What AI Cites When It Describes Your Company — The Reputation Supply Chain for PR

When generative AI describes your company, what appears in the answer as citations is neither the corporate website that communications teams control most tightly, nor the news coverage they have worked hardest to earn. It is a far wider set of sources that contains both — and whose composition differs from one AI service to another.

In Japan, generative AI adoption inside corporate communications departments rose from 37.2% (2024) to 77.0% (2025) across two surveys by the Study Group on Generative AI in Public Relations at the Japan Society for Corporate Communication Studies (121 and 104 valid responses). The uses reported, however, are concentrated in producing, summarizing, and supporting planning.

The supply structure, meanwhile, does not match the assumptions communications work has been built on. McKinsey's analysis estimates that a brand's own website typically accounts for around 5–10% of the sources AI search draws on (a reference figure, as the calculation method is not disclosed). And of the URLs cited in AI Overviews, 37.9% were within the first 10 SERP blocks for the same query (Ahrefs; 863,000 SERPs and 4 million AI Overviews URLs). The cast of cited sources also differed substantially between services within the datasets examined.

Then, in May 2026, AMEC — the international body for communication measurement and evaluation — published the AMEC GEO Principles. This article maps that supply structure from a PR and communications perspective and checks it against AMEC's principles. It is not about responding to errors; it is about where accurate information is drawn from, and how it is handled, in normal times.

The observation scope of this article (important)

What this article treats as measurable is the citation links and supporting links displayed in an AI answer. These are publicly visible clues to what an answer rests on. They are not the full set of sources the model consulted or weighed internally, nor do they show the causal contribution of any individual page. Google uses the term "supporting links" for these in its own official documentation.

AMEC's GEO Principles take the same position: observed AI outputs are directional indicators, to be validated transparently across tools, prompts, markets, languages, and time. The question in this article's title — what AI cites — should be read under that limitation.

Who this is for

Communications and PR departments at operating companies, PR agencies, and heads of corporate communications. Recruitment communications are out of scope (see Candidates research your company with AI).

A note on terminology

This article treats labels such as AIO, GEO, and LLMO as practitioner terms rather than official standards. Google Search states in its official documentation that no additional requirements and no special optimization are needed to appear in generative AI features, and that optimizing for generative AI search is optimizing for the search experience — which is still SEO. For the terms themselves, see What AIO is and What LLMO is.

"AIPM (AI Perception Management)," used in the second half of this article, is likewise not an industry-standard term; it is how Vaipm frames this area. It stands on the same footing as AIO, GEO, and LLMO.

1.1 From 37.2% to 77.0% across two surveys

The Study Group on Generative AI in Public Relations at the Japan Society for Corporate Communication Studies has run a survey on generative AI use in communications departments in two consecutive years. The first survey ran from October 12 to November 11, 2024, covering communications staff and managers at Japanese companies and organizations, practitioners at PR agencies, and researchers in public relations, and drew 128 responses (121 of them valid). Generative AI adoption at that point stood at 37.2% — 44.8% at companies with capital of ¥100 million or more, and 31.6% at those below that threshold.

The second survey was conducted about a year later, from October to November 2025, as a fixed-point study of comparable timing and scale using a partly shared list of targets, with 104 valid responses. Adding "established practice" at 18.3% and "adopted, working toward establishing it" at 58.7% gives an adopted group of 77.0%, while those who had only "heard of it" fell from 45.5% to 13.5%.

Read these numbers carefully. The two surveys are not a panel study following the same respondents. The accurate statement is not "the same organizations changed" but "across two survey waves, the figure rose from 37.2% to 77.0%." Both were internet surveys, and respondents include not only in-house communications staff but also PR agency practitioners and public relations researchers.

1.2 Uses are concentrated in production, summarization, and planning support

Looking at what the tools are used for sharpens the outline. The top uses in the second survey were drafting press releases and similar materials, generating copy and release headlines, and thinking through and generating ideas for plans. Compared with the previous wave, minute-taking rose from 11 responses (24.4%) to 29 (36.3%), and translation from 3 (6.7%) to 14 (17.5%).

Source: Japan Society for Corporate Communication Studies, second survey on the use of generative AI in public relations (October–November 2025, 104 valid responses). Shares use adopters as the denominator. Multiple responses allowed.

What this distribution supports is the fact that respondents' generative AI uses were concentrated in production, summarization, and planning support — and no more than that. The survey did not ask directly whether organizations measure how AI describes them, so the prevalence of AI perception measurement among Japanese communications departments cannot be determined from it.

Writing a press release with generative AI and knowing what generative AI says about your company are different activities. The first is completed inside your own organization. The second happens outside it — and inside several AI services — and remains invisible unless you go and ask. The 77.0% adoption figure shows the spread of the first activity. It says nothing about the second. This article is about the second.

1.3 What this article does not cover

This article deals with the territory that is not error. Where AI states something incorrect about your company, the definitions, categories, remedies, and legal avenues are handled by a separate set of articles.

• The overall picture and a framework for response: AI misinformation about your brand
• Whether it can be challenged legally: Can AI misinformation be challenged legally
• The practice of seeking correction and removal: Correcting and removing AI misinformation

Job seekers researching companies with AI, and the employer-review sources consulted in that context, are covered by Candidates research your company with AI.

