AIO & LLMO for IR | Does Your Earnings Call Reach AI? — Video, Transcripts, and What Never Becomes Text

The earnings call sits at the center of a listed company's IR activity. Yet most of what is said there exists only inside the video and the slide deck PDF. The Q&A session leaves no record at all unless it is transcribed.

Asking whether the content of your earnings call reaches AI is asking whether text corresponding to what was said at that call exists in a place from which it can be retrieved. Where no corresponding text exists, then at least as regards the text route, you cannot confirm whether that content is in a state where it can be retrieved or referenced. What cannot be confirmed cannot be managed either.

This article does not argue that you should stop producing video. What it can address goes only as far as one question: does text corresponding to what was said exist in a place from which it can be retrieved? As for how much citation increases if you publish a transcript, we have not been able to confirm published data supporting any such figure.

• Why "the text exists but cannot be retrieved" and "no text exists in the first place" are different problems
• How far the IR activity of Japanese listed companies has moved online, and what is out on the web (with denominators)
• Of the information generated around a briefing, what exists as text and what does not
• Where the Q&A session sits within the rules, and why ownership of it tends to become ambiguous
• How earnings call materials, video, and transcripts are treated in TSE's list of information requiring timely disclosure
• The questions that arise when you publish a transcript (who produces it, how much to include, corrections, English)
• A procedure for checking, yourself, whether what you said reaches AI

IR staff at listed companies, and in particular companies that hold their earnings calls online and publish the video on the web. Companies that outsource video production and delivery to an IR support firm are included. This is not an article about how the rules should be designed; it is an article for taking stock of material you have already put out. It does not address individual legal judgments.

• 73.9% of companies hold their domestic earnings calls online, and 73.1% disclose them on the web (Japan Investor Relations Association, 33rd Survey on IR Activities; base: companies that conduct IR activities)
• 65.8% answered that they either use generative AI in their work or are trialing it in their work for producing meeting minutes for briefings and similar events (8.1% in the previous, 2024 survey; base: companies that conduct IR activities)

1-1. Separating two kinds of "not being read"

Not being read by AI looks like a single phenomenon, but the causes fall into at least two layers, and the remedies differ.

This distinction is the starting point of this article. No amount of work on retrievability helps where no text exists, because there is nothing to retrieve. Conversely, text that is produced but cannot be retrieved does not arrive either. Both are needed, but what should be checked first is whether the thing to be retrieved exists at all.

The retrievability side — the state in which text does exist but is not read — is handled by a separate article in the IR lane (which covers the technical structure of IR sites and retrieval by AI). This article does not enter that general territory.

1-2. The state of having no way to check

Statutory disclosure and timely disclosure have a correct answer. Annual securities reports, earnings reports (*kessan tanshin*), and timely disclosure are published through prescribed routes and can be retrieved mechanically. That ground was covered in the parent article (How AI describes your financials).

The problem lies outside it. The business outlook management described in its own words, what investors were concerned about in the Q&A session and how those questions were answered, the spoken commentary attached to a chart — these play a substantive part in understanding the published materials. And yet, where no corresponding text exists, there is no means of confirming whether that content is in a state where it can be retrieved or referenced as text.

The point is not that something is conveyed wrongly; it is that there is no way to check. An error can be detected by matching against published materials. But where no corresponding text exists, there is nothing to match against in the first place. From the outside, you cannot tell a thin answer from an absence of material.

1-3. What this article does not address

The scope is stated up front.

2-1. The premises of the survey (denominators first)

All the figures below come from the Japan Investor Relations Association's 33rd Survey on IR Activities (published May 2026). It surveyed all 4,088 listed companies as of January 2026, and 948 companies responded (a response rate of 23.2%). Of those, 917 companies, or 96.7%, answered that they conduct IR activities.

The base differs from question to question in this survey. Below, the base is stated alongside each figure. Please do not add together figures that rest on different bases.

2-2. Moving online, and disclosure on the web

The base for all of these is companies that conduct IR activities, and "previously" refers to the 2024 survey. Disclosure on the web rose against the previous survey across every option.

2-3. What is placed on the company's own website

The base is companies that answered that they have a website explicitly labeled "for investors" or "IR".

