AI Perception Management by Department: Survey Figures on How Candidates, Investors, Buyers and Journalists Use AI

AI perception management by department means working out, for each department, the questions its stakeholders (candidates, investors, corporate buyers, journalists, business partners and consumers) are likely to put to AI, then measuring and managing how your company is described in the AI's answers to those questions. This article covers five departments: HR, investor relations (IR), sales and marketing, PR, and legal. For each, it sets out how the people the department deals with have begun to use AI, giving the survey figures together with the conditions under which they were collected.

The figures mean different things in each department. In a Japanese survey of 419 job-hunting students, 76.2% said they "often" or "sometimes" use generative AI to gather information for job hunting, and 37.9% said they used it at the company-research stage. In a Japanese survey of individual investors, 31.8% of individual investors currently investing in stocks used AI to look for stocks to buy. In a Japanese survey of 1,573 people involved in corporate purchasing, 41.6% said the influence of generative AI and AI search on their purchase decisions had increased. For journalists, within the scope of this review, we could not confirm a figure limited to Japan; in a global survey of 1,899 journalists, 21% said they had never used AI in their work.

The four departments above have figures showing that their stakeholders use AI. Legal does not. Within the scope of this review, we could not confirm any direct figure showing that business partners or consumers research companies with AI. For legal, what we can show goes no further than a survey of how generative AI users check the answers they get, and the kinds of problems reported in AI answers about news outside Japan. In a white paper from Japan's Ministry of Internal Affairs and Communications (MIC), 38.2% of generative AI users in Japan said they check whether the answer is correct by other means (multiple answers allowed). The share of business partners or consumers who took an AI answer about a company as fact is not known.

None of these figures supports saying that most stakeholders research companies with AI. Whether a figure is a share of all respondents or a share of AI users also differs from survey to survey. What can be said is that in each of the four departments, some stakeholders have begun to add AI to the material they use to make decisions. Even so, since some stakeholders may come across an AI answer before they decide, how AI describes your company becomes something each department needs to check. Different stakeholders need different information to make their decisions.

What this article covers

• For HR, IR, sales and marketing, and PR: survey figures showing what stakeholders do with AI, with the surveyor, period, target and number of respondents
• For legal: that we could not confirm any direct figure showing stakeholders research companies with AI, and the indirect evidence that can be shown instead
• Which figures are from Japan and which are from elsewhere, and which figures we could not find
• What can happen in each department when an AI answer is wrong
• Why AI's description of the same company can change depending on the question
• How to build questions for each department, how to set up phrasing sets, and what the mention rate and the alignment rate mean
• What the figures and measurements in this article cannot tell you

Who this article is for

Heads of HR, IR, sales and marketing, PR and legal, and anyone who explains AI perception management inside their company. It is an overview for readers who want to check whether the same thing is happening outside their own department, including the conditions behind the figures. The detail for each department is left to the in-depth articles introduced in each section.

The figures up front

The figures in Table 1 cannot be compared with one another. The form of the question, the base and the way respondents were selected differ from survey to survey. Shares of all respondents are mixed with shares limited to AI users. The table puts them side by side to show, condition by condition, how far each department can say that its stakeholders are using AI. The legal row alone is not a figure for stakeholders' AI use.

1-1. "Using AI" means something different in each survey

The surveys in Table 1 do not ask the same thing, even when the result reads as "using AI". The student survey asks whether AI was used to gather information, and the investor survey asks whether it was used to look for stocks. The purchasing survey asks whether its influence on purchase decisions has increased. The journalist survey asks whether respondents have never used AI in their work. The white paper figure used in the legal section asks how AI users handle the answers they get.

None of the surveys asks whether stakeholders researched "your company" with AI. What they researched was companies in general, stocks in general, and products and services in general. This article therefore does not write that "your stakeholders are researching your company with AI". What it can write is that AI has begun to enter the process by which stakeholders gather the material for their decisions.

1-2. Why the survey conditions are set out

There are three reasons for setting out the conditions of each survey.

First, many of the surveyors run businesses in this field. Nyle supports companies in responding to AI search; Siemple uses the same release to promote a guide to AI optimization for listed companies; HubSpot sells sales and marketing software; and Cision sells PR distribution and media information. The MIC white paper and the study led by public service broadcasters, used in the legal section, are not surveys by such businesses. A survey is not wrong just because the surveyor has an interest. But if the figures alone are passed around a company without readers knowing about that interest, readers lose the means to check the conditions.

Second, whether a figure is a share of all respondents or a share of AI users differs from survey to survey, and some figures do not state their base at all. In the purchasing survey, the base for 41.6% is all 1,573 respondents, but the base for 50.3% is 1,155 AI users. For the investors' 31.8%, the release does not let us confirm which stage it comes from, the 8,000-person pre-screening or the 260-person main survey, or how many people form the base. For the students' 37.9%, it is not stated whether the question was put to everyone or only to generative AI users.

Third, mixing figures from Japan with figures from elsewhere leads to misreading. The journalist figure comes from a global survey, and we could not confirm whether journalists in Japan were included. This figure does not allow us to say the same of journalists in Japan.

2-1. Who the stakeholders are

HR's stakeholders are candidates, both new graduates and mid-career hires. Candidates research a company before applying, research it again before an interview, and compare it with others after receiving an offer.

2-2. What they research with AI

Nyle Inc. ran an online survey from March 25 to 30, 2026 (published April 28, 2026) of 419 students graduating in 2026 to 2028 who had job-hunted or were job-hunting. In it, 34.4% said they "often" use generative AI to gather information when job hunting and 41.8% said they "sometimes" do. The release describes this as "more than 70%"; the total of 76.2% is the sum this article calculated.