What this article covers is the supply structure of reputation in normal times. Even where the information is accurate: where is it drawn from, how is it summarized, and who is it placed alongside. That is the structure examined here.

2.1 About half are using AI search

McKinsey reported results from its AI Discovery Survey in an article published on October 16, 2025, "New front door to the internet." The survey was fielded in August 2025 to a representative US consumer panel, with n=1,927. In it, roughly half of consumers said they deliberately use AI search engines. Use spans all age groups, and more than 70% of users were reported to be asking upper-funnel questions aimed at learning about categories, brands, and products.

The frequently quoted "44%" requires care with its denominator. It is not all consumers but users of AI search, and it is the share who said AI search is their primary and preferred source of information. On the same question, traditional search came to 31%, retailer and brand sites to 9%, and review sites to 6%.

2.2 16% of brands are tracking

In the same article, McKinsey refers to a second survey: a September 2025 study of CMOs at Fortune 500 consumer brands (n≈30), in which 16% said they systematically track their performance in AI search. This is a small survey at n≈30, limited to large US consumer goods brands. It cannot be transferred as-is to Japanese companies, still less to their communications departments.

Note that the figures in this section come from a consulting firm's own research and have not been peer reviewed. This article treats them as directional reference values, not as precise levels.

3.1 An estimate of 5–10%

In the same article, McKinsey states that a brand's own website typically accounts for around 5–10% of the sources AI search draws on. Google AI Overview and the firm's own analysis are given as the basis, and the details of the calculation are not published. It is an estimate, not a precise measurement.

The structure it points to still matters. Where conventional SEO practice tended to center on optimizing your own site, AI search gathers material from a diverse pool of sources that includes affiliate content and user-generated content. That composition shifts with the language model, the region, the category, and the type of question.

3.2 In some industries, external sources exceed 65%

According to the same article, in industries such as consumer goods and financial services, more than 65% of the sources referenced by AI search answers are publishers (magazines and microsites), user-generated content, and affiliate sites. Here too the basis of calculation is undisclosed, so it cannot be treated as a level; but the observation that the supply structure differs substantially by industry points in the same direction as the service-by-service citation data discussed below.

The newsroom, the company profile, the sustainability report. Even if the area communications teams control most directly is in perfect order, its weight among the sources cited in AI answers may be limited. McKinsey itself writes that the strength of a traditional brand is not an indicator of readiness to compete in the world of AI search.

3.3 The technical requirements still stand

To avoid a misreading, a note in the opposite direction.

Google Search's official documentation states that for a page of your own site to appear as a supporting link in generative AI features, that page must be indexed and eligible to appear in Google Search with a snippet. It must also be included in the generative AI features coverage shown in Search Console. These are not additional optimizations; they are long-standing technical requirements.

That said, this is a condition for a page of your own site to appear as a citation link, not a necessary condition for the company itself to be mentioned in an AI answer. It is entirely possible for your company to be described using only third-party coverage, Wikipedia, reviews, or forum posts as material. Even so, unless crawling is permitted and important content is provided as text, your own pages will not enter the pool of citation candidates.

4.1 37.9% of citations come from the first 10 blocks

This is where received wisdom breaks down most sharply. In research published on March 2, 2026, Ahrefs analyzed 863,000 keyword SERPs and a total of 4 million AI Overviews URLs (data from the company's own AI visibility tool). Of the URLs cited in AI Overviews, 37.9% were within the first 10 SERP blocks. A block here means a SERP feature — ads, featured snippets, People Also Ask, video packs and so on — each counted as one block. It is not the same as "page one" or "the top 10 positions." The rest splits almost evenly: 31.2% in blocks 11 to 100, and 31.0% beyond block 100. A second analysis restricted to ordinary blue links gave 37.1% in the top 10, 26.2% in positions 11 to 100, and 36.7% outside the top 100.

The same company reported in July 2025 that around 76% came from the top 10, which means the figure has roughly halved. Ahrefs connects this shift to AI Overviews running on Gemini 3 from January 2026, but that is the company's interpretation, not Google's explanation.

4.2 29.8% of cited domains appear nowhere on page one

Measurements from the academic side point the same way. A study by Haofei Xu, Umar Iqbal, and Jacob M. Montgomery of Washington University in St. Louis (arXiv:2605.14021, submitted May 13, 2026; a preprint, not peer-reviewed) issued 55,393 trending queries across 19 categories to Google over the 40 days from March 13 to April 21, 2026. AI Overviews appeared for 7,583 of them (13.7%): 64.7% for question-form queries and 9.5% for non-question queries.

On sources, the study reports the following. Of the domains referenced by AI Overviews, the overlap with the organic results on the same SERP was 25.0% for the top 5, 41.4% for the top 10, and 70.2% for the whole of page one. Turned around, 29.8% of referenced domains appear nowhere on the corresponding page one. At URL level the figure is 28.5% (17,451 of 61,206). Moreover, the average domain quality score of these "not on page one" references was higher than that of references that did appear on page one.