This is where the subject of this article appears. More than nine in ten place the presentation "materials" on their own website. But the materials are not the same thing as what was said. The slides are the skeleton; the commentary that puts flesh on them, and the Q&A session, sit outside the deck.

2-4. Generative AI for meeting minutes spread quickly

The same survey also asked about generative AI use in IR-related work. The base is companies that conduct IR activities, and the values are the share that answered either that they use it in their work or that they are trialing it in their work.

Producing meeting minutes stands at 65.8%. That is an increase of 57.7 points from 8.1% in the previous survey. It shows that use and trial of generative AI for producing minutes has spread quickly.

That, however, is as far as this figure goes. The share of companies that hold an internally available transcript of publishable quality, and the share for which every briefing has been turned into text, are not measured by this question. What can be written is one step: at companies that already produce minutes, whether to publish them becomes the next question.

Note also that among the challenges of adopting generative AI, "accuracy of information (so-called hallucination)" ranked highest at 77.2% (base: companies that conduct IR activities). That is consistent with the practical instinct that automatically generated minutes cannot necessarily be published as they are.

2-5. Some companies outsource video production and delivery

The same survey covers use of IR support firms as well (base: companies that conduct IR activities). 81.9% use one (previously 79.9%). Among the services currently in use, "video production and delivery" stands at 41.0% (previously 38.0%), and 32.7% (previously 30.5%) name it as a service they would like to use in future. The survey's own summary characterizes this as work that is difficult to handle with generative AI at present.

A structure becomes visible here. Some companies use an outside service for video production and delivery, while generative AI is used in-house for the minutes. At companies where outsourced video and in-house minutes are separate processes, the people responsible and the places the outputs live are separate too. The seam of that division of labor is where gaps tend to open.

2-6. The heavier the explanation, the more likely it sits in the presentation materials

One more from the same survey. Asked where they describe their work on the "Action to Implement Management that is Conscious of Cost of Capital and Stock Price" requested by TSE, companies most often named "earnings call materials for investors", at 80.7% (previously 78.3%). The Corporate Governance Report followed at 68.0%, the integrated report at 58.8%, and the IR site at 55.3%. The base is companies that answered, on disclosure of that work, either that they are responding appropriately, that they are responding but there is room for improvement, or that their response remains a formality.

In other words, the medium in which companies most often place their account of how they are responding to TSE's request is the earnings call deck. And the spoken explanation attached to that deck does not survive unless it is transcribed.

Hold an earnings call once, and information in several formats is generated at the same time. Here it is broken down by format.

3-1. The "numbers that do not exist" inside the deck PDF

Item 1 needs care. Even where you publish a PDF, if the key points are built as images, the numbers inside them do not exist as text. The material is public, and yet precisely the numbers that matter are missing. How to check this is in §8-5.

3-2. Text that accompanies a video page

On item 4, Google's official documentation is instructive. Its optimization guide for generative AI features (updated July 10, 2026) positions images and video as supporting the text content.

The documentation on video recommendations (updated December 18, 2025) states, as a condition for a video to be eligible for video features, that the page whose primary purpose is to watch that video must be indexed, and gives a list page presenting several videos as equals as an example of a page that does not count in this sense. It also explains making the page title and description unique to each video, and indicating important points through structured data or timestamps in the description.

What follows from this goes only one step. Text that can be placed alongside a video — the page title, the description, the chapter markers — does exist, and the official documentation recommends putting it in order. But none of it is the content that was spoken inside the video. Even a thorough set of chapter markers does not turn what was said in that chapter into text.

Note also that, within the scope of this review, we could not confirm any statement in that documentation listing captions or transcripts as a condition for indexing. This article does not write that "adding captions will get you read".

3-3. The unit of the inventory is not the event but what was said

The practical trap is getting the unit of the inventory wrong. At the level of "we have an earnings call page" or "we publish the video", the gaps are invisible. The unit to check is each individual thing that was said there. How to go about it is set out in §8-4.

4-1. It sits outside the reach of the rules

The Q&A session shows, at the same time, what investors are concerned about and how management answered. And yet this part sits outside the area where the rules set out an explicit disclosure obligation.