The stage at which the most students used it was preparing for selection (application forms and interviews), at 48.0%, followed by company research at 37.9% and industry research at 33.2%. The company information they looked up with generative AI was the business description (36.7%), strengths and characteristics (33.5%), and differences from competitors (29.8%). What candidates ask AI about is a company's business, its strengths and how it differs from others. All of these should already be on the company's careers page.

2-3. What can happen when the AI answer is wrong

The business may be described in an outdated form, strengths may be confused with those of another company, or the company may come off badly in a comparison with others. Candidates then read the careers page with that as their first impression. In the same survey, about 90% of students said they check the facts on the official website or careers page after researching with generative AI. However, the survey does not tell us how an impression formed before checking changes after checking. The other questions in this survey are covered in more detail in the HR hub article.

2-4. What HR should look at first

What AI says in response to the questions candidates are likely to ask, such as "What kind of business is Company X in?" or "What is the difference between Company X and Company Y?", and whether that description matches the careers page and the company profile. HR should also check whether the company's name comes up when the question is asked without naming it, such as "Which companies in the [industry] are good at developing young employees?"

2-5. In-depth articles

Data on how candidates use AI, and how to design recruitment communications in practice, are covered in Candidates Are Researching You With AI. What job posting structured data does for AI is covered in Does JobPosting Structured Data Help in AI Search?, and whom AI recognizes as the employer in placement and staffing is covered in Who Does AI Think Is the Employer?

3-1. Who the stakeholders are

IR's stakeholders are individual investors, institutional investors and analysts. The only group for which this article could confirm a Japanese figure was individual investors.

3-2. What they research with AI

A survey published on September 1, 2026 by Siemple Inc. and the Digital Crisis Research Institute covered individual investors in their 20s to 60s across Japan, in two stages: a pre-screening of 8,000 people and a main survey of 260. The release states that 31.8% of individual investors currently investing in stocks already use AI to look for stocks, and 58.9% of those in their 20s. From the body of the release, we could not confirm which stage and how many people form the base for 31.8%, the survey period, or the wording of the question. The same release also states that 51.5% of investors went directly to the IR page after AI had presented a company.

Within the scope of this review, we could not confirm any Japanese figure for institutional investors and analysts. A survey of institutional investors outside Japan is covered in the IR hub article.

This article does not use another figure in the same release, that "AI presents only 0.16% of listed companies". The number of companies AI presents at a time is given as an average of 4 to 5 in the body of the release and an average of 6.6 in its key-points section, and we found no description of how it was measured.

3-3. What can happen when the AI answer is wrong

In IR, the problem tends to be not reputation but numbers, and which company is being discussed. AI may mix up fiscal periods, present a forecast from before a revision as the current figure, or blend the company with what it was before a change of trade name. Individual investors then read this at the stage of narrowing down candidate stocks. A distinctive feature of IR is that statutory disclosure provides an institutional reference to check against, which makes it easier to judge whether an answer is wrong. The IR hub article covers this point in detail.

3-4. What IR should look at first

Whether the answers to questions such as "What are Company X's latest results?" or "What is Company X's dividend policy?" match the earnings report (kessan tanshin) and the annual securities report. IR should also check whether the company comes up, and which companies it is listed alongside, when the question is asked without naming it, such as "Which stocks in the [field] pay high dividends?"

3-5. In-depth articles

The entry point is How AI Describes Your Financials. How to design a check of AI answers is covered in How to Verify What AI Says About Your Financials; which version of a disclosure AI describes, in Which Version of Your Disclosure Is AI Describing?; being mistaken for another company, in When AI Mistakes You for Another Company; and the technical side, in Do AI Crawlers Read Your IR Site's JavaScript? and Should You Block AI Crawlers on Your IR Site? Whether AI can read the contents of disclosure documents is covered in an article available in Japanese only. Earnings presentations are covered in Does Your Earnings Call Reach AI?

4-1. Who the stakeholders are

The stakeholders for sales and marketing are corporate buyers, the people on the ground who are involved in purchasing, and the people who sign off.

4-2. What they research with AI

HubSpot Japan planned an online survey that Macromill carried out from June 11 to 16, 2026 (published July 16, 2026). It covered 1,573 people working at companies with 50 or more employees who had been involved in comparing, selecting, requesting approval for, approving or signing off on B2B products or services in the previous 12 months.

• Compared with past purchases, the influence of generative AI services and AI search on purchase decisions increased: 41.6% (base: all 1,573 respondents)
• Named using generative AI services or AI search as an information-gathering behavior that had increased over the past two to three years: 33.5% (also all respondents)
• Use ChatGPT "frequently" or "sometimes" to gather information when considering a purchase: 58.0% combined (base not stated)
• Of 1,155 users of generative AI or AI search, used it to list candidate products and services: 50.3%. Used it to organize each company's features and differences: 47.1%; to create comparison tables: 39.2%

Pay attention to the base of each share. Of the 1,573 respondents, 1,155 were counted as users of generative AI or AI search, and 1,266 used AI in this particular purchase. These are two different questions. The 50.3% is a share "of users", not a share of all respondents.

In a separate question, of 1,431 buyers who contacted a seller, 63.5% had already progressed at least as far as listing several candidates by the time they made contact. The 50.3% and the 63.5% come from different questions with different bases. It cannot be said that listing candidates with AI changed when buyers contacted sellers.