Again: this is a preprint, and it covers English-language queries issued from the United States. There is no guarantee that behavior in Japanese is the same.

4.3 Citation is spread, not concentrated

The same study contains a figure with even more practical implications for communications. Across 7,583 answers, AI Overviews citations reached 61,212 referenced URLs and 7,479 unique hosts. The median number of references per answer was 8, and the distribution is as follows.

Source: Xu, Iqbal & Montgomery (arXiv:2605.14021, preprint). The page-one aggregation includes roughly 3% Google-service URLs.

Page one of a search result concentrates on a small number of domains; AI Overviews citations spread out more flatly. Hosts cited exactly once during the 40-day observation window numbered 4,212, or 56.3% of all unique hosts. The assumption that watching the top 10 domains will give you the whole picture does not hold.

4.4 Why this happens — query fan-out, per Google's own documentation

The structural reason is set out in Google's own official documentation. Google Search Central explains that AI Overviews and AI Mode may use a technique called "query fan-out": from the original query, the model generates several related queries at once and retrieves additional search results. The example given is a query about fixing a lawn full of weeds producing derived queries such as the best weed killer for lawns and weeding without chemicals. Through this process, Google says, a wider and more varied set of helpful links can be surfaced than in classic web search.

In other words, the URLs shown as citations in an answer are not supplied only by the SERP for the words the user typed. They come from the SERPs of several questions the AI rephrased internally. Your pages, or a competitor's, may be pulled from search results for phrasings your communications team never anticipated.

"Rank first for the main keyword" cannot fully explain how you appear inside an AI answer. That does not make SEO meaningless. Google states officially that generative AI features are rooted in its core Search ranking and quality systems, and that SEO best practices continue to apply. The accurate reading is that the foundation has not changed, and that on top of it a second axis of observation — where answers are being drawn from — has become necessary.

5.1 In Profound's 2024–2025 data, the composition of top cited sources differed by service

Profound, which provides an AI visibility tool, analyzed citation patterns in ChatGPT, Google AI Overviews, and Perplexity from August 2024 to June 2025 (published June 5, 2025; updated August 2025). The dataset is described as being on the scale of 680 million citations. However, while the citation counts are published, the number of independent questions (prompts), the number of companies covered, the language and region composition, and the sampling method are not disclosed. The study presents two different kinds of figure, and they are easy to confuse, so they need to be read separately.

First, share of all citations in the dataset.

Reddit in ChatGPT stands at 1.8%.

Second, relative share within each service's top 10 sources. Here the denominator is the combined citations of the top 10 sources, not all citations.

Source: Profound, "AI Platform Citation Patterns" (August 2024 – June 2025). These are relative shares within the top 10 sources, not shares of all citations.

Reading this as "about half of ChatGPT's citations are Wikipedia" is wrong. The correct statement is that Wikipedia accounts for 47.9% within the top 10 sources, while its share of all citations is 7.8%. When circulating these numbers internally, always attach the denominator.

With that in place: in this vendor's data, for this period and this dataset, the composition of top cited sources differed considerably. ChatGPT leaned toward encyclopedic sources, Perplexity toward community sources, and Google AI Overviews mixed the two. Whether the same composition holds today cannot be determined from this study.

5.2 A different cast appears at the top in AI Mode

On June 20, 2025, SE Ranking studied Google's AI Mode using 10,000 keywords issued from the United States (published August 29, 2025). The data was collected in a logged-out state.

AI Mode answers contained an average of 12.6 links each. The most frequent cited source was www.google.com at 5.7% of all citations — but these were not links to search results: 97.9% of them were links to Google Maps business profiles displayed within the answer. That is 902 of 9,734 answers, equivalent to an answer appearance rate of 9.2%.

Excluding Google, the leaders were www.indeed.com (1.8%), en.wikipedia.org (1.6%), www.reddit.com (1.5%), youtube.com (1.4%), and www.nerdwallet.com (1.2%). Even combined, the top 10 domains accounted for 11.9% of all citations. That information such as a company's location and opening hours reaches the answer through a route that is neither news coverage nor the owned site is a point communications teams easily overlook.

5.3 Google itself writes that AI Overviews and AI Mode differ

The difference is not only a matter of how outside research is interpreted. Google Search Central's official documentation states plainly that because AI Mode and AI Overviews may use different models and technologies, the set of answers and links shown will vary. Even two features from the same company, built on the same search index, are officially described as producing different cited sources. SE Ranking's study likewise found that the overlap of cited URLs for the same keyword averaged 10.7%, and 16% at domain level.

5.4 Ask three times on the same day, and cited URLs overlap by 9.2% on average

The same SE Ranking study contains a figure that cannot be ignored when thinking about measurement.

The same 10,000 keywords were issued three times independently on the same day. For the 9,451 keywords that returned an answer in all three runs, the Jaccard coefficient across the three sets of cited URLs averaged 9.2%, and 14.7% at domain level. Pairwise comparison rises to 18.5–19% for URLs and 26.6–27% for domains, but the overlap remains limited.

The "9.2%" in the previous section is an answer appearance rate; this one is the average Jaccard coefficient across three sets. The numbers are identical and the meanings are not, so take care not to swap them when sharing internally.

