of a corrective disclosure is itself information about disclosure quality.

3-3. Two columns to hold on the reference table side

Rather than rewriting the row in the reference table, add a new row and make its relationship to the old row explicit. Two columns serve this purpose.

• supersedes: the identifier of the earlier row that this row replaced
• superseded_by: the identifier of the later row that replaced this one. Empty for the current version

Several rows exist for the same question, and the row whose superseded_by is empty is the current version. Judgment is made against that row. Do not delete the old rows. Delete them and it becomes impossible to work out afterward which version AI was answering with.

This design means making the reference table append-only rather than update-in-place. The number of rows grows, but the way it grows becomes something to manage. If versions have piled up many times over on a single question, that topic may be one that invites misunderstanding outside the company.

3-4. The line not to cross

Talk of corrective disclosure invites a certain thought: "AI is speaking an old figure, so put out the right figure and fix it."

Correcting an AI answer is limited to re-presenting published materials. You must not correct an AI answer with material non-public information. This is a premise that does not move across this article group. The fact that an AI answer is wrong is not a reason to release non-public information. If it is necessary, carry out a lawful publication procedure first. On what changes before and after publication, the parent article How AI Describes Your Financials covers the fair disclosure boundary.

For the same reason, do not put non-public figures into the reference table itself. The reference table is an internal document, but it passes through outside contractors, tools, and cloud services. Every row in the reference table must correspond to a published source document and a publication date. A row with no corresponding published material does not belong in the reference table.

4-1. There is more than one primary source

Information about a listed company is spread across several official documents. And each differs in the point in time it describes, the purpose of the description, the definition of the figures, and the scope covered.

As a point of reference, the corporate governance report is also the place where the state of IR system development is disclosed. The Exchange requires listed companies to develop an IR system under the Code of Corporate Conduct, and the state of that development is required to be disclosed in that report (TSE, "IR Systems and IR Activities"). Care is needed here. What is required is the development of an IR system, not the measurement and management of how the company is perceived in AI space. What this article can address goes one step and no further: how far, within the existing framework of the IR system, inconsistencies in disclosure as reflected in AI can be treated as something to manage.

It should also be noted that the July 2026 edition of the Corporate Governance Code has been published, and the partial amendment to the Securities Listing Regulations relating to the revision took effect on July 21, 2026 (TSE, "コーポレートガバナンス・コード（2026年7月版）の公表について (Japanese only; unofficial translation: "Publication of the Corporate Governance Code (July 2026 Edition)")"). The scope of application by market segment has not been checked in this article, so no segment-by-segment description is offered. The Code is a comply or explain framework, not a law.

4-2. Decide the precedence rule before a conflict happens

When primary sources conflict, debating on the spot which one is right makes the basis of judgment wobble from round to round. Once it wobbles, comparison over time stops working. Last month's mismatch and this month's mismatch were not judged on the same basis.

So decide precedence in advance and write it into the reference table. That is the role of the precedence column. What follows is a practical proposal, not an optimum derived from research. Rearrange it to fit your own disclosure practice.

However, this ranking has to coexist with "the newer one wins." Even though the annual securities report ranks above the earnings report, if the earnings report has since been the subject of a corrective disclosure, the corrective disclosure is the current version. So the judgment runs in two stages.

1. First keep only the rows whose superseded_by is empty (the version judgment)
2. Among what remains, take the smallest precedence (the document judgment)

4-3. Do not ask about forecasts and actuals in the same question

Deciding precedence means nothing if the wording of the question does not specify the status. "What was revenue for the fiscal year ending March 2026?" holds together whether the answer is the initial forecast, the revised forecast, or the actual.

So write the status into the question in the reference table.

• "What is the actual consolidated revenue for the fiscal year ending March 2026?"
• "What is the latest company forecast for consolidated revenue for the fiscal year ending March 2027?"

This is not how people naturally ask. That is exactly why you also place, as a separate question, what comes back when an investor asks in the natural way. Holding a status-specified question and an unspecified question as a pair lets you separate "AI is supplying the status" from "AI is getting the status wrong."

Within the scope of this review, we could not confirm public data measuring how often AI mistakes the version of a disclosure. This article therefore gives no figure for how often it happens. It shows the path by which it can happen structurally, and a way to check for it in your own company, and stops there.