There is no item for a Q&A session in TSE's list of information requiring timely disclosure (§5-1). Nor is the Fair Disclosure Rule a provision that uniformly requires the content of briefings to be published (§5-2).

Separate form from content here. What is not uniformly subject to a publication obligation is the form: a full text or transcript of the Q&A session. Whether publication is required when material information has been conveyed there is judged separately. Because there is no obligation as to form, it is unlikely to be treated as something to be managed; and because it is not managed, ownership becomes ambiguous.

That said, it is not unrelated to the rules. The Cabinet Office Order that sets out the details of this rule lists, among those with a high likelihood of trading, attendees at a meeting held for the purpose of providing information about the operations, business, or property of a listed company to particular investors, and confines this to the time they are attending that meeting. In other words, attendees at a briefing that meets these requirements may be included among "transaction-related parties" while they are present.

The Financial Services Agency's guidelines, meanwhile, set out the position that information of the kind generally provided at a segment briefing — information that can be used for investment decisions when combined with other information but that cannot be said by itself to have an immediate effect on such decisions (so-called mosaic information) — is not, in itself, considered to fall within the information covered by this rule.

Put these two together and the position of the Q&A session comes into view. Attendees may qualify as transaction-related parties. Yet much of what is said there may be treated as not amounting to material information in itself. The result is that the Q&A session occupies a middle ground: not subject to a disclosure obligation, but used in investment decisions.

Note that these guidelines state at their own outset that they present a general interpretation of the law, do not answer whether they apply to any individual case, and do not bind judicial determinations. For how they apply to your own circumstances, always confirm with your legal department or outside counsel.

4-2. The scope of access has long been the company's to decide

The structure in which a company can decide who is present is nothing new. A 2003 study published in a peer-reviewed accounting journal (Bushee, Matsumoto, and Miller, *Journal of Accounting and Economics*, vol. 34) examined the determinants and effects of providing unrestricted real-time access to conference calls (that is, holding open calls), and reports that this decision was associated to some degree with investor composition and the complexity of financial information, and that open calls were associated with increased small trades and greater price volatility during the call period.

Do not apply this study directly to the present. Its subjects were US companies around the year 2000, and the question was real-time audio access. It is not research about AI, and what it reports is association, not proof of causation.

Even so, one structure has carried over. What decides who receives the information is not the rules but the company's own practice. Once, the issue was whether to open the phone line. Now, whether what was said remains as text and is placed somewhere it can be retrieved occupies the same position.

4-3. TSE itself once called it a gap

On October 24, 2019, the Tokyo Stock Exchange announced the start of a pilot program on the delivery of transcripts of investor events held by listed companies. The announcement described the background as follows. In Europe and the United States, transcripts of investor events have long been delivered globally, providing information with a high degree of transparency and fairness. In Asia, including Japan, such delivery has not progressed, and an information gap has arisen between those who attend an event and other investors.

This is an announcement from 2019 and cannot be cited as TSE's current view. What remains, however, is the fact that a state in which only those present know something was recognized as a problem.

Back then, the other side of the gap was the investor who was not there. The detour such an investor takes afterward runs through news articles and third-party summaries. Where no corresponding text exists on the company's side, the record survives only along that detour. Regardless of who — or what — follows that record, what can be referenced is not the account the company gave, but the account a third party summarized.

4-4. The research side shows the same bias

A review article on earnings calls and large language models published in a peer-reviewed journal (Bongale and Shrivastava, *Discover Artificial Intelligence*, vol. 6, article 688, 2026, open access) notes that research in this field is concentrated on the US market, and gives as reasons the standardization of disclosure and the comparative availability of transcript data.

This is a description of how research is distributed, not a statement about how well AI answers. It is suggestive nonetheless. In this research field, readily available transcripts are one of the conditions that support the formation of the objects of study.

5-1. What is on the timely disclosure list

The Tokyo Stock Exchange publishes a list of corporate information requiring timely disclosure (as of July 10, 2026; verified September 1, 2026). The categories and the number of items are as follows.

What is listed as financial information is the earnings report and the quarterly earnings reports. No item naming earnings call materials, video, or transcripts is placed on this list.