By contrast, the cast of sites appearing at the top was relatively stable across all three runs: www.google.com, www.indeed.com, en.wikipedia.org, and www.reddit.com kept much the same order. Individual pages rotate, but there is a consistent tendency in the set of sites that get relied on.

This property shows the danger of drawing conclusions from a single observation. Asking once and not appearing does not establish that you are absent; appearing once does not establish that you are present.

The opposition communications teams have long worked with is earned coverage versus owned channels. Among the URLs shown as citations, a large share falls outside both. Encyclopedias, forums, Q&A sites, video, reviews, business profiles on maps. Treating "AI" as a single thing makes these differences invisible.

6.1 9% of citations in that dataset

An analysis by Kai-Cheng Yang of Binghamton University (arXiv:2507.05301, submitted July 7, 2025; a preprint, not peer-reviewed) draws on data from AI Search Arena, a head-to-head evaluation platform. The base comprises more than 24,000 conversations, more than 65,000 responses, and more than 366,000 citations involving 12 AI search models from three providers: OpenAI, Perplexity, and Google.

Of these, citations referencing news sources accounted for 9%. This is the share within that dataset. It is not the share across AI search as a whole, nor for questions about companies, nor for questions asked in Japanese.

6.2 Different providers rely on different sets of outlets

The study has two main findings. First, models from different providers cite different news sources. Second, citation behavior nonetheless shows shared patterns. News citations are heavily concentrated in a small number of outlets, and low-reliability sources were rarely cited. The study also analyzes the political leanings of cited sources, but that is not the subject of this article and is not covered here. What matters for communications practice is the single point that the set of outlets being relied on differs by provider.

6.3 Earned media matters, but it is not the only route

A piece published on PRSA's PRsay blog on June 18, 2025 argues that visibility comes not from links but from becoming a source. The author, however, is the co-founder and CEO of the PR software company Muck Rack, so this is an argument by an interested party. It is an opinion piece rather than a data paper, and the figures it contains are not used in this article.

Set against the data in the preceding sections, a more measured conclusion follows. News accounted for 9% of citations in that dataset, is concentrated in a small number of outlets, and varies in cast by provider. Earned media is an important route, but it is not the whole of the material from which AI describes your company.

7.1 The AMEC GEO Principles (May 20, 2026, Dublin)

AMEC, the international association for communication measurement and evaluation, published the AMEC GEO Principles together with a practitioner guide, "A Practitioner's Guide to GEO Measurement," at the AMEC Global Summit held in Dublin on May 20, 2026. It is a framework for measuring how organizations are found, interpreted, and represented within AI-led discovery.

The principles were developed over more than six months through collaboration within the AMEC Agency Group, board review, academic validation, feedback from vendors and practitioners, and iterative testing. The lead authors include James Crawford of PR Agency One, Mary Elizabeth Germaine of Ketchum, and Ben Levine of FleishmanHillard TRUE Global Intelligence, with input from AMEC's Academic Advisory Group.

The purpose is stated at the head of the principles: GEO measurement focuses on whether the information stakeholders encounter is accurate, useful, current, reliable, and trustworthy — not simply on increasing visibility.

7.2 The seven principles

AMEC's seven principles are as follows (summarized by this article).

1. AI-led discovery should be measured against communication objectives and stakeholder information needs
2. GEO measurement must assess the upstream information environment before interpreting AI-generated answers
3. Search and content readiness should be evaluated as evidence of whether trustworthy information can be discovered, understood, and cited
4. Observed AI outputs are directional indicators and should be validated transparently across tools, prompts, markets, languages, and time
5. GEO measurement should distinguish visibility from outcomes, connecting AI-led discovery to awareness, trust, action, and impact
6. Sources that are trustworthy and current matter more than volume, promotion, or short-term visibility
7. Ethical GEO improves the public information environment and must not bring about manipulation, disguise, or flooding

7.3 Three evidence domains, and a minimum standard of evidence

The principles are applied across three domains of evidence.

As the minimum standard of evidence for this, AMEC lists the following.

• A managed query library tied to stakeholder questions
• Documentation of the tools, prompts, platforms, dates, markets, and languages used
• Repeatable testing with disclosed variability, and retention of outputs as evidence
• Source quality review and a risk log
• A clear separation of visibility, outcomes, and impact

Ethical guardrails are stated explicitly as well: do not flood the web with low-quality content, do not disguise promotion as independent evidence, do not manipulate reviews or forums, do not treat AI outputs as fact without verification. And it is written that no single score, tool, or set of prompts can prove the whole of AI visibility or of communication impact.

7.4 "Why AI visibility alone is the next AVE"

Alongside the publication, AMEC also released an article titled "How to Measure GEO: Why AI Visibility Alone Is the Next AVE." AVE means advertising value equivalency, a metric AMEC has argued for years should not be used.

In his comment at publication, James Crawford said that while measurement of GEO and LLM outputs is being demanded rapidly, there is also variability in standards, overclaiming, vanity metrics, and methodologies that are not sufficiently transparent, and that the most useful measurement comes from triangulating evidence. AMEC's position is not "do not measure AI visibility." It is "measure it — but do not reduce it to a single score."