On the other hand, when the same survey asked which of 10 kinds of information respondents found trustworthy, information compiled by generative AI services or AI search ranked last, at 37.8%. Of the 1,266 who used AI in this purchase, 24.9% used AI information "as is, or almost as is", while 48.3% used it after checking the facts or adding to it themselves.

4-3. What can happen when the AI answer is wrong

If a buyer lines up candidates based on an AI answer, a company whose name does not come up may be left out of the initial comparison. It may still be added from other sources of information, so it does not necessarily stay left out. Even if the company does come up, it can be put at a disadvantage in the comparison if it is described with an old price range or a discontinued feature.

4-4. What sales and marketing should look at first

Whether the company comes up, which companies it is listed alongside, and which features it is described by, when the question is asked without naming it, such as "Which services solve [problem]?" or "Which [product type] suits a company of about 300 employees?" Sales and marketing should also check whether the description given when the company is named matches the product pages and the pricing page.

4-5. In-depth articles

There is not yet an in-depth article for sales and marketing. How the department should approach the subject is summarized in How Sales & Marketing should approach AI perception.

5-1. Who the stakeholders are

PR's stakeholders are journalists and editors. Journalists research a company before an interview and check facts while writing.

5-2. What they research with AI

Cision, which sells PR distribution and media information, ran its 2026 journalist survey from January to February 2026 among 1,899 journalists in North America, Europe/Middle East/Africa and Asia-Pacific (published May 2026). The release refers to 19 markets, but we could not confirm whether Japan was among them.

According to the company's commentary article, only 21% of journalists said they had never used AI in their work. The reasons given for using it included generating ideas (48%), research and fact-checking (43%), and transcription and summarizing (41%). Within what has been published, we could not confirm the base of these shares (all respondents or AI users).

The 2025 edition (more than 3,000 journalists worldwide) put the share of journalists who said they use generative AI at 53%, and listed its main uses as research (25%), transcription (23%) and summarizing (20%). The 2025 and 2026 editions ask the question differently, so we do not set them side by side to say the figure went up or down.

5-3. What can happen when the AI answer is wrong

When journalists use AI for research or fact-checking, an AI answer that gives a former representative's name or a business that has ended can become part of the premise of an interview. After a scandal, a matter that has been resolved may also be described as if it were still ongoing.

5-4. What PR should look at first

Whether the answers to questions such as "What has Company X been doing recently?" or "Who heads Company X?" match the press releases and the company profile. PR should also look at which sources the answer cites.

5-5. In-depth articles

Which sources AI draws on when it describes the company is covered in What AI Cites When It Describes Your Company, and how long information stays in AI answers after a crisis is covered in How Long Does Information Stay in AI Answers.

6-1. Who the stakeholders are

Legal's stakeholders are business partners, consumers and, if it comes to a dispute, the opposing party. The concern here is people who take an AI answer as fact and make decisions based on it.

6-2. What they research with AI

Within the scope of this review, we could not confirm any primary Japanese figure showing what stakeholders ask AI about companies. Instead, there are figures on how people handle AI answers.

MIC's "White Paper on Information and Communications in Japan, 2026 edition" (July 2026) draws on an online questionnaire carried out from January to February 2026 in Japan, the United States, Germany and China (1,236 people aged 15 to 69 in Japan and 624 in each of the other three countries). It reports what people who use generative AI services take care about in order to use AI appropriately, with multiple answers allowed. In Japan, "I check whether the generated content is correct by other means" was chosen by 38.2%, more than any other option. "Nothing in particular" was chosen by 24.6%. Because multiple answers were allowed, the share of people who do not check cannot be derived from these figures. We could not confirm, on the relevant page of the white paper, the number of generative AI users in Japan who formed the base.

In the same white paper, 34.8% said they feel it is "a serious risk" and 34.5% said they feel it is "somewhat of a risk" that "AI answers to questions are not factual".

What kinds of problems can occur in AI answers? Within the scope of this review, we could not confirm any published figure measuring the rate of problems in descriptions of companies. For news, there is a study led by the European Broadcasting Union (EBU) and the BBC, with 22 public service broadcasters from 18 countries working in 14 languages (published October 2025). It put news questions based on real users' questions to the free versions of ChatGPT, Copilot, Gemini and Perplexity from late May to early June 2025, and journalists evaluated the responses. Of 2,709 responses to the core questions, 45% had at least one significant issue in accuracy, sourcing, distinguishing opinion from fact, editorializing, or context. By criterion, 20% had significant accuracy issues and 31% had significant sourcing issues.

The 45% is not "the share of answers that were factually wrong". It includes problems with sourcing and context. All of these are figures from outside Japan for questions about news, and no Japanese broadcaster took part. It cannot be said that the same rates apply to descriptions of companies. What can be said is that AI answers have other kinds of problems besides factual errors, such as how sources are presented and missing context.

6-3. What can happen when the AI answer is wrong

Return conditions, whether the company holds a license or permit, its representative, or whether it is involved in litigation may be described incorrectly, and a stakeholder may take this as fact and make a decision. If AI's description conflicts with the company's official information, the stakeholder may decide on the basis of a misunderstanding, or the company may receive inquiries or complaints asking it to confirm. What can be done legally when an error affects the company's reputation or credibility is covered in the in-depth articles.

6-4. What legal should look at first

Whether, when the company is asked about by name, AI's description of contract terms, licenses and permits, and basic company information matches the company's official information. Choose the documents to check against according to the type of transaction. For consumer transactions, the terms of use and the notice required under the Act on Specified Commercial Transactions; for business-to-business transactions, the published terms of trade; for industries that require a license or permit, the published license information; and in every case, the company profile. If an answer is wrong, keep a record of the question asked, the date and time, the service used and the original text of the answer, without rewriting it.

