Shuichiro Ogawa
日本語

Notes · updated 2026-10-11

Industry Trends in AIO (Optimization for AI Search): What Is Measured, What Is Sold, and What Is Contested

This note examines AIO, the work of getting a company or its information cited and recommended in AI search and generative AI answers (including GEO, AEO, and LLMO).

Contents (10)
  1. The names AIO, GEO, AEO, and LLMO
  2. How much did AI summaries reduce clicks?
  3. The scale of referrals from AI to websites
  4. What gets cited?
  5. The evidence for citation-friendly writing
  6. The productization of citation measurement
  7. Control over, and payment for, AI use of content
  8. Growing use and policy debate in Japan
  9. What academic research adds
  10. Where this account does not apply

The names AIO, GEO, AEO, and LLMO

The work of making a company’s name or pages appear in AI answers has at least four names: AIO (AI Optimization), GEO (Generative Engine Optimization), AEO (Answer Engine Optimization), and LLMO (LLM Optimization). GEO was coined by a paper released in 2023 and accepted at KDD 2024, which used the term for a framework that increases the visibility of a creator’s content within generative AI answers (Aggarwal et al. 2024). In Japan, advertising companies use these names for their research units. CyberAgent set up a “GEO Lab” in July 2025, and Hakuhodo DY ONE runs the “ONE-AIO Lab”, which glosses AIO as “AI optimization” (CyberAgent 2026; Hakuhodo DY ONE 2026).

Despite the different names, the object is the same: being cited, recommended, and mentioned in AI answers. The providers treat it in the same way. A Bing product manager at Microsoft wrote, “Whether you call it GEO, AIO, or SEO, one thing hasn’t changed: visibility is everything” (Madhavan 2025). Google’s official guide names AEO and GEO and states that, from Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, “and thus still SEO” (Google Search Central 2026a).

However, treating this work only as an extension of SEO misses two industry developments: measurement that counts citations, and the choice of whether to let AI use content at all. Between 2025 and 2026, these two developments moved further than writing techniques did.

How much did AI summaries reduce clicks?

All sources agree on the direction: clicks fall when an AI summary appears. The most direct measurement comes from Pew Research Center. In the March 2025 browsing records of 900 U.S. adults, 12,593 of 68,879 Google searches produced an AI summary. Users clicked a traditional search result link in 8% of visits to searches with an AI summary, about half of the 15% for searches without one. They clicked a link inside the AI summary in 1% of visits to pages with a summary (Pew Research Center 2025). The share of searches that produce an AI summary is also not stable. In Semrush’s tracking of more than 10 million keywords, the share of keywords triggering an AI summary rose from 6.49% in January 2025 to nearly 25% in July and then fell to 15.69% in November (Semrush 2025a).

Measurements by SEO tool companies estimate a larger decline. Ahrefs compared the click-through rate of the top-ranking page for 300,000 keywords. For keywords with an AI summary, the rate was 34.5% lower than the value predicted from the trend of informational keywords without one (March 2024 compared with March 2025; Ahrefs 2025a). An update using the same method for December 2023 and December 2025 found a 58% gap (Ahrefs 2026a).

These figures need careful reading. In the Ahrefs update, the position-one click-through rate for informational keywords without an AI summary also fell by about half, from 0.076 to 0.039 (Ahrefs 2026a). Seer Interactive tracked 3,119 informational queries across 42 client organizations. In the third quarter of 2025, organic click-through rates fell year over year by 65.2% for queries with an AI summary that did not cite the client, by 49.4% for those that did, and by 46.2% even for queries without an AI summary (Seer Interactive 2025). Factors other than AI summaries therefore also contribute to the decline, and the whole decline cannot be attributed to AI summaries. The Ahrefs figures of 34.5% and 58% are the additional gap after removing this shared decline, whereas the Seer figures are values before such removal.