That said, the 39th decision item and the 29th occurrence item are catch-all items covering important matters and facts concerning the operations, business, or property of a listed company. Whether any particular content falls within them is a case-by-case judgment, and this article offers no such judgment.

It cannot be written that TSE requires the measurement or management of AI-space perception. Putting an IR framework in place is an obligation under the Code of Corporate Conduct, and the state of that framework is required to be disclosed in the Corporate Governance Report. But measuring and managing how your company is perceived in AI space is not an obligation. What this article addresses is one step: within the existing IR framework, how far the gap between what was said and what exists as text should be treated as something to manage.

5-2. Do not mix up the nature of each instrument

The Financial Services Agency's guidelines give the purpose of this rule as securing fair disclosure of information to investors, and state that listed companies subject to it are expected to disclose information proactively. This is an expectation, not a provision that obliges companies to publish the content of their briefings.

5-3. Routes for placing text exist on TSE's side as well

There are practical routes around the rules. All of the following were verified on September 1, 2026.

TSE IR Movie Square (page updated August 28, 2026; 東証IRムービー・スクエア, unofficial translation). A service through which the Tokyo Stock Exchange delivers investor-facing video from listed companies — company introductions, messages from the president, and the like — on a TSE channel on a video platform. The videos are grouped into business introductions, company introductions, earnings explanations, and messages from the president, and the guidance for listed companies explains that use is free of charge.

The Event Transcript Service (page updated April 1, 2026; イベントトランスクリプト提供サービス, unofficial translation). Under a partner agreement with a listed company, a group company of JPX Market Innovation & Research, Inc. attends earnings calls, ESG briefings, mid-term management plan briefings, shareholder meetings, and similar events, and produces comprehensive transcript articles in Japanese or another local language and in English. The data supplied consists of transcript articles (PDF, TXT, XML, JSON) and event audio (mp3). The service is described as covering events at more than roughly 1,300 companies in the Japanese market as of April 2026. On the investor side it sits in the paid information category.

What matters to a listed company is the following. The transcript articles produced are supplied through a portal site for listed companies, and once supplied, may be posted on the corporate website (with editing and correction left to the company) or handed out at one-on-one meetings with institutional investors. A contract is required.

In short, a route for turning what you said into text and placing it on your own site exists on the exchange group's side as well. Indeed, on Japan Exchange Group's own IR site, the presentation materials and the full transcript of its own earnings calls are offered side by side.

That said, within the scope of this review, we could not confirm the costs or contract terms on the listed company's side (§9).

This is not a matter of saying "let us publish transcripts" and stopping there. Actually doing it calls for judgment. The questions are set out below. Which answer is right is not settled in this article.

6-1. Who produces it

6-2. How much to include

Full text or summary is a substantive choice. A full text preserves the phrasing as it was said, including hesitations and imprecise expressions. With a summary, the party deciding what to keep and what to drop becomes your own company.

It cannot be written that one is better than the other. What can be written goes only this far: which one you choose changes the range of text AI could read. If you choose a summary, you need at a minimum to know internally what you dropped.

6-3. What to do when an error is found

Publish a transcript and errors may come to light afterward.

Where a correction would add previously unpublished information, confirm with your legal department or equivalent whether that information constitutes material information and whether publication is required. Where it does constitute material information, a response in line with the Fair Disclosure Rule becomes necessary. An AI answer being wrong is not a reason to bring out unpublished information without going through that confirmation.

Managing the fact that versions change after publication — which is the latest, and as of when — is the province of a separate article in the IR lane (which covers managing disclosure versions and states).

6-4. Where this sits under fair disclosure

The Financial Services Agency's guidelines state that where a transaction-related party points out that information conveyed to them may constitute material information, possible responses through dialogue include: (i) publishing promptly where the company agrees with the observation; (ii) not publishing where the conclusion is that it does not apply; and (iii) where it does apply but publication is not appropriate, having the party assume a duty of confidentiality and a duty not to trade, limited to the period until publication becomes possible.

Unintended communication is also provided for in the Cabinet Office Order, and the guidelines give as an example a case where an officer happens, in the flow of conversation, to convey material information that there was no plan to convey.