6-5. In-depth articles

The pillar article is AI Misinformation and Misattribution: Detect, Correct, Prevent. How to seek corrections is covered in Can You Remove or Correct AI Misinformation?, legal remedies in Can You Sue for AI Misinformation?, and the difference between claims by legal persons and natural persons in Does the Right to Be Forgotten Reach AI Answers.

Setting the five departments side by side shows something they share. When stakeholders make decisions for different purposes, they want to check different things. A candidate is likely to ask "What kind of company is this?", an investor "How are its results?", a buyer "Which service suits [need]?", and a journalist "What has it been doing recently?" These, however, are assumptions. Within the scope of this review, we could not confirm any survey that collected the questions stakeholders actually typed into AI.

AI builds its answer around the question. Google explains that AI Overviews and AI Mode may use different models and techniques, so the responses and links they show differ. Even for the same company, the description can change depending on which service is asked, what question is asked, and how it is phrased.

So the result one department has checked cannot be read as another department's result. Even if PR asks by company name and gets a correct description, whether the name comes up when a buyer asks from a problem without naming the company is a separate question. Even if the business description HR has checked is correct, that does not mean the figures IR asks about are correct.

One more point concerns how well known a company is. A study by Mallen and colleagues (ACL 2023) reports that language models struggle to answer questions about less well-known facts when they rely only on their internal knowledge. The same study also reports that bringing in outside information, such as through retrieval, improves answers to such questions. The study looked at questions about general facts; it did not directly measure descriptions of companies or AI search answers. It cannot be said that the less well known a company is, the more likely AI search is to describe it wrongly. What can be said is that the quality of answers can vary with how well known something is, so that is no reason to skip checking your own company.

The figures so far can also be read the other way. Here, first, is the material that weakens this article's argument.

First, stakeholders do not simply believe AI answers. In the purchasing survey, trust in information compiled by AI ranked last of 10 kinds of information. Of those who used AI in this purchase, 24.9% used AI information almost as is. About 90% of job-hunting students said they check the official website, and 51.5% of individual investors went directly to the IR page. These surveys do not tell us how far AI answers sway decisions.

Second, all of these are self-reports. That stakeholders said they used AI is a different matter from which questions they actually asked and how many times.

Third, many of the surveyors have an interest in the field, as described in §1-2. Within the scope of this review, we could not confirm whether surveys by organizations without such an interest would produce the same results.

Fourth, "have begun to use" does not mean a majority. Whether a figure is a share of all respondents or of AI users differs from survey to survey. And none of the surveys asked whether stakeholders researched a specific company with AI. None of these figures supports saying that most stakeholders research companies with AI.

Even so, there is still a reason to measure. There are buyers who said they used AI to list candidates, and job-hunting students who said they used AI for company research. How much this changed their final choice is not known, but these surveys do show that some stakeholders encounter AI answers at a stage before they check. And the questions stakeholders actually asked, and the answers they saw, cannot be known from the company's side. Even so, if you decide on the questions stakeholders are likely to ask and measure them repeatedly, you can check how AI describes your company under the conditions you set.

9-1. What two ways of asking tell you

Checking by department starts with two ways of asking. What is described here is a general procedure set out in this article. It is not the same as the way our own Vaipm asks (Vaipm's ways of asking are described in §9-7).

A question that uses the name comes from stakeholders who already know your company: a candidate with an interview coming up, an investor considering holding the stock, a journalist who has decided to cover the company. A question that does not use the name comes from stakeholders who do not yet know your company: a candidate looking for companies by industry, an investor looking for candidate stocks, a buyer looking for products from a problem.

9-2. How to build questions for each department

Build the questions for each department in the following order.

1. Settle on one stakeholder. Not "investors", but something like "an individual investor looking for candidate stocks": decide who, and at what stage.
2. Write down, one at a time, what that stakeholder wants to know. The two middle columns of Table 2 are examples.
3. Divide them into questions that use the name and questions that do not. In questions without the name, include the conditions the stakeholder is likely to add (industry, size, region, problem).
4. Decide which official information to check against. This is the right-hand column of Table 2. If you cannot decide what to check a question against, you cannot judge whether the answer is correct.
5. Bring the departments' questions together. Putting the five departments' questions into one table shows both where the same question is being measured twice and where questions are left that nobody is measuring. For each question, write down separately which department owns the question, which owns the documents to check against, and which fixes things when an answer is wrong, so that work does not stall after measurement. For example, the question "What is the dividend policy?" may be owned by IR, the documents by accounting or IR, and the fix by whoever runs the IR site.

9-3. Phrasing sets

Even questions that seek the same information can get different answers when the phrasing changes. One study showing that the form of the input moves the results is by Sclar and colleagues (ICLR 2024). It reports that, in few-shot prompts, formatting differences that do not change the meaning altered accuracy by up to 76 percentage points for a particular model and task. This study, however, examined formatting differences; it did not directly measure natural rephrasing or AI search answers. It is cited not as direct evidence for building sets of rephrasings, but as a reference for the idea that the result of a single input should not be read as the result for the question as a whole.

So, for each question, prepare a set with varied phrasing. For example, "What kind of company is Company X?", "Tell me about Company X's business" and "I want to know about Company X" form one set. Read the results set by set, and also set out the results for each phrasing side by side. Once you have fixed a set for ongoing comparison, do not change it. If you change the set, you can no longer tell whether a difference between earlier and later results comes from a change in the AI or a change in the question. When a new product or policy means you want to add a question, record the new question as a separate set, and present its results apart from those of the existing sets.