The status axis has one more short topic. A matter to be put to a general shareholders meeting is not "resolved at the shareholders meeting" at the point it appears in the notice of convocation.

Dividend from surplus, amendment to the articles of incorporation, and acquisition of all classified stocks subject to whole acquisition clause are all set out as decisions by a listed company (TSE, "Corporate Information required for Timely Disclosure"). Once a matter appears as a proposal in the notice of convocation, its content is published information. But it is the information that the company is proposing this, not the information that it has been resolved at the shareholders meeting.

In verifying AI answers, this difference tells in practice. Ask "what is the annual dividend?" during the period before the meeting and the amount in the proposal may come back. The value is correct. The source exists. But the status is "proposed," not "resolved."

In the reference table, handle it as follows.

• Hold matters to be put to the meeting under a separate status (for example, proposed)
• Record in advance that the expected value for the same question can change across the meeting (use effective_from)
• Once the outcome of the resolution is disclosed after the meeting, add a new row and fill in the superseded_by of the earlier row

This article does not enter the legal question, under the Companies Act, of when a resolution takes effect. That is a matter to be left to the legal department and outside counsel for case-by-case judgment. What this article addresses goes one step and no further: holding "the status of a proposal" and "the status after a resolution" separately as a matter of reference table operation.

This is the core of the implementation. Add five columns to the existing reference table.

6-1. The five columns

6-2. How they relate to the existing columns

The reference table in the previous article has announced_date (the date the value was published), last_verified (the date the reference table itself was last verified), and source_document (the name of the source document). The new columns do not overlap in role with these.

A word on why announced_date and effective_from are separate. Corrective disclosure replaces figures for a past period with a later date attached. The amended report may be filed in August 2026 while the period covered by the replaced figures is the fiscal year ending March 2025. The fiscal year covered is carried by fiscal_period, and effective_from carries the date from which this row is treated as the current version for verification. The publication date alone cannot express these two separately.

6-3. An implementation example in CSV

As in the previous article, the reference table can be a spreadsheet or a CSV. Here the column definitions are shown as a CSV header. Do not make it an image. The point is to allow mechanical matching in later steps.

row_id,question_id,question_en,expected_value,status,fiscal_period,period_type,consolidation,accounting_standard,currency,unit,continuing_ops,announced_date,effective_from,supersedes,superseded_by,precedence,source_document,source_url,source_page,tolerance,last_verified
R001,Q010,What is the actual consolidated revenue for the fiscal year ending March 2026,1234567,actual,2026-03,FY,consolidated,IFRS,JPY,million,continuing,2026-05-12,2026-05-12,,R004,3,kessan_tanshin,https://example.co.jp/ir/tanshin_2026q4.pdf,p.1,0,2026-08-16
R002,Q011,What is the latest company forecast for consolidated revenue for the fiscal year ending March 2027,1300000,initial_forecast,2027-03,FY,consolidated,IFRS,JPY,million,continuing,2026-05-12,2026-05-12,,R003,3,kessan_tanshin,https://example.co.jp/ir/tanshin_2026q4.pdf,p.1,0,2026-08-16
R003,Q011,What is the latest company forecast for consolidated revenue for the fiscal year ending March 2027,1250000,revised_forecast,2027-03,FY,consolidated,IFRS,JPY,million,continuing,2026-08-05,2026-08-05,R002,,4,gyoseki_yoso_shusei,https://example.co.jp/ir/rev_20260805.pdf,p.1,0,2026-08-16
R004,Q010,What is the actual consolidated revenue for the fiscal year ending March 2026,1229000,restated,2026-03,FY,consolidated,IFRS,JPY,million,continuing,2026-08-20,2026-08-20,R001,,1,teisei_hokokusho,https://example.co.jp/ir/teisei_20260820.pdf,p.3,0,2026-08-21

Read out what these four rows are saying.

• Q010 (actual) has two rows. R001 holds R004 in superseded_by, so it is not the current version
• R004 derives from the amended report and has precedence 1, the highest. This one is the current version
• Q011 (forecast) also has two rows, and R003 (the revised forecast) is the current version
• Because R001 has not been deleted, when AI answers 1234567 it can be judged as answering with the pre-corrective-disclosure value

That last point is the practical payoff of this design. Because the old row is still there, the type of error can be identified. Had it been deleted, the record would say no more than "mismatch."