What can be read from this is the following. What the rules contemplate is whether material information was conveyed — not whether the entire content of a briefing was published. It cannot be written that not publishing a transcript is, in itself, a breach of this rule. Conversely, where a Q&A session strays unintentionally into material territory, whether publication is required has to be judged regardless of whether a transcript exists.

In that sense, producing a transcript (or minutes) has practical value in its own right, separate from whether you publish it. It leaves a record against which what was said can later be checked and matched.

6-5. What to do about English

English-language disclosure of presentation materials stands at 68.0% (previously 61.8%; base: companies that answered that they have a website explicitly labeled "for investors" or "IR"). The outside service in §5-3 produces transcript articles in both Japanese and English.

Where you publish in English, the question narrows to whether the scope, the point in time, the figures, and the content of the summary diverge between two sets of official information. How AI answers differ between Japanese and English is the province of a separate lane covering cross-border questions.

According to the Japan Investor Relations Association's 33rd survey, 28.9% of companies use social media for IR (previously 24.6%, and 14.9% the time before), with 13.9% considering it (previously 11.6%). The base is companies that answered that they have a website explicitly labeled "for investors" or "IR".

On the question of which media they use (base: companies that answered that they use social media or are considering using it), video platforms came to 51.2% (previously 54.6%), X to 36.6% (previously 33.1%), and Facebook to 17.1% (previously 18.1%).

7-1. Content placed outside the company

The defining characteristic of what you post on social media is plain. It sits in a place your company does not control. The TSE service in §5-3 delivers on a video platform as well.

Display conditions, whether retrieval is possible, and how long content is retained all depend on each platform's terms and technology, and the range a company can control is limited. How long, and in what form, what you posted remains referenceable is not determined by the company's intentions.

7-2. This is as far as it goes

What this article can write goes only one step: it depends on the platform. How much content from which platform appears in AI answers is handled by the communications lane's What AI cites about your company.

What can be checked in practice is a single point. Does text corresponding to what you posted on social media also exist on your own site? If it does, one referenceable version exists under your control. If it does not, that content exists only outside your control.

This is the core of the article. The general design of verification testing that matches AI answers against published materials — the columns of a reference table, what to log, how to define the indicators — belongs to the article on how to implement verification testing. This section is confined to the checks specific to whether what was said has become text.

8-1. Establishing the starting point

In the Japan Investor Relations Association's 33rd survey, 51.1% of companies named "the number of analysts and investors attending briefings" as an indicator for measuring the effect of IR activity (previously 47.3%; base: companies that conduct IR activities). The survey notes that every option outside the top six fell below 30%, throwing into relief how difficult it is to measure contribution.

Briefings are measured by attendance. That is a natural indicator, but it does not measure what reached those who were not there. How far what was said became referenceable afterward cannot be read from attendance.

8-2. Building questions in three layers

This is the core procedure of the section. Divide what was said at the briefing into three layers according to where the text lives.

Build questions from each layer: three to five from Layer A, three to five from Layer B, two to three from Layer C. Phrase the questions the way an investor who wanted that information would actually phrase them. With internal jargon or in-house abbreviations, getting no answer is a foregone conclusion, and you are not measuring anything.

For Layer C, decide in advance how you will interpret an answer if one comes back. If an answer consistent with content that should exist nowhere does come back, first check that it is not a wrong answer or a guess, and then check the cited URLs (§8-6). If a record did exist outside your company, that is a discovery, not an achievement.

8-3. Reading which layer the answers reach

Sort the answers into three kinds.

1. Cannot answer (an answer stating that it does not hold the relevant information, or one that stays at the level of generalities)
2. Answers within Layer A (states only what is within the material written in the deck)
3. Goes into Layer B or Layer C (touches on content outside the deck)

Always separate "cannot answer" from "answers wrongly". The first is a state in which text does not exist or is not arriving; the second is a state in which wrong content has been taken in from another source, and the remedies differ. The first is about producing text; the second is about matching against published materials (the province of the parent article).

Stopping at Layer A is not itself a failure. It means content at the Layer A level is appearing in the answers (whether the deck itself was referenced is checked in §8-6). What becomes a problem is where Layer B and Layer C content matters to investment decisions and yet no text for it exists anywhere.