9-4. The mention rate and the alignment rate

Measurement results are easier to organize if read through two rates.

The mention rate is the share of AI answers in which your company's name comes up when the question is asked without naming it. For questions asked without the name, count the share of answers, out of the number of times asked, in which your company's name appeared. For example, if you ask 20 times and the name comes up 7 times, that is 7 out of 20, or 35%. Do not count questions that use the name, or questions that list the company by name alongside competitors. If the name is in the question, it is only natural for the name to appear in the answer. For answers in which the name came up, check, where necessary, whether the description matches the official information. Note that When AI Mistakes You for Another Company uses an indicator called the "former-trade-name mention rate". That indicator, specific to that article, looks at the share of answers that use a former trade name in the present tense. It is different from the mention rate described here.

The alignment rate is the share of answers to questions that use the name that matched the official information. The unit of judgment is "one answer". Classify each answer as one of four: matches; ambiguous (weak basis or vague wording); differs (contrary to the facts); or not covered by official information. The alignment rate is the number of "matches" divided by the sum of "matches", "ambiguous" and "differs".

There are three things to decide in advance.

• How to judge an answer that contains several facts. For example, "if even one fact is wrong, the answer counts as 'differs'". Decide and write down this rule before measuring. If you decide after measuring, the criterion can be moved to fit the results.
• Answers about matters not covered by official information are taken out of the denominator, counted separately and reported alongside. If something that can be called neither matching nor differing is counted as "not matching", you can no longer tell whether the answer failed to align or simply could not be judged. If every answer falls into this group, write "cannot be judged", not 0%.
• Do not silently drop answers that failed to identify your company even though it was asked about by name. Answers that said "I don't know" or that confused the company with another one should be counted and reported separately. If they are dropped and only the alignment rate is reported, the failure to identify the company disappears from view.

What the alignment rate measures is "agreement with the official information". It does not in itself show that something is factually correct. If the official information is out of date, answers that agree with it are out of date too. Keeping the documents you check against up to date is a precondition for measuring.

The mention rate is an indicator for questions asked without the name; the alignment rate is an indicator for questions that use the name. The two answer different questions. A high mention rate does not mean the description is correct. A high alignment rate does not mean stakeholders who do not know your name are finding you.

9-5. What to record

Whether you came up, and whether the answer was correct: that alone does not tell you what to do next. Across departments, record the following.

If you drop the last item, you may learn that an answer was wrong but lose the clue to where to look. A displayed source, however, is a clue, not necessarily the cause of the answer. If the source is an old page of your own, you can fix it yourself, but fixing it does not necessarily change the answer. After fixing, measure the same question set again under the same conditions to see whether anything changed.

9-6. Once is not enough: keep measuring

AI answers can change each time the same question is asked. Not coming up once does not prove that you do not come up. Being described correctly once does not prove that you keep being described correctly. Decide on a number of repetitions, repeat, and hold the result as a share.

When repeating, fix the conditions: do not carry over previous exchanges, span multiple AI services, keep the language and region settings the same, keep the use or non-use of search the same, and do not change the phrasing sets from §9-3. If a login state or history remains, you cannot tell whether you are measuring your company or the history of the person doing the measuring.

A share obtained by repetition is an observation under the questions and conditions you set. It is neither the share of all real users for whom your company comes up nor an effect on behavior such as applications or purchases.

And keep measuring. Each department has events that can change how AI describes the company. What follows is a general procedure set out in this article.

Measuring the same set before and after such a point lets you check whether anything changed. But to say that a difference between before and after was caused by that event, you need ways of holding other factors constant. That a difference appeared and that its cause is known are two different things.

9-7. How we measure

Vaipm, operated by our company, Vaigate Inc., measures AI-space perception, for each question, through a total of 25 stateless queries across multiple AI engines. Measurement is monthly. The procedure in §9-6 of measuring again around events is presented as a general procedure, and is separate from Vaipm's monthly measurement. Stateless means measuring without carrying over login state or previous exchanges. This count is our operational design choice. It is not an optimal value derived from research.

Vaipm asks about a subject in four ways: by name; by category, without naming it; alongside competitors; and by the name of a specific initiative. The way a question is asked changes the AI's answer.

The two ways of asking in §9-1 are close in idea to two of these four, but they are not the same. AI answers to each department's questions can be measured with Vaipm by placing those questions within these four ways of asking.

9-8. Questions this measurement does not answer

First, it does not tell you what stakeholders actually asked. What is measured is the answers to questions stakeholders are likely to ask. The questions candidates and investors actually asked cannot be seen from the company's side.

Second, it does not tell you how much the answers affected stakeholders' decisions. What is measured is how your company is described in AI answers. How that connected to applications, investment, orders or articles belongs to other metrics.

Third, the answers are not necessarily the same as those on a stakeholder's screen. Stakeholders may be asking while logged in, with a history. Measuring without carrying over history is a condition for getting comparable results, not a reproduction of a stakeholder's screen.

We consider not overstating what measurement can answer to be part of the design.

Misconception 1: "Most of our stakeholders research companies with AI." None of these figures supports saying that most stakeholders research companies with AI. Whether the figures in Table 1 are shares of all respondents or of AI users differs from survey to survey, and none of the surveys asked whether stakeholders researched a specific company with AI. The journalist figure comes from a global survey and cannot be applied to journalists in Japan either.

Misconception 2: "If AI answers correctly when asked by company name, there is no problem." A question that uses the name checks the description given to stakeholders who already know your company. Whether stakeholders who do not know your name can find you can only be learned from questions that do not use the name.