7.5 Awareness has spread; ownership and process have not caught up

Practice has not yet caught up with the framework. Muck Rack's "State of PR" report was conducted from May 14 to June 12, 2026, distributed mainly by email to 1,115 PR practitioners, with 971 responses included in the final analysis after excluding low-quality, duplicate, spam, and outlier responses. The wording of questions was updated from previous years, so year-on-year comparison may not apply directly.

Source: Muck Rack, "State of PR" 2026 edition (May 14 – June 12, 2026, 971 valid responses). Muck Rack supplies software for PR, so this is a vendor's own study.

While 73% consider it important, 39% are not measuring at all and 29% have not settled who owns it. The accurate reading is that it is not that a framework is missing, but that awareness has spread while internal ownership and process are not yet in place. Note that the survey period spans May 20, 2026, the day AMEC published its GEO Principles, so whether respondents answered with awareness of the principles is not known.

7.6 This does not mean a single company-wide KPI now exists

An industry body has indeed set out a framework. That does not mean a fixed standard has come into being in the sense that every company uses identical KPIs. AMEC itself cautions against reliance on a single score or tool. What the principles call for is not common numbers but a common discipline.

Barcelona Principles V4.0 (June 2025, Vienna) point the same way: measurement and evaluation should report outputs, outcomes, and impact, and ethics, governance, and transparency in data, methodology, and technology build trust and encourage learning. The GEO Principles sit on that line.

Separately, "The Responsible Use of AI in PR," published by the UK's CIPR in June 2026, translates the Global Alliance's Venice Pledge into guidance for PR practice. Where AMEC's GEO Principles address how to measure the way AI describes you, CIPR's guide addresses the ethics of how PR uses AI. The two are complementary, and keeping them distinct makes both easier to work with.

8.1 Put primary information where machines can read it too

Start with what is certain within what Google states officially. Its documentation lists conventional SEO fundamentals: that crawling is not blocked by robots.txt or at the CDN, that important content is provided in text form, that pages are easy to find through internal links, that structured data, where used, matches the visible text on the page, and that the page offers a good experience.

Translated into communications practice, this becomes: do the company profile, history, business description, officer information, and headline figures exist only inside images or PDFs? Are press releases placed at permanent URLs in a form that can be retrieved as text? Have duplicate pages proliferated? This corresponds to what AMEC's GEO Principles call search and content readiness.

Google states officially that semantic HTML does not need to be perfect, but that using it where possible is generally a good idea, and that it makes pages easier for other users, such as those using screen readers, to parse and navigate. That this points in the same direction as accessibility is a useful argument when persuading colleagues internally.

8.2 How to handle third-party surfaces

Most of the URLs shown as citations are outside your control, but there is a limited amount communications teams can do. Review sites, communities, video, business profiles on maps. Among these, the factual entries that are under your control — the address and opening hours on a business profile, for instance — can be kept accurate. Given that SE Ranking's study found links to Google Maps business profiles appearing in AI Mode answers, this is not work to treat lightly.

Attempting to manipulate what third parties have written, on the other hand, falls under the prohibitions in the next section.

8.3 What not to do

First, manipulating word-of-mouth and reviews. AMEC's GEO Principles state that ethical GEO improves the public information environment and must not bring about manipulation, disguise, or flooding, and the guardrails explicitly include not manipulating reviews or forums and not disguising promotion as independent evidence. Google Search's official guide also lists seeking unnatural "mentions" among the things that should not be expected to work.

Second, in Japan, the stealth marketing rules under the Act against Unjustifiable Premiums and Misleading Representations, administered by Japan's Consumer Affairs Agency. Posts made by employees on their own initiative do not automatically fall under the regulation, but cases are judged comprehensively: the employee's position and duties, the purpose of the post, the company's involvement in deciding the content, and whether compensation or benefits were provided (see Q6 in the FAQ). Consult your legal department or a professional on individual cases.

Third, reliance on AI-specific files. Google Search states officially that you do not need to create new machine-readable files, AI text files, markup, or Markdown to appear in Google Search, including its generative AI features, and that Google Search itself does not use them. It is written that llms.txt may be maintained for other services, but that Google Search ignores it, neither harming nor helping visibility there. No special schema.org structured data is required either, and as for FAQPage, Google ended the display of FAQ rich results on May 7, 2026 (see Q7 in the FAQ).

Fourth, over-trusting a single score. AMEC's GEO Principles state that no single score, tool, or set of prompts can prove the whole of AI visibility or of communication impact. Where a summary metric is used, it needs to be reported alongside its component values by service and by question, together with the measurement conditions. Google itself warns in its official guide that no third-party tool has access to Google's internal ranking or AI systems. What can be observed from outside is output, not internal ranking.