8-4. Building an inventory table of "what was said"

Alongside the measurement, check your own stock. Build it by content, not by event.

The purpose of this table is to make the location of the gaps visible. Rows where location is "nowhere" and the layer is B or above are the gaps this article has been pointing to. Whether to fill them is a management decision informed by how much the content matters and by the questions in §6.

Recall §2-4. At companies that already produce minutes, rows reading "published: internal only" may line up. In that case, the next question is not the production process but the decision to publish.

8-5. Whether the figures and tables in the slides are images

Even where you publish a deck as a PDF, the key points may be embedded in images. Checking is simple. Look at it by figure, not by page.

Open the PDF you publish in a browser or viewer, and try selecting and copying the numbers inside a chart or table. If you cannot select them, then at least in that viewer they cannot be confirmed as selectable text. Segment results, progress against the mid-term management plan, charts on cost of capital and return on capital — the parts investors care most about are often the ones turned into images for the sake of the layout.

As noted in §2-6, the medium most often chosen for describing how a company is responding to TSE's request is the earnings call deck. It is worth checking whether that account is embedded as an image.

Note that text existing does not guarantee it will be retrieved. Retrieval routes are the province of the technical article in the IR lane. What is being checked here is only the prior step: whether the thing to be retrieved exists.

8-6. Where the URLs cited in an answer point

When an AI answer shows source URLs, record where they lead.

A run of "third-party articles only" is the state this article has been dealing with. As set out in §4-3, where what was said has not become text on your side, what gets referenced is the account a third party summarized.

That said, the question of how much each type of source is cited overall — the composition of citations — is not addressed here. It is the province of What AI cites about your company. What is being looked at here is your own side: whether text corresponding to what your company said is on the reference route at all.

8-7. Tie the timing of re-measurement to the event

A single measurement tells you nothing. Repeat the same questions under the same conditions, at fixed points.

• Before the earnings call (how the previous period's content is being described)
• Immediately after the earnings call
• After the deck PDF is published
• After the video is published
• After the transcript or summary is published (where you publish one)
• And again after a set interval

It cannot be written that "publishing a transcript increases citation by X%". Within this review, we could not confirm published data showing such an effect (§9). What can be written goes only as far as the procedure: observe before and after publication, and record what changed. Whether an observed change was caused by publication cannot be separated from other factors (the results themselves, the volume of press coverage, market conditions).

8-8. How many times, and against what, to measure

The answer to a single question put once to a single model cannot be treated as a standing perception. Answers vary from run to run even for the same question, and tendencies differ by model. You need to measure several times, against several AI engines, in a state where no prior conversation carries over.

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.

How you design the number of runs and what you run them against depends on what you want to measure. What matters is fixing the conditions and repeating, not any particular number in itself.

8-9. Things to watch when you measure

On the last row: within the scope of this review, we could not confirm primary material from a regulator or professional body recommending exposure volume in AI space as a formal IR indicator. Rather than exposure volume, looking at which layer of what you said the answers reach, and at what those answers cite, is better suited to practice — that is this article's inference.

So that the claims in this article are used correctly, the items we could not confirm are set out explicitly. If another source states any of these definitively, please check the basis for it.

Q1. Should we stop publishing video and publish only transcripts?

That cannot be written. Video conveys things that expression and tone carry with them, and some investors expect to watch. What this article points to is whether text corresponding to what was said exists in a place from which it can be retrieved. The two are compatible.

Q2. Is it that AI cannot understand the content of video?

This article makes no such assertion. The specifications for how the major AI services handle video and audio are not published, and we have not been able to confirm primary material on the point. What can be checked is a single thing: where content exists as text, you can observe whether it is being referenced.

Q3. Isn't publishing the deck PDF enough?

The deck is not the same as what was said. Assumptions and background not written in the deck, and points raised in the Q&A session, sit outside it. In addition, where figures and tables are images, their numbers cannot be confirmed as selectable text (how to check this is in §8-5).

Q4. Is there an obligation to publish earnings call transcripts?