6-4. Finer-grained judgments

Holding a status column raises the granularity of the judgment. The verification in the previous article records match or mismatch; adding the status axis splits mismatch further.

• Value and status both match: no problem
• Value matches, status differs: the version has been mistaken. AI is answering with a real published value of the company, but not the current version
• Value does not match: the traditional kind of error. The number itself does not match a value the company published
• Value does not match and matches no past version either: the origin may not be a published document of the company at all

The second and the fourth call for entirely different responses. A mistaken version can be treated as a problem in how the company presents information. If the causes sit on the company's side — the current version is hard to find, prior-year pages do not state the year, old pages remain reachable from search — then the company can act on them. The fourth suggests that the company's materials may not be being referred to at all.

7-1. Combining periodic measurement with event-driven measurement

The previous article dealt with the design of repeated, stateless measurement. What this article adds is the idea of tying "when to re-measure" to disclosure events.

The reason is simple: the company knows the moment the status changes. The day an amendment to the performance estimate was disclosed, the day the earnings report went out, the day the annual securities report was filed, the day a corrective disclosure was made. These are on the company's own calendar. With periodic measurement alone, the interval between a change of status and the next measurement is a blank.

7-2. How to set the checking timings

What follows is one example of frequencies. Set your own according to the size of the risk, the type of disclosure event, and the running cost. These are not optimal values derived from research.

The choice of "two weeks later" has no fixed basis. Two points in time are placed simply to see whether the result differs right after disclosure and after some time has passed. Run it once at your own company, and if no change shows up, the interval can be widened.

7-3. Measurement conditions carry over from the previous article

A single answer is one observation. From an answer obtained by putting the same question once, you cannot infer the stable tendency of what that AI returns about your company. The reason for this, the need for repeated and stateless measurement, and the design of what to record are in How to Verify What AI Says About Your Financials. This article takes that as given.

It is worth holding onto the point that stability of measurement is not evidence of correctness. On a benchmark in the financial domain, even in cases where eight re-answers to the same question all agreed, 15 to 23% were reported to be wrong (a study of high-confidence errors using FinQA, a July 2026 preprint. It has not been peer reviewed. It is a result on a benchmark of foundation models, not of publicly available AI search products). That an answer does not waver does not mean the answer is right.

Vaipm measures AI-space perception through a total of 25 stateless queries across multiple AI engines. This is not an optimum derived from research; it is Vaipm's operational design.

8-1. External publication venues have time limits

How long the company's disclosure materials stay in external publication venues is set by the rules. Materials published through TDnet are viewable on the Company Announcements Disclosure Service for 31 days including the date of disclosure (including Saturdays, Sundays, and national holidays), and the past ten years are viewable through Listed Company Search. In addition, the past five years of disclosed data are stated to be viewable through a paid database service (TSE, "Overview of TDnet").

One practical consequence follows from this. You need a place of your own where materials can be referred to continuously. A reference table holding a Company Announcements Disclosure Service URL as its source loses the trail after 31 days. It is easier to work with a reference table whose source_url puts a permanent URL on the company's own IR site first, and holds the URL on the rules side as secondary.

8-2. What you can do on your own site

From the standpoint of version management, this is the range in which the company's own site can act.

• State the year on prior-year pages: write which business year the information belongs to in the page title, the heading, and the opening of the body. Do not rely on the file name or the directory level alone
• Put a route to the current version on the prior-year page itself: let a reader who lands on an old page move to the latest one
• For items subject to a corrective disclosure, post the post-corrective-disclosure material alongside: the fact of the corrective disclosure is itself information
• Provide the key figures as HTML text as well: avoid providing them only as PDF
• Reduce duplication: tidy up states in which the same content is reachable from several URLs. Google explains that having the same content reachable from many URLs is confusing for users and makes it harder to track how the content performs (Google Search Central, "What is canonicalization", updated August 20, 2026)

On provision in HTML, the evidence is limited. A practitioner report covering 20 listed companies in Europe states an answer accuracy of 71% for companies with an annual report in HTML format and 54% for companies centered on PDF (GenAI as a Reader. Subsample for accuracy n=200). However, this is a joint study that includes a digital IR reporting vendor, so it has an interested party. Its peer review status has not been confirmed either. Even in the HTML group 29% were judged inaccurate, so it cannot be said that "moving to HTML makes it correct." Differences in company size, IR quality, and update practices may be mixed in.