8.4 A checklist for communications teams

1. Is the company's primary information — history, business description, locations, headline figures — at permanent URLs and retrievable as text?
2. Do press releases remain as an archive, and are they crawlable?
3. Is crawling being blocked unintentionally by robots.txt or CDN settings?
4. Where structured data is used, does it match the visible text on the page?
5. Is the factual information on third-party surfaces you can control, such as a Google Business Profile, up to date?
6. Have you built and maintained a list of anticipated questions tied to stakeholder questions?
7. Are you actually putting those questions to several AI services, and recording and retaining the answers and cited sources?
8. When recording, are you documenting the tool, prompt, platform, date, market, and language used?
9. When sharing numbers internally, do you attach the denominator and the measurement conditions, and distinguish visibility from outcomes?
10. Is there an agreed process for what happens when the information contains an error?

Items 6 to 9 correspond to AMEC's minimum standard of evidence: a managed query library, documentation of conditions, retention of outputs, and the separation of visibility from outcomes. There are five angles for anticipated questions: discovery questions that do not include your company name ("who are the major companies in this industry"), confirmation questions that do, comparison questions that place you alongside competitors, reputation questions that ask about impressions, and questions touching on issues specific to your company, such as past incidents or disputes.

Note that in the Xu et al. study cited earlier, the 64.7% activation rate of AI Overviews for question-form queries is a figure for the whole surveyed population, not the result of extracting questions about corporate reputation. Whether the same gap holds for corporate reputation questions cannot be confirmed from this study alone. Even so, there is value in anticipating questions phrased as sentences — that is this article's inference.

Once you actually measure, you will encounter statements that differ from fact. What is then required is a different set of judgments. What counts as an "error," what means exist to seek correction or removal, and whether it can be contested legally. The following articles cover these.

• The overall picture and a framework for response: AI misinformation about your brand
• Legal liability and the means to contest it: Can AI misinformation be challenged legally
• The practice of correction and removal: Correcting and removing AI misinformation

One addition. The Xu et al. study cited earlier (a preprint) decomposed AI Overviews answers into 98,020 atomic claims and reports that 11.0% were not supported by the page cited ("omission" 7.0%, "error" 2.7%, "ambiguous" 1.4%). Even where the cited source is a reliable outlet, the sentence assembled from it does not necessarily match what that outlet wrote. That AMEC's principles hold that AI outputs must not be treated as fact without verification answers to this structure.

10.1 A single score alone cannot evaluate the whole

Set out from the standpoint of measurement, what this article has covered comes to three things. First, the composition of cited sources differs markedly between services — and even between AI Overviews and AI Mode inside Google, the overlap of cited URLs averaged 10.7%. Second, the same service varies from run to run: for the same keywords issued three times on the same day, the Jaccard coefficient across the sets of cited URLs averaged 9.2%. Third, because the tail of citation is wide and flat, monitoring a small number of domains will not give you the whole picture.

It follows that a single score alone cannot evaluate the whole. Using a summary metric is not itself ruled out, but where one is used, it must be reported alongside its component values by service and by question, together with the measurement conditions. This is consistent with what AMEC's GEO Principles require: repeatable testing with disclosed variability, and the separation of visibility, outcomes, and impact.

10.2 The conditions worth recording

If you intend to share measurements internally and compare them over time, record at least the following. It is close to identical to AMEC's minimum standard of evidence.

• Which AI service (name and mode), when (date and time), and from which language and market
• Whether the result was affected by login state or by prior conversation history
• How many times you asked, and in how many of those your company appeared
• How you were handled within the answer (mentioned only, mentioned with a citation link, recommended, or placed unfavorably in a comparison)
• What was shown as the cited source

Over the long run, being able to explain how a number was obtained matters more than the number itself.

Note also that even where the movement of an observed figure lines up in time with the execution of a communications initiative, that alone does not establish a causal relationship. Cited sources also move for reasons unrelated to your activity. Keep correlation and causation separate in how you write things up.

10.3 AIPM as a way of framing this

Understanding on a continuing basis how your company appears inside AI, and managing that appearance, is work that cuts across SEO, media analysis, and reputation management. Vaipm calls this area "AI Perception Management" (AIPM). As noted earlier, it is not an industry-standard term; it is how Vaipm frames this area. The concept is covered in detail in What AIPM is.

For communications teams, this layer is not exotic. Understanding how the organization is seen by society, and working on that perception, is what PR has always done. What changed is that the layer mediating how you are seen now contains several AI services that behave differently from one another. The first step is to understand the supply structure.

Q1. Is there a way to make AI speak well of my company?

No method that directly guarantees favorable description can be confirmed from publicly available primary sources. Google Search states officially that no additional requirements and no special optimization are needed to appear in generative AI features, and that conventional SEO best practices apply; it also states plainly that seeking unnatural mentions across the web is less useful than it appears. AMEC's GEO Principles likewise hold that sources that are trustworthy and current matter more than volume or promotion. What is realistically available to you is to provide your own primary information accurately and in a retrievable form, to check on a continuing basis how you are actually being described, and to act where the description departs from fact.

Q2. Is a well-developed corporate website enough?

Probably not, is this article's conclusion. McKinsey's analysis estimates that a company's own website typically accounts for around 5–10% of the sources AI search draws on (a reference figure, as the calculation method is not disclosed). That does not mean your own site is unnecessary. For one of your pages to appear as a supporting link in generative AI features, that page must be indexed and eligible to appear with a snippet. This is a condition for a page of your own site to be cited, not a condition for the company itself to be mentioned in an AI answer.