No item naming earnings call materials, video, or transcripts is placed on TSE's list of information requiring timely disclosure (as of July 10, 2026). Nor is the Fair Disclosure Rule a provision that uniformly obliges companies to publish the content of briefings. This, however, is a matter of form. Whether publication is required when material information has been conveyed there is judged separately.

Q5. Does publishing a transcript create a problem under fair disclosure?

What this rule governs is the communication and publication of material information; it is not a provision that prohibits publishing the content of a briefing as such. If anything, a record against which what was said can later be checked has practical value when judging whether publication is required after an unintended move into material territory. For judgments on individual cases, confirm with your legal department or outside counsel.

Q6. Which is better, a full text or a summary?

This article does not settle it. A full text preserves the phrasing as it was said; with a summary, your company decides what to keep. The choice changes the range of text AI could read. If you choose a summary, we recommend knowing internally what you dropped.

Q7. We produce minutes internally. Can we just publish those?

Not as they are. Minutes written for internal use are not written on the assumption of external publication. That said, use and trial of generative AI for producing minutes is spreading (65.8% in the 33rd survey; base: companies that conduct IR activities). At companies that already produce minutes, the next question is not the production process but the decision to publish.

Q8. If we outsource transcription, can we put the result on our own site?

Under the Event Transcript Service provided by a group company of JPX Market Innovation & Research, Inc., transcript articles are supplied through a portal for listed companies, and once supplied may be posted on the corporate website (with editing and correction left to the company), as the service describes it (verified September 1, 2026). A contract is required.

Q9. Does publishing a transcript increase the number of times AI cites us?

We could not confirm published data showing such an effect. What can be written goes only as far as the procedure: observe before and after publication, and record what changed. An observed change cannot be explained as the effect of publication.

Q10. Where should we start?

We recommend building the inventory table in §8-4 for a single recent earnings call. Break what was said into one-line units and fill in where the text lives — that is all. The rows reading "nowhere" are the gaps this article has been pointing to. From there, build the three-layer questions in §8-2.

The earnings call sits at the center of a listed company's IR activity, and it is also the medium most often chosen for describing how a company is responding to TSE's request (80.7% in the 33rd survey). Moving online has advanced, and so has disclosure on the web.

Even so, a gap remains between what was said and what exists as text. Neither the commentary around the presentation nor the Q&A session exists anywhere but in the audio and video unless it is transcribed. At least as regards the text route, you cannot confirm whether that content is in a state where it can be retrieved or referenced.

There are footholds, though. Use and trial of generative AI for producing minutes has spread (65.8%; base: companies that conduct IR activities), and a route for putting transcripts on your own site exists on the exchange group's side. What companies that already produce minutes face next is the inventory, and the decision to publish.

This article's proposal is simple. Take stock by content rather than by event, check where the text lives, and observe which layer the answers reach using questions divided into three layers. What comes after that — how much to publish, and where to stop — is a management decision informed by how much the content matters and by your own organization, and it is not for this article to make.

For the overall approach to managing AI-space perception on a continuing basis, see How AI describes your financials; for how to implement matching against published materials, see the article on how to implement verification testing.