Misconception 3: "If PR checks, that covers the whole company." Each set of stakeholders implies different questions and different references to check against. As set out in §7, one department's result cannot be read as another department's result.

Misconception 4: "Once checked, we are fine for a while." Answers can change each time a question is asked, and the description can also change at the points listed in Table 5.

Misconception 5: "Figures from outside Japan apply in Japan too." The journalist figure comes from a global survey, and we could not confirm whether Japan was included.

Misconception 6: "Special measures are needed to appear in AI." Google writes that there are no additional requirements or special optimizations needed to appear in AI Overviews or AI Mode. This is an explanation of the AI features in Google Search and cannot simply be extended to other AI services. The core of what each department can do is to publish the official information that serves as the reference, accurately and in a readable form.

HR, IR, sales and marketing, and PR have figures showing that their stakeholders use AI. In Japanese surveys, 37.9% of job-hunting students used AI for company research, 31.8% of individual investors used AI to look for stocks, and 41.6% of corporate buyers felt that AI's influence on purchase decisions had increased; in a global survey, 21% of journalists had never used AI in their work. Legal has no direct figure showing that stakeholders research companies with AI. What can be shown goes no further than a survey of how generative AI users check the answers they get, and the kinds of problems reported in AI answers about news outside Japan. Of generative AI users in Japan, 38.2% said they check whether the answer is correct by other means (multiple answers allowed). The share of business partners or consumers who took an AI answer about a company as fact is not known.

None of these figures supports saying that most stakeholders research companies with AI. Whether a figure is a share of all respondents or of users differs from survey to survey, many of the surveyors have an interest in the field, and figures from Japan and elsewhere are mixed. This article has set them out together with those conditions.

Different stakeholders need different information to make their decisions. So each department should work out separately the questions its stakeholders are likely to ask, and check how AI describes the company. Ask by name and see whether the answer matches the official information. Ask without the name and see whether the name comes up, and alongside whom. Fix the phrasing sets, hold the conditions constant, repeat, and keep measuring at each key point.

We call this practical field of measuring and managing perception in AI space AIPM (AI Perception Management). AI answers to each department's questions can be measured with Vaipm by placing the questions within the four ways of asking, for each question through a total of 25 stateless queries across multiple AI engines.

Q1. Is it true that our stakeholders research companies with AI?

Within the range of the surveys, some stakeholders have begun to add AI to the material they use for decisions. In a Japanese survey of 419 job-hunting students (Nyle, March 2026), 37.9% used AI at the company-research stage. In a Japanese survey of individual investors (Siemple and the Digital Crisis Research Institute, published September 2026), 31.8% of individual investors currently investing in stocks used AI to look for stocks. In a Japanese survey of 1,573 people involved in B2B purchasing (HubSpot Japan, June 2026), 41.6% said the influence of AI on purchase decisions had increased. However, none of the surveys asked whether stakeholders researched a specific company with AI, and none of the figures supports saying that most stakeholders do.

Q2. Can we compare the figures across departments?

Please do not. The form of the question, the base and the way respondents were selected differ from survey to survey. Some figures are shares of all respondents and others are shares limited to AI users; in the purchasing survey, for example, the base for 41.6% is all 1,573 respondents, while the base for 50.3% is 1,155 AI users. Some figures do not state their base at all, and some come from Japan while others come from global surveys. This article sets the figures side by side only to show, condition by condition, how far each department can say that its stakeholders are using AI.

Q3. Are there figures on AI use by institutional investors and analysts?

Within the scope of this review, we could not confirm any Japanese figure. The only IR stakeholders for whom this article could confirm a Japanese figure were individual investors, from the Siemple and Digital Crisis Research Institute survey, and even there the base for 31.8%, the survey period and the wording of the question could not be confirmed from the release. A survey of institutional investors outside Japan is covered in the IR hub article, How AI Describes Your Financials. That survey is outside Japan, so it should not be read as a figure for Japanese institutional investors.

Q4. Do the journalist figures apply to journalists in Japan?

This is not known. Cision's 2026 survey covered 1,899 journalists in North America, Europe/Middle East/Africa and Asia-Pacific, carried out from January to February 2026. The release refers to 19 markets, but within what has been published we could not confirm whether Japan was among them. The bases of its shares, such as the 21% who have never used AI in their work and the 43% who use it for research and fact-checking, could not be confirmed either. The figures therefore do not allow us to say the same of journalists in Japan.

Q5. Do stakeholders trust AI answers?

Not entirely. In a Japanese purchasing survey (HubSpot Japan), trust in information compiled by AI ranked last of 10 kinds of information, at 37.8%. Of the 1,266 respondents who used AI in this purchase, 24.9% used AI information almost as is, while 48.3% checked the facts or added to it before using it. About 90% of job-hunting students said they check the official website, and 51.5% of individual investors went directly to the IR page. On the other hand, some respondents said they used AI to list candidates. How much this changed their final decisions cannot be told from these surveys.

Q6. What is the difference between asking by name and asking without the name?

Asking by name tells you whether AI's description matches the official information, where it is wrong, and what it cites as sources. Asking by conditions without the name tells you whether your company's name comes up, which companies it is listed alongside, and which features it is described by. The first is the question of stakeholders who already know your company, such as a candidate with an interview coming up. The second is the question of stakeholders who do not yet know you, such as a buyer looking for products from a problem. Neither replaces the other, so it helps to use both.

Q7. How do the mention rate and the alignment rate differ?