Q3. If I rank first in search, will AI cite me?

Not necessarily. In Ahrefs' research (863,000 SERPs, 4 million AI Overviews URLs), 37.9% of the URLs cited in AI Overviews were within the first 10 blocks of the search results, and 31.0% were beyond block 100. A preprint measurement study likewise found that 29.8% of referenced domains do not appear on page one for the same query. The reason is query fan-out, which Google explains officially: AI also gathers material from the search results of several questions derived from the original query.

Q4. Are the tactics the same for ChatGPT and Google?

In Profound's analysis (August 2024 – June 2025), Wikipedia accounted for 47.9% within ChatGPT's top 10 sources, while Google AI Overviews was led by Reddit (21.0%), YouTube (18.8%), and Quora (14.3%), and Perplexity by Reddit at 46.7%. All of these are relative shares within the top 10 sources, not shares of all citations. The number of prompts and the sampling method are not disclosed, and the data period is limited. Google itself states officially that AI Overviews and AI Mode may use different models and technologies, so the answers and links shown will vary.

Q5. Will more press coverage improve my reputation inside AI?

News is one important route, but not the only one. A preprint analysis found that 9% of citations referenced news sources within that AI Search Arena dataset. This is not the share across AI search as a whole, nor for questions about companies. News citations are also reported to be concentrated in a small number of outlets and to differ in cast by provider. Equating the effort to earn coverage with a whole strategy for AI visibility drops other routes — communities, video, reviews, encyclopedias — out of view.

Q6. Is it effective to have employees post on social media?

Posts made by employees on their own initiative do not automatically fall under the regulation. In Japan, however, the stealth marketing rules under the Act against Unjustifiable Premiums and Misleading Representations, administered by the Consumer Affairs Agency, apply, and cases are judged comprehensively: the employee's position and duties, the purpose of the post, whether the company was involved in deciding the content, and whether compensation or benefits were provided. Where a post amounts to a representation by the business, indicating affiliation alone is not enough; it must be made clear to ordinary consumers that the post is advertising or was made on request. Consult your legal department or a professional on individual cases.

Q7. If I put up llms.txt or FAQ structured data, will AI pick me up?

For Google Search this is officially denied. Google Search states plainly that you do not need to create new machine-readable files or AI text files to appear in Google Search, including its generative AI features, and that Google Search itself does not use them. It is written that llms.txt may be maintained for other services, but that Google Search ignores it, neither harming nor helping visibility there. Nor is any special schema.org structured data required for generative AI search. As for FAQPage, Google ended the display of FAQ rich results on May 7, 2026, so it cannot be expected to produce display treatment (the FAQ content itself still has value for readers). Where structured data is used, it must match the visible text on the page.

Q8. If my company does not appear when I ask once, does that mean it is not appearing?

It does not. In SE Ranking's study, the same 10,000 keywords were issued three times on the same day, and for the 9,451 keywords that returned an answer in all three runs, the Jaccard coefficient across the sets of cited URLs averaged 9.2%, and 14.7% at domain level. A single observation tells you only what was selected on that occasion. AMEC's GEO Principles likewise list repeatable testing with disclosed variability among the minimum standards of evidence.

Q9. Is there an industry framework for AI visibility?

There is. AMEC published the AMEC GEO Principles and A Practitioner's Guide to GEO Measurement in Dublin on May 20, 2026. They set out three evidence domains — upstream reputation signals, search and content readiness, and downstream AI outputs — together with minimum standards of evidence such as a managed query library and documentation of measurement conditions. It is not, however, a fixed standard in the sense that every company uses identical KPIs. AMEC itself states that no single score, tool, or set of prompts can prove the whole of AI visibility.

Q10. Isn't this visible in Search Console now?

Search Console's performance report for generative AI features has been rolling out to some sites since it was announced on June 3, 2026. As of August 2026 it is still not available on every site. On sites where it is available, you can see performance in the generative AI features of Google Search and Discover. What you can see, however, is limited to Google's own surfaces: ChatGPT, Perplexity, Claude, and the Gemini app are not included, and it does not show how your company was treated within an answer.

Sources were verified on August 4, 2026. Figures around AI search change quickly, so please check the current state with the publisher when citing them.