8-3. What you cannot do

Which URL is treated as the representative one is not something the company can decide. On specifying the canonical URL, Google states plainly that you can indicate a preference, but that it is a hint and not a rule. A page other than the one specified may be chosen as canonical (the same "What is canonicalization" cited above).

On the conditions for appearing in generative AI features, Google cites that the page must be indexed and eligible to appear in search results with a snippet, and additionally that inclusion in generative AI features must be enabled in Search Console. On top of that, it states plainly that even meeting all the requirements, best practices, and policies does not guarantee crawling, indexing, or appearance (Google Search Central, "Optimizing your website for generative AI features on Google Search", updated July 10, 2026).

In the same document, Google states that machine-readable files such as llms.txt and special markup are not required in order to appear in Google Search (including generative AI features), and that Google Search does not use them. On structured data too, it states that this is not required for generative AI search. This article therefore does not recommend these as measures that work on Google. Structured data can be eligible for rich results, so it is worth continuing as part of ordinary SEO, but positioning it as an extra measure aimed at AI does not square with what Google says.

It should be noted that Google describes the mechanism of generative AI features as retrieval-augmented generation (RAG), which retrieves relevant pages from the search index, and query fan-out, in which the model issues several related queries at once. "No special technical requirements or markup are needed for AI" and "no dedicated processing is involved" are two different statements. As noted above, Google cites, in addition to the ordinary search requirements, that inclusion in generative AI features must be enabled in Search Console as a condition of eligibility to appear. What is said to be unnecessary is additional technical measures aimed at AI, such as llms.txt; it is not shown that ordinary search results and AI answers go through the same processing.

When a change on the company's site is reflected in each AI's answers is not controllable from the company's side. The time to reflection, and how long information that has once appeared continues to be referred to, are covered in How Long Information Stays in AI Answers. This article's interest runs the other way: not the behavior of the AI side, but which version the company itself presents as current.

Q1. If the figure matches, is there really a need to look at the status as well?

Yes. Even when the figure matches, a different status makes it a different piece of information for an investment decision. If AI answers with the revenue in the initial forecast, for instance, that figure is one the company actually published, so verification on value alone judges it a match. But if that figure is circulating after the actual has come out, the information reaching investors is an old outlook. Verification of value does not substitute for verification of status. Only by holding both can you separate the types of error.

Q2. Are the four buckets — initial forecast, revised forecast, actual, restated — a classification settled in the industry?

No. These four buckets are this article group's own organization, not a classification established in the industry. Nor do they map one-to-one onto the disclosure categories in the rules. The purpose is to bring the distinction down to a granularity that can be judged mechanically as a single column in the reference table. If they do not fit your own disclosure practice, add buckets or remove them. What matters is not the names of the buckets but holding a column for status separately from the column for value.

Q3. How often does AI mistake the version of a disclosure?

Within the scope of this review, we could not confirm public data measuring that rate. This article therefore does not indicate how often it occurs. What can be shown is the path by which it can happen structurally, and a way to check for it in your own company. If you want to know the frequency, the practical route is to build your own reference table and produce your own figure from measurement results over a set period. Borrowing another company's figure does not carry over, because the way disclosure is issued and the structure of the materials differ.

Q4. Can old rows in the reference table be deleted?

We recommend not deleting them. It is because the old rows are still there that you can tell whether a figure AI answered with is a pre-corrective-disclosure value or a value of entirely unknown origin. Delete them and both are recorded as no more than a mismatch, leaving no way to decide which direction to respond in. Making the reference table append-only rather than update-in-place, and treating rows whose superseded_by is empty as the current version, is the easier design to work with.

Q5. After making a corrective disclosure, how should verification of AI answers change?

Three pieces of work are needed. First, add a new row to the reference table and fill in the superseded_by of the earlier row. Second, leave past measurement records as they are without overwriting them. A judgment recorded as a match before the corrective disclosure was correct at that time. Third, for the questions covered by the corrective disclosure, re-measure at two points in time: after the disclosure, and again after an interval. The point is to see whether the pre-corrective-disclosure value keeps coming back.

Q6. When AI is speaking an old figure, can we set it right with non-public information?