Separating the shared decline from the effect of AI summaries requires varying the presence of AI summaries at random. Agarwal and Sen randomly assigned users with a Chrome extension either to the standard Google Search with AI summaries or to a version with the summaries removed. According to the abstract, when an AI summary appeared, outbound organic clicks fell by 38% and the probability of a zero-click search rose by 33%, while sponsored clicks and search frequency did not change (Agarwal and Sen 2026). The paper is a working paper that has not been peer reviewed; its full text could not be obtained, so this account stays within the abstract.

Google presents a different view. In August 2025, Liz Reid, Head of Google Search, wrote that total organic click volume from Google Search to websites had been “relatively stable” year over year and that Google was sending “slightly more” quality clicks, meaning clicks after which users do not quickly return (Reid 2025). No figures or calculation methods were given. Reid also acknowledged that, even with stable totals, traffic is shifting between sites.

The scale of referrals from AI to websites

AI does send people to websites in place of the lost clicks, but the volume is still small. Similarweb estimated that AI platforms generated more than 1.1 billion referral visits in June 2025, up 357% year over year. The same release states that about 95% of ChatGPT users still use Google and that all figures are estimates based on third-party sources (Similarweb 2025). Adobe reported that traffic from generative AI tools to U.S. retail sites in the 2025 holiday season rose 693.4% year over year, while noting that the base of users remains modest (Adobe 2026). In BCG’s consumer survey across nine countries, shopping-related use of generative AI grew by 35% from February to November 2025 (Boston Consulting Group 2026).

Surveys of users also show that few people move from AI answers to the original sources. Six in ten U.S. adults say they read AI summaries in search results (Pew Research Center 2026). In Bain’s December 2024 survey (n = 1,117), about 80% of consumers said they rely on zero-click results in at least 40% of their searches (Bain & Company 2025). In the Reuters Institute’s 2026 survey across 27 markets, 4% of all respondents said they always or often click through from AI to underlying news sources, compared with 19% for search and 17% for social media (Reuters Institute 2026). This gap also reflects the smaller number of people who use AI for news in the first place. In a 2025 survey of six countries including Japan, 54% had seen an AI-generated answer to a search in the past week, and 33% of them said they always or often click the links in the summary (Simon et al. 2025). Both reports note that self-reported behavior may differ from actual behavior.

The ratio of crawling to referrals shows the same imbalance from the side of websites. In Cloudflare’s observations of its own network in July 2025, the number of crawls per referral was 38,065.7 for Anthropic, 1,091.4 for OpenAI, 194.8 for Perplexity, and 5.4 for Google (Cloudflare 2025b). Cloudflare acknowledges that referrals from native apps may lack referrer information, so the ratios may be overstated (Cloudflare 2025a).

Some claim that AI referrals are few but valuable. Semrush stated that a visitor from a non-Google AI search source such as ChatGPT is 4.4 times as valuable as an average visitor from traditional organic search (Semrush 2025b). The calculation is said to be based on conversion rates, but the sample and period are not given.

What gets cited?

Many AIO techniques assume that being cited brings a benefit. However, what gets cited differs from the results of traditional SEO. In Ahrefs’s analysis of 15,000 prompts, on average only 12% of the links cited by ChatGPT, Gemini, and Copilot appeared in Google’s top 10 results for the same prompt. Perplexity was the exception, with nearly one in three citations overlapping the top 10 (Ahrefs 2025b). For Google’s AI summaries, 37.9% of cited URLs also appeared in the top 10 blocks of the same search, and Ahrefs describes this as a drop from about 76% in its July 2025 study (Ahrefs 2026b). Because the detection method changed between the two studies, the size of this drop should be discounted.

Citations concentrate on a few sites. In Ahrefs’s June 2025 data, Wikipedia accounted for 16.3% of mentions in ChatGPT, YouTube for 16.1% in Perplexity, and YouTube 9.5%, Wikipedia 8.4%, and Reddit 7.4% in Google’s AI summaries (Ahrefs 2025c). Profound analyzed about 730,000 ChatGPT conversations with at least one web citation from U.S. English-language users (October to December 2025). Wikipedia accounted for 5% of all citations and appeared in 18% of conversations with citations, and Reddit accounted for 3% and 13%, respectively (Profound 2026).