Sources verified: September 1, 2026

Primary and official

1. Japan Investor Relations Association, "Summary of Results of the 33rd Survey on IR Activities" (published May 2026; covering all 4,088 listed companies as of January 2026, 948 responses, response rate 23.2%) — https://www.jira.or.jp/download/survey/202605_summary.pdf
2. Japan Investor Relations Association, "Surveys and Research" (survey overview) — https://www.jira.or.jp/activity/research.html
3. Tokyo Stock Exchange, "Corporate Information Requiring Timely Disclosure" (page updated July 3, 2026 / content as of July 10, 2026) — https://www.jpx.co.jp/equities/listing/disclosure/info/
4. Japan Exchange Group, "TSE IR Movie Square" (page updated August 28, 2026) — https://www.jpx.co.jp/listing/ir-clips/ir-movie/index.html
5. Japan Exchange Group, "Event Transcript Service" (page updated April 1, 2026) — https://www.jpx.co.jp/markets/paid-info-listing/transcripts/index.html
6. Tokyo Stock Exchange, Navigation System for Listed Companies, "Services for Streamlining IR Operations" — https://faq.jpx.co.jp/disclo/tse/web/knowledge8521.html
7. Tokyo Stock Exchange, Navigation System for Listed Companies, "TSE IR Movie Square" — https://faq.jpx.co.jp/disclo/tse/web/knowledge6989.html
8. Japan Exchange Group, "SCRIPTS Asia and the Tokyo Stock Exchange Begin a Pilot Program on the Delivery of Transcripts of Investor Events Held by Listed Companies" (October 24, 2019) — https://www.jpx.co.jp/corporate/news/news-releases/0060/20191024-02.html
9. Japan Exchange Group, "Launch of the Event Transcript Service" (April 3, 2020) — https://www.jpx.co.jp/corporate/news/news-releases/0060/20200403-01.html
10. Japan Exchange Group, "IR Briefings" (an example of publishing its own presentation materials and full transcripts side by side) — https://www.jpx.co.jp/corporate/investor-relations/ir-library/events/index.html
11. Japan Exchange Group, "For Listed Companies" (list of services for listed companies) — https://www.jpx.co.jp/listed-companies/
12. Financial Services Agency, Planning and Coordination Bureau, "Points to Note Regarding the Provisions of Article 27-36 of the Financial Instruments and Exchange Act (Fair Disclosure Rule Guidelines)" (established April 1, 2018) — https://www.fsa.go.jp/common/law/kaiji/20180206-2.pdf
13. "Cabinet Office Order on the Publication of Material Information under the Provisions of Chapter II-6 of the Financial Instruments and Exchange Act" (Japanese Law Translation Database System) — https://www.japaneselawtranslation.go.jp/ja/laws/view/3431
14. Japan Exchange Group, "Establishing an IR Framework" — https://www.jpx.co.jp/equities/listing/investor-relations/index.html
15. 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
16. Google Search Central, "Video SEO best practices" (updated December 18, 2025) — https://developers.google.com/search/docs/appearance/video
17. Google Search Central, "AI Features and Your Website" — https://developers.google.com/search/docs/appearance/ai-features

Peer-reviewed research

1. Bushee, B. J., Matsumoto, D. A., & Miller, G. S. (2003). Open versus closed conference calls: the determinants and effects of broadening access to disclosure. *Journal of Accounting and Economics*, 34(1-3), 149-180. — https://www.sciencedirect.com/science/article/abs/pii/S0165410102000733
2. Bongale, A., & Shrivastava, K. (2026). A comprehensive review of open source large language models for the analysis of earnings call reports and financial documents including applications, datasets, and long term challenges. *Discover Artificial Intelligence*, 6, 688. (open-access review) — https://link.springer.com/article/10.1007/s44163-026-01334-9
3. Shah, A., Ye, M., Jaskowski, S., Xu, W., & Chava, S. Beyond the Reported Cutoff: Where Large Language Models Fall Short on Financial Knowledge. (accepted at COLM 2025; the arXiv version was consulted) — https://arxiv.org/abs/2504.00042

Pre-peer-review and interested-party studies (cited with qualifications)

1. Brunswick, "U.S. Investor Survey 2026" (100 US institutional investors. A self-reported study by an IR support firm, an interested party) — https://review.brunswickgroup.com/article/investor-survey-2026/
2. QUICK, "SCRIPTS Asia" service overview (the delivery provider's own description, covering the formats supplied and delivery times) — https://corporate.quick.co.jp/products/scripts-asia/

• Sources verified: September 1, 2026
• Bases for the figures: the Japan Investor Relations Association's 33rd Survey on IR Activities uses a different population for each question. The base is stated alongside each figure in the body. Please do not add together figures that rest on different bases.
• Statements about law and regulation: this article is for general information and is not legal advice. For judgments on individual cases, always confirm with your legal department or an attorney.
• Where AIO, LLMO, and AIPM stand: these are not names of official standards; they are practitioner terms. This article treats them as a practical area concerned with improving citation, mention, display, and impression within generative AI answers. Google states officially that no additional requirements and no dedicated markup are needed to appear in its generative AI features, and that conventional SEO fundamentals are the foundation.