Primary and official documentation

1. AMEC, "AMEC GEO Principles" (PDF; published at the AMEC Global Summit in Dublin, May 20, 2026)
   https://amecorg.com/wp-content/uploads/2026/05/AMEC-GEO-Principles.pdf
2. AMEC, "A Practitioner's Guide to GEO Measurement" (PDF, May 2026)
   https://amecorg.com/wp-content/uploads/2026/05/AMEC-Practitioners-Guide-to-GEO-Measurement.pdf
3. AMEC, "AMEC launches GEO Principles to bring rigour to AI-led communications measurement" (May 20, 2026)
   https://amecorg.com/2026/05/amec-launches-geo-principles-to-bring-rigour-to-ai-led-communications-measurement/
4. AMEC, "The AMEC GEO Hub" (hub page collecting the principles, practitioner guide, and data quality principles)
   https://amecorg.com/amec-geo-hub/
5. AMEC, "Barcelona Principles 4.0" (announced at the AMEC Global Summit in Vienna, June 2025)
   https://amecorg.com/resources/barcelona-principles-4-0/
6. Google Search Central, "AI features and your website" (last updated December 10, 2025)
   https://developers.google.com/search/docs/appearance/ai-features
7. Google Search Central, "Optimizing your website for generative AI features on Google Search" (last updated July 10, 2026)
   https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
8. Japan Society for Corporate Communication Studies, Study Group on Generative AI in Public Relations — generative AI adoption in communications departments at 37.2% (survey period October 12 – November 11, 2024; 128 responses, 121 valid; internet survey)
   https://www.jsccs.jp/activity/report/ai372.html
9. Japan Society for Corporate Communication Studies, Study Group on Generative AI in Public Relations — second survey on the use of generative AI in public relations, with adoption in communications departments at 77.0% (survey period October – November 2025; 104 valid responses; internet survey)
   https://www.jsccs.jp/activity/report/ai77-20252-ai.html
10. CIPR, "AI in PR guides" / The Responsible Use of AI in PR (June 2026; translates the Global Alliance's Venice Pledge for practitioners)
    https://cipr.co.uk/CIPR/CIPR/Our_work/Policy/AI_in_PR_/AI_in_PR_guides.aspx

Academic and measurement research (all preprints, not peer-reviewed)

11. Haofei Xu, Umar Iqbal, Jacob M. Montgomery, "Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact," arXiv:2605.14021 (submitted May 13, 2026; Washington University in St. Louis; 55,393 trending queries, 19 categories, 40 days [March 13 – April 21, 2026], 98,020 atomic claims)
    https://arxiv.org/abs/2605.14021
12. Kai-Cheng Yang, "News Source Citing Patterns in AI Search Systems," arXiv:2507.05301 (submitted July 7, 2025; Binghamton University; AI Search Arena data, more than 24,000 conversations, more than 65,000 responses, more than 366,000 citations)
    https://arxiv.org/abs/2507.05301

Vendor and consulting studies (reference values; note the interested-party position)

13. McKinsey & Company, "New front door to the internet: Winning in the age of AI search" (October 16, 2025. AI Discovery Survey: US consumer panel, August 2025, n=1,927 / CMO survey: CMOs at Fortune 500 consumer brands, September 2025, n≈30. Figures such as 5–10% are estimates whose calculation method is not disclosed)
    https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/new-front-door-to-the-internet-winning-in-the-age-of-ai-search
14. Ahrefs, "Update: 38% of AI Overview Citations Pull From The Top 10" (published March 2, 2026; updated May 31, 2026. 863,000 keyword SERPs, 4 million AI Overviews URLs. Analysis of the company's own data from its AI visibility tool)
    https://ahrefs.com/blog/ai-overview-citations-top-10/
15. Profound, "AI Platform Citation Patterns" (published June 5, 2025; updated August 2025. August 2024 – June 2025, on the scale of 680 million citations. Number of prompts, number of companies, language and region composition, and sampling method are not disclosed. Analysis of an AI visibility tool vendor's own data)
    https://www.tryprofound.com/blog/ai-platform-citation-patterns
16. SE Ranking, "AI Mode research: Volatility, source patterns, and differences from AIO and organic results" (published August 29, 2025. Data collected June 20, 2025; 10,000 keywords issued from the United States; logged-out state. Overlap measured by Jaccard coefficient. Analysis of an SEO tool vendor's own data)
    https://seranking.com/blog/ai-mode-research/
17. Muck Rack, "State of PR" 2026 edition (survey period May 14 – June 12, 2026; distributed to 1,115 PR practitioners; 971 valid responses. A study by a vendor supplying software for PR) / reported by PRSA PRsay
    https://prsay.prsa.org/2026/07/15/long-hours-fewer-reporter-responses-ai-optimization-2026-state-of-pr/
18. Gregory Galant, "LLMs Just Made PR the New Power Player in Search," PRSA PRsay (June 18, 2025. The author is co-founder and CEO of Muck Rack. An opinion piece, not a data paper)
    https://prsay.prsa.org/2025/06/18/llms-just-made-pr-the-new-power-player-in-search/

What this article treats as measurable is the citation links and supporting links displayed in an AI answer. They are not the full set of sources the model consulted or weighed internally, nor do they show the causal contribution of any individual page.

Among the data cited, vendor studies are identified in the body as coming from interested parties, and academic work is identified as peer-reviewed or not. Surveys with a small base (n) and estimates whose calculation method is not published are treated as directional reference values rather than as levels. The summaries of AMEC's GEO Principles are this article's own; for the original text, see the PDFs in the source list.

The situation around AI search changes quickly, and the figures in this article are as of the verification date. For decisions in practice, please check the latest information with the publisher. For matters involving law and regulation, consult your legal department or a professional.

Updated: August 4, 2026 / Sources verified: August 4, 2026

By Vaipm (which measures AI-space perception through a total of 25 stateless queries across multiple AI engines)