This concentration is not fixed. Semrush checked more than 230,000 prompts weekly for 13 weeks and found that the share of ChatGPT responses citing Reddit fell from close to 60% in early August 2025 to about 10% by mid-September (Semrush 2025c). SEO practitioners attributed the change to a change in Google Search, but a Semrush specialist stated that this was unlikely to be the only cause.

Recommendations themselves do not reproduce. SparkToro had about 600 volunteers enter 12 prompts into ChatGPT, Claude, and Google’s AI a combined 2,961 times. The probability of receiving the same list of brands for the same prompt was below 1 in 100, and the same order appeared about once in 1,000 runs (Fishkin 2026). Fishkin also wrote that the frequency with which a brand appears across many runs carries some statistical validity. He disclosed that his co-researcher had joined a company that tracks AI visibility.

Being cited and receiving visitors therefore need to be treated as separate quantities. In Pew’s measurement, the link inside an AI summary was clicked in 1% of visits to pages with a summary (Pew Research Center 2025). In Seer’s study, queries that cited the client had a higher organic click-through rate (0.70%) than queries that did not (0.52%), but Seer notes that it cannot prove that citation causes higher click-through (Seer Interactive 2025).

The evidence for citation-friendly writing

At the center of AIO advice is a way of writing that is likely to be cited. One starting point is the GEO paper. In an experimental setup where GPT-3.5 generated answers from the top five Google results, adding statistics, adding quotations, and citing sources increased visibility. The best methods improved on the baseline by up to 40%, and by up to 37% on Perplexity, whereas keyword stuffing from traditional SEO did not work (Aggarwal et al. 2024).

Later evaluation did not support this effect. C-SEO Bench tested conversational search optimization methods, including the GEO methods, on product recommendation and question answering in several domains with recent LLMs. Only 3 of 54 conditions showed a significant ranking improvement, and many methods had no effect or even lowered the ranking. What worked was the traditional SEO factor of ranking high in search and being placed early in the LLM context. In addition, overall gains decreased as more parties adopted the methods, which indicates a congested, zero-sum situation (Puerto et al. 2025).

Manipulation that goes beyond writing has been reported to work on deployed systems. Adding a strategic text sequence to a product description raised the likelihood of becoming the LLM’s top recommendation (Kumar and Lakkaraju 2024, not peer-reviewed). Prompt injection promoted low-ranked products, and the attack transferred to Perplexity (Pfrommer et al. 2024). Attackers could make Bing and Perplexity promote their products, and when every party attacks, answer quality falls for everyone, producing a prisoner’s dilemma (Nestaas et al. 2025). Google’s statement that mass-producing pages primarily to manipulate generative AI responses violates its scaled content abuse policy follows the same line (Google Search Central 2026a).

The providers’ official positions reject many techniques that spread in the industry. On May 15, 2026, Google added a guide on optimizing for generative AI features, including a section that corrects common “AEO/GEO” misconceptions (Google Search Central 2026b). The guide states that files such as llms.txt neither harm nor help visibility in Google Search because Google Search ignores them (Google Search Central 2026a). It also states that there is no requirement to break content into tiny pieces for AI and that seeking inauthentic mentions across the web is not as helpful as it might seem. Another Google page states that there are no additional requirements or special optimizations for appearing in AI Overviews or AI Mode, and that no new machine-readable files or markup are needed (Google Search Central 2026c). Microsoft states that there is no secret strategy for being selected by AI systems, while recommending headings, lists, tables, Q&A blocks, and structured data (Madhavan 2025). It gives no figures showing an effect.

llms.txt is often discussed in the industry as an AIO technique, but the proposal has a different purpose. Jeremy Howard’s proposal of September 2024 collects, in one file, information that helps agents use a website (Howard 2026). The proposal notes that OpenAI, Anthropic, and Gemini publish llms.txt files for their own developer documentation, but it does not state that their AI search products read llms.txt to build answers.