No. Correcting an AI answer is limited to re-presenting published materials. You must not correct an AI answer with material non-public information. The fact that an AI answer is wrong is not a reason to release non-public information. If it is necessary, carry out a lawful publication procedure first. For the same reason, do not put non-public figures into the reference table itself. The reference table is an internal document, but it passes through outside contractors and tools.

Q7. When the annual securities report and the earnings report differ, which should be the correct answer?

Rather than deciding after a conflict arises, we recommend deciding precedence in advance and writing it into the reference table. If the judgment wobbles from round to round, comparison over time stops working. As a practical proposal this article gives the order: amended report and corrective disclosure, annual securities report, earnings report, timely disclosure, earnings presentation materials. It is not an optimum derived from research. Alongside that, make the version judgment (whether it is the current version) before the document judgment.

Q8. Can a proposal to be put to the general shareholders meeting go into the reference table?

It can, but separate the status. Once it appears in the notice of convocation the information is published, but it is the information that the company is proposing this, not the information that it has been resolved at the shareholders meeting. Hold status as proposed, and once the outcome of the resolution is disclosed after the meeting, add a new row. Note that the question under the Companies Act of when a resolution takes effect is not covered in this article. Please confirm case-by-case judgments with your legal department.

Q9. Is the Tokyo Stock Exchange asking companies to manage how they are perceived in AI space?

No. What the Exchange requires under the Code of Corporate Conduct is the development of an IR system, and the disclosure of the state of that development in the corporate governance report. Measuring and managing how the company is perceived in AI space is not an obligation. What this article can address goes one step and no further: how far, within the existing framework of the IR system, inconsistencies in disclosure as reflected in AI can be treated as something to manage. Please do not mix obligations under the rules with voluntary initiatives.

Q10. If we fix the IR site, when will AI's answers change?

That is not controllable from the company's side. When a change is reflected, and how long information that has once appeared continues to be referred to, are covered in How Long Information Stays in AI Answers. This article's interest runs the other way: not the behavior of the AI side, but which version the company itself presents as current. Before waiting for reflection, check first whether the current version is easy to find on your own site.

Q11. If we put up an llms.txt, can we tell AI which version is right?

As far as Google is concerned, do not expect that. Google states that machine-readable files such as llms.txt and special markup are not required in order to appear in Google Search (including generative AI features), and that Google Search does not use them. On structured data too, it states that this is not required for generative AI search. How other services treat the same file is something to check in each provider's own explanation. This article does not recommend this as a measure that works on Google.

Q12. Where should we start?

Adding one status column to the existing reference table is enough to start with. Reclassify the existing questions by which of initial forecast, revised forecast, actual, or restated they are asking about. That work alone shows how many questions do not specify the status. Next, pick one recent disclosure event and re-measure before and after it. One round of measurement gives a sense of whether version management is a live issue at your company.

The claim of this article comes down to one thing. In verifying AI answers, looking only at whether the value matches is not enough. Hold which status and which version the figure belongs to as columns in the reference table.

To put the points in order.

1. The Tokyo Stock Exchange places amendments to performance estimates and differences in estimates and earnings values, and dividend estimate or amendment to dividend estimate, as separate categories of timely disclosure. On the rules side too, a change of status is treated as an object of disclosure in its own right
2. Disclosure gets overwritten. Forecasts are revised, replaced by actuals, and rewritten retroactively by corrective disclosure. The premise itself — that a published primary source is the correct answer — breaks down
3. The four buckets of initial forecast, revised forecast, actual, and restated are this article group's own organization, not a classification established in the industry
4. Add status / supersedes / superseded_by / effective_from / precedence to the reference table. Do not delete old rows; make it append-only
5. Decide precedence among conflicting primary sources before a conflict arises. Make the version judgment before the document judgment
6. Make disclosure events the trigger for re-measurement. The company knows the moment the status changes
7. What the company's own site can do stops at making the current version easy to find. Neither which URL is treated as representative nor when it is reflected on the AI side is controllable from the company
8. Correcting an AI answer is limited to re-presenting published materials. You must not correct an AI answer with material non-public information

There are three actions you can take next.