The mention rate is the share of answers, out of the number of times a question was asked without the name, in which your company's name came up. The alignment rate classifies each answer to a question that uses the name as matches, ambiguous, differs, or not covered by official information, and divides the number of "matches" by the total of "matches", "ambiguous" and "differs". Answers not covered by official information are taken out of the denominator and counted separately, and answers that failed to identify the company are also counted separately. The two rates answer different questions.

Q8. Is checking once enough?

No. AI answers can change each time a question is asked. Not coming up once does not prove that you do not come up, and being described correctly once does not prove that you keep being described correctly. Decide on a number of repetitions and repeat, with the conditions fixed: no carry-over of previous exchanges, the same language and region settings, and the same use or non-use of search. Then measure again with the same question sets at key points such as earnings announcements and new product launches.

Q9. Should each department measure separately, or should the whole company measure together?

A workable approach is to build questions department by department and bring them together into one table. Stakeholders and the references to check against differ by department, but bringing the questions together shows both where the same question is being measured twice and where questions are left that nobody is measuring. For each question, it also helps to write down which department owns the question, which owns the documents to check against, and which fixes things when an answer is wrong, so that work does not stall after measurement.

Q10. Are special measures needed to appear in AI?

Google writes that there are no additional requirements or special optimizations needed to appear in AI Overviews or AI Mode. However, this is an explanation of the AI features in Google Search and cannot simply be extended to other AI services. Google also explains that AI Overviews and AI Mode may use different models and techniques, so the responses and links they show differ. The core of what each department can do is to publish the official information that AI's descriptions are checked against, accurately and in a readable form.

Q11. Is there an in-depth article for sales and marketing?

There is not yet an in-depth article for sales and marketing. How the department should approach the subject is summarized in How Sales & Marketing should approach AI perception. In this article, the sales and marketing section uses a Japanese survey by HubSpot Japan (fieldwork by Macromill, June 2026) of 1,573 people involved in B2B purchasing, and suggests checking whether your company comes up, and alongside whom, when buyers ask about a problem without naming any company, as well as whether the description given by name matches your product and pricing pages.

Q12. Can Vaipm measure each department's questions?

Yes. Vaipm asks in four ways: by name; by category, without naming it; alongside competitors; and by the name of a specific initiative. For each question, it measures through a total of 25 stateless queries across multiple AI engines, and measurement is monthly. Each department's questions are measured by placing them within these four ways of asking. What is measured is how your company is described in AI answers to the questions stakeholders are likely to ask, not what any stakeholder actually asked, and not how much the answers affected their decisions.

Q13. Is there a figure showing that legal's stakeholders research companies with AI?

Within the scope of this review, we could not confirm one. For legal, this article does not have figures of the same kind as for the other four departments. What this article showed are indirect figures: how generative AI users in Japan check answers, from the MIC white paper (38.2% check by other means, multiple answers allowed), and the rates of problems in answers to news questions outside Japan, from the EBU and BBC study. Neither shows what business partners or consumers ask AI about companies.