The productization of citation measurement

While the evidence for writing techniques has weakened, measurement has grown. Google’s Search Console now has a generative AI performance report that shows how many times links to a site were shown in AI Overviews and AI Mode (Google 2026a). In February 2026, Bing released a feature that shows how often a site’s pages are cited in Copilot and Bing’s AI summaries, noting that the count reflects citation frequency, not page importance or ranking (Bing Webmaster Team 2026a). In June, Bing added the share of citations that a site holds for a given query and described it as an observational metric, not a ranking system or a competitive scoreboard (Bing Webmaster Team 2026b). Both tools count impressions or citations and show no effect on clicks or sales.

Funding has flowed to measurement companies. Profound, which measures AI visibility, announced a USD 35 million round led by Sequoia Capital in August 2025 (Profound 2025), and Peec AI announced a USD 21 million round in November 2025 (Peec AI 2025). SEO tool companies have also added AI visibility tracking to their products, and many of the studies used in this note are in-house studies by companies with a commercial interest in selling such tools.

Measurement figures inherit the instability of citation sources and the low reproducibility of recommendations. Because the result of a single query does not reproduce, measurement can only be read as frequencies over many repetitions (Fishkin 2026). The providers’ own reports also stop at impressions and citations and do not show what people did after reading an AI answer.

Control over, and payment for, AI use of content

Alongside measurement, mechanisms for choosing whether to let AI use content have developed. AI providers have separated crawling for search from crawling for training. OpenAI allows sites to block OAI-SearchBot, which surfaces sites in ChatGPT search, independently of GPTBot, which is used for training, and states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers (OpenAI 2026). Anthropic separates Claude-SearchBot, and Perplexity separates PerplexityBot, as crawlers for search (Anthropic 2026; Perplexity 2026).

For Google’s AI summaries, the means of refusing use separately from ordinary search became a point of dispute. In an April 2026 statement, the Japan Newspaper Publishers & Editors Association argued that crawler controls for ordinary search and AI search are identical, so that refusing use in AI search also removes content from ordinary search. On this basis, it pointed to a suspected abuse of a superior bargaining position (Japan Newspaper Publishers & Editors Association 2026).

The UK Competition and Markets Authority (CMA) designated Google as having strategic market status in general search services on October 10, 2025, and decided on a conduct requirement for publishers on June 3, 2026. The requirement obliges Google to provide effective controls that let publishers withhold their content from search generative AI features such as AI Overviews and AI Mode and from broader generative AI services such as Gemini. It also requires transparency, including metrics on user engagement, and clear and accurate attribution of content, with a nine-month implementation period (CMA 2026a; CMA 2026b). Google’s Search Console help states that, as of August 31, 2026, a control for excluding a site from AI Overviews, AI Mode, and generative AI features in Discover has been rolled out to all websites worldwide (Google 2026b). On July 23, 2026, the European Commission fined Google a total of EUR 890 million for breaches of the Digital Markets Act (self-preferencing in Search and steering restrictions on Google Play). The Commission stated that dialogue would continue on Google’s proposals for applying the principles of the decision to AI Overviews and AI Mode (European Commission 2026).

Standardization and payment mechanisms are developing in parallel. An IETF working group aims to standardize building blocks for expressing preferences about how content is used for AI (IETF n.d.). Its September 2026 draft separates training (train-ai), AI use (ai-use), and search (search), and states that the text does not yet have consensus (Keller and Thomson 2026). A companion draft that conveys these preferences through an HTTP header field and robots.txt is written by authors from Google and Mozilla (Illyes and Thomson 2026). The original robots.txt standard states that its rules are not a form of access authorization (Koster et al. 2022), so whether a stated preference is honored is left to the crawler. In July 2025, Cloudflare began pay per crawl, which returns HTTP 402 to AI crawlers to request payment, as a private beta (Cloudflare 2025c). In September 2026, it moved to a beta of a model in which buyers offer a price for each use of content and self-report each use (Cloudflare 2026).