• Today: open the existing reference table and classify each row by which of initial forecast, revised forecast, actual, or restated it is asking about. Rows you cannot classify are rows whose question does not specify the status
• At the next disclosure event: an amendment to the performance estimate or an earnings announcement will do. On the day of disclosure and two weeks later, re-measure the questions concerned and see whether the earlier version is coming back
• Every quarter: extract only the rows whose superseded_by is empty and check against the primary sources that they really are the current version

The overall design of measurement — the basics of the reference table, repeated and stateless measurement, what to record, metrics and reporting to management — is in How to Verify What AI Says About Your Financials, and the problem framing of how AI describes your financial information, together with the four types of error, is in How AI Describes Your Financials. The idea of managing how a company is perceived in AI space on an ongoing basis is organized in What AI Perception Management Is.

Disclaimer: This article provides an organization of publicly available information as a reference for IR practice at listed companies. It is not intended for use in legal or accounting judgments, in judgments about whether disclosure is required, or in individual investment decisions. The column design, the precedence order, and the checking timings shown in this article are all practical proposals, not optima derived from research. On the application of the rules, please confirm with the relevant stock exchange, the supervisory authorities, and your own legal counsel and accounting auditor.

Primary and official (confirmed August 31, 2026)

1. Tokyo Stock Exchange, "Corporate Information required for Timely Disclosure" (as of July 10, 2026; page updated July 3, 2026) — https://www.jpx.co.jp/equities/listing/disclosure/info/
2. Tokyo Stock Exchange, "IR Systems and IR Activities" (the obligation to develop an IR system under the Code of Corporate Conduct, and disclosure in the corporate governance report) — https://www.jpx.co.jp/equities/listing/investor-relations/index.html
3. Tokyo Stock Exchange, "Examinations pertaining to Disclosure of Corporate Information" (the five perspectives of disclosure examination. Securities Listing Regulations, Rule 415, Paragraph 1 and Rule 3, Paragraph 2) — https://www.jpx.co.jp/equities/listing/disclosure/examination/index.html
4. Tokyo Stock Exchange, "Overview of TDnet" (Company Announcements Disclosure Service, 31 days; Listed Company Search, the past ten years; paid database service, the past five years) — https://www.jpx.co.jp/equities/listing/disclosure/tdnet/index.html
5. Japan Exchange Group, "コーポレートガバナンス・コード（2026年7月版）の公表について (Japanese only; unofficial translation: "Publication of the Corporate Governance Code (July 2026 Edition)")" (partial amendment to the Securities Listing Regulations, effective July 21, 2026) — https://www.jpx.co.jp/corporate/news/news-releases/1020/20260721-01.html
6. Google Search Central, "Optimizing your website for generative AI features on Google Search" (updated July 10, 2026) — https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
7. Google Search Central, "What is canonicalization" (updated August 20, 2026) — https://developers.google.com/search/docs/crawling-indexing/canonicalization

Peer-reviewed and academic (confidence: medium to high. Reservations stated alongside)

1. Nakagawa, Hirano, Fujimoto, "Evaluating Company-specific Biases in Financial Sentiment Analysis using Large Language Models" (earnings reports on TDnet, 2019–2023. 10,249 observations with GPT-4o. An IEEE version exists. The direction of the bias is not consistent across models) — https://arxiv.org/html/2411.00420v1

Preprint (not peer reviewed. Treated with hedging)

1. A study of high-confidence errors using FinQA (even where all eight re-answers to the same question agreed, 15 to 23% were wrong. A July 2026 preprint. It is a benchmark of foundation models, not an evaluation of publicly available AI search products) — https://arxiv.org/abs/2607.11414

Practitioner report (interested party. Independent replication needed)

1. GenAI as a Reader (USTP, HHL and nexxar, 2026. 20 listed companies in Europe. Answer accuracy 71% for the HTML group and 54% for the PDF-centered group; subsample for accuracy n=200. Even in the HTML group 29% were inaccurate. nexxar is a digital IR reporting vendor and is therefore an interested party. Its peer review status has not been confirmed) — https://digital-investor-relations.com/_assets/downloads/DIR_GenAI-as-Reader.pdf?h=E0wP6LL9

Matters this article could not confirm

• Public data measuring how often AI mistakes the version of a disclosure — within the scope of this review, we could not confirm it
• The effective date of the obligation to develop an IR system — it could not be confirmed in a primary source, so this article does not refer to it
• The scope of application by market segment of the July 2026 edition of the Corporate Governance Code — this article has not checked it, so no segment-by-segment description is offered