1. Survey by a business operator. Nyle Inc., 「【就活生のAI利用はもう常識!?】就職活動での生成AI活用実態調査」 ("Is AI now a given for job-hunting students? Survey on the use of generative AI in job hunting"), PR TIMES, April 28, 2026. Survey period March 25–30, 2026; online survey; 419 students graduating in 2026, 2027 or 2028 who have job-hunted or are job-hunting. Use of generative AI to gather information ("often" 34.4%, "sometimes" 41.8%; the release describes this as "more than 70%"); stages at which it was used (preparing for selection 48.0%, company research 37.9%, industry research 33.2% and others; the base of each question is not stated); company information looked up (business description 36.7%, strengths and characteristics 33.5%, differences from competitors 29.8%); checking facts on the official website or careers page (about 90%). Survey by Nyle's "SEO Consultation Room". https://prtimes.jp/main/html/rd/p/000000655.000055900.html
2. Survey by a business operator. Siemple Inc., 「個人投資家の3人に1人が銘柄探しにAIを利用 ── AIが提示するのは上場企業のわずか0.16％」 ("One in three individual investors uses AI to look for stocks: AI presents only 0.16% of listed companies"), September 1, 2026 (a survey with the Digital Crisis Research Institute). Individual investors in their 20s to 60s across Japan, in two stages: pre-screening of 8,000 and a main survey of 260. 31.8% of individual investors currently investing in stocks use AI to look for stocks; 58.9% of those in their 20s; 51.5% went directly to the IR page after AI had presented a company. The base for 31.8%, the survey period and the wording of the question are not given in the body of the release. https://digitalpr.jp/r/142401
3. Survey by a business operator. HubSpot Japan, 「日本のBtoB購買に関する意識・実態調査2026」 ("Survey on attitudes and practices in B2B purchasing in Japan 2026"), press release, July 16, 2026. Fieldwork by Macromill; survey period June 11–16, 2026; online survey; 1,573 employees of companies with 50 or more employees who had been involved in comparing, selecting, requesting approval for, approving or signing off on B2B products or services in the previous 12 months. Influence of generative AI and AI search on purchase decisions increased 41.6%; generative AI and AI search named as an increased information-gathering behavior 33.5%; of 1,155 users, listing candidates 50.3%, organizing each company's features and differences 47.1%, creating comparison tables 39.2%; information compiled by AI ranked last of 10 kinds of trustworthy information at 37.8%; used AI information as is or almost as is 24.9%, used it after checking facts or adding to it 48.3% (1,266 who used AI in this purchase); had progressed at least to listing several candidates when contacting a seller 63.5% (1,431 who contacted a seller); use ChatGPT frequently or sometimes to gather information when considering a purchase 58.0% (base not stated). https://www.hubspot.jp/hubfs/APAC_JAPAN/6.%20press%20release/%E3%80%90%E3%83%97%E3%83%AC%E3%82%B9%E3%83%AA%E3%83%AA%E3%83%BC%E3%82%B9%E3%80%91HubSpot%20Japan%E3%81%8C%E3%80%8C%E6%97%A5%E6%9C%AC%E3%81%AEBtoB%E8%B3%BC%E8%B2%B7%E3%81%AB%E9%96%A2%E3%81%99%E3%82%8B%20%E6%84%8F%E8%AD%98%E3%83%BB%E5%AE%9F%E6%85%8B%E8%AA%BF%E6%9F%BB2026%E3%80%8D%E3%82%92%E7%99%BA%E8%A1%A8.pdf
4. Survey by a business operator. Cision, "PR Emerges as the Primary Source for Journalists in High-Pressure Newsrooms," May 13, 2026. A survey of 1,899 journalists in North America, Europe/Middle East/Africa and Asia-Pacific, January–February 2026 (19 markets according to the text). No list of countries is given. https://www.cision.com/about/press-releases/2026-press-releases/pr-emerges-as-the-primary-source-for-journalists-in-high-pressure-newsrooms-302770936/
5. Commentary by a business operator. Rocky Parker, "Quick Takeaways from the 2026 State of the Media Report," Cision PR Newswire Beyond Bylines, May 27, 2026. "only 21% of respondents say they have not used AI in their work"; reasons for using AI: generating ideas 48%, research and fact-checking 43%, transcription and summarizing 41%. Base not stated. https://mediablog.prnewswire.com/2026/05/27/2026-state-of-the-media-takeaways/
6. Commentary by a business operator. Rocky Parker, "3,000 Journalists Told Us How They're Using AI, Social Media and More," Cision PR Newswire Beyond Bylines, May 14, 2025. In the 2025 journalist survey (more than 3,000 respondents), 53% use generative AI; uses: research 25%, transcription 23%, summarizing 20%. Base not stated. https://mediablog.prnewswire.com/2025/05/14/2025-state-of-the-media-takeaways/
7. Ministry of Internal Affairs and Communications (MIC), 『令和8年版 情報通信白書』 ("White Paper on Information and Communications in Japan, 2026 edition"), Part I, Chapter 1, Section 1, 「個人におけるAI利用の現状」 ("The current state of AI use by individuals"), July 2026. Figure I-1-1-18, 「AIの適切な利用のために気を付けていること（国別）」 ("What people take care about in order to use AI appropriately, by country") (Japan: check whether the generated content is correct by other means 38.2%, nothing in particular 24.6%, multiple answers allowed); Figure I-1-1-16, 「AI利用のリスクに関する考え方（日本）」 ("Views on the risks of using AI, Japan") (AI answers to questions are not factual: a serious risk 34.8%, somewhat of a risk 34.5%). The figures' source is MIC (2026), 「国内外における最新の情報通信技術の研究開発及びデジタル活用の動向に関する調査研究」 ("Research study on trends in the latest ICT research and development and digital utilization in Japan and abroad"). Values checked in the white paper's published data (CSV). Survey outline (online questionnaire, January–February 2026, Japan 1,236 and the United States, Germany and China 624 each, aged 15 to 69) checked in the white paper's notes. https://www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd111420.html
8. European Broadcasting Union, "AI's systemic distortion of news is consistent across languages and territories – international study by public service broadcasters," October 22, 2025, and the report "News Integrity in AI Assistants". 22 public service broadcasters, 18 countries, 14 languages; ChatGPT, Copilot, Gemini and Perplexity; responses generated from late May to early June 2025 and evaluated by journalists. In the report, with 2,709 responses to the core questions as the base: at least one significant issue 45%, significant accuracy issues 20%, significant sourcing issues 31%, context 14%. Japan is not among the participating countries. https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste https://www.ebu.ch/files/live/sites/ebu/files/Publications/MIS/open/EBU-MIS-BBC_News_Integrity_in_AI_Assistants_Report_2025.pdf
9. Google Search Central, "AI features and your website," last updated December 10, 2025. "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary." https://developers.google.com/search/docs/appearance/ai-features
10. Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D., Hajishirzi, H., "When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories," ACL 2023 (arXiv:2212.10511, submitted December 20, 2022, revised July 2, 2023). A dataset of 14,000 questions (PopQA) and 10 models. "LMs struggle with less popular factual knowledge". https://arxiv.org/abs/2212.10511
11. Sclar, M., Choi, Y., Tsvetkov, Y., Suhr, A., "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting," ICLR 2024 (arXiv:2310.11324, submitted October 17, 2023, revised July 1, 2024). Formatting differences that do not change the meaning produced a difference of up to 76 points in the evaluation of LLaMA-2-13B. https://arxiv.org/abs/2310.11324

Sources checked: the sources in this article were checked directly against the primary sources by the research for the Japanese edition on October 11, 2026, and source 3 (the body of HubSpot Japan's press release), source 7 (the white paper's published data and notes) and source 8 (the body of the EBU and BBC report) were opened again and rechecked on October 12, 2026. The 38.2%, 24.6%, 34.8% and 34.5% in source 7 were confirmed in the white paper's published data (CSV). For source 2, the survey period and the base for 31.8%, and for sources 4 to 6, the base of each share, could not be confirmed within what has been published. Within the scope of this review, we could not confirm a Japanese figure for AI use by institutional investors and analysts, a Japanese figure showing what legal's stakeholders ask AI about companies, or a rate of AI errors in descriptions of companies.