What publishers and regulators are contesting is less the number of citations than the choice of whether to let content be used in exchange for citations, and the payment for that use.

Growing use and policy debate in Japan

Japanese surveys show growth in the use of AI for search. In a CyberAgent survey (9,278 people in their teens to sixties nationwide, February 2026), 37.0% said they use generative AI for search, up from 21.3% in May 2025 and 31.1% in October 2025. In the same survey, 21.0% used AI Mode in Google Search (CyberAgent 2026). In a Hakuhodo DY ONE survey (2,800 people aged 18 to 69 nationwide, March 2026), 72.9% of AI search users said their use had increased over the past year; about half of search users use AI search routinely, while nearly 80% still use web search (Hakuhodo DY ONE 2026). Both surveys were conducted by advertising companies that sell AIO or GEO services, and neither gives the definition of AI search users or the denominators of the percentages.

On the policy side, in December 2025 the Japan Fair Trade Commission announced a follow-up to its fact-finding survey on news content distribution, citing as one reason the concern that generative AI search and AI summaries use news content without permission to generate answers. At the press conference, the Secretary General also said that this in itself does not necessarily lead directly to an Antimonopoly Act issue (Japan Fair Trade Commission 2025). In July 2026, the Commission began a questionnaire of about 370 news content providers (Japan Fair Trade Commission 2026). The results had not been published at the time of this note.

What academic research adds

Academic research supplements two points that industry sources do not measure. The first is whether citations are correct. A 2023 audit of four generative search engines found that, on average, only 51.5% of generated sentences were fully supported by citations and only 74.5% of citations supported their associated sentences (Liu et al. 2023). In a preregistered randomized experiment with 4,927 participants representative of the U.S. adult population, people trusted generative AI search less than traditional search on average, but reference links and citations increased trust even when they were incorrect or hallucinated (Li and Aral 2025, not peer-reviewed). AIO, which aims at being cited, therefore relies on a mechanism in which citations generate trust, while the correctness of those citations is not guaranteed.

The second is the direction of bias in citation sources. For consumer product queries, AI search was reported to cite third-party reviews and media far more than Google Search did, and brand-owned sites and social media less (Chen et al. 2025, not peer-reviewed). A study comparing 11,000 real search queries across five systems reported that AI summaries overrepresent Wikipedia and longer documents among the cited sources and underrepresent social media and negatively framed sources (Huang et al. 2026, not peer-reviewed; listed as accepted at EMNLP 2026). Both studies suggest that being written about by third parties may matter more for citation than rewriting a company’s own site, but, as with industry tallies of citation sources, this depends on the type of query examined.

Where this account does not apply

Many of the click-through and citation figures come from in-house studies by SEO tool and measurement companies with a commercial interest. This note used only the factual parts with stated methods, but the selection of keywords depends on each company’s clients and products. The Pew study, which measured actual browsing records, is limited to one month of Google searches in the United States. A 2026 field experiment that randomly hid AI summaries was used only within its abstract because its full text could not be obtained; it covers Google Search by users who installed a Chrome extension. Gartner’s forecast of search volume and McKinsey’s consumer survey on AI search were not used because their publisher pages could not be reached. U.S. publishers’ lawsuits against Google were not used because the texts of the judgments and complaints could not be reached. No survey with a representative sample measuring the adoption of AIO in Japan was found.

Unverified items

No unverified statements remain in the main claims. Sources that could not be reached (Gartner’s search volume forecast, McKinsey’s consumer survey on AI search, Chartbeat’s publisher traffic report, the complaints and judgments in the Penske Media and Chegg lawsuits against Google, the remedies judgment in United States v. Google, and the text of the European Commission’s December 2025 announcement of an investigation) are recorded in the ledger with the routes attempted.

References

Author: Shuichiro Ogawa (Design Researcher / Consultant) About me →