SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.

Recall Strategy in the AI Era: Is it true that "once AI becomes widespread, the importance of awareness and recall will decrease"?

2026年5月26日 07:17

Recently, marketing media, conferences, and even internal company discussions have been dominated by generative AI. While there is a lot of talk about how "Claude is amazing" or "this job will be taken by AI," some people are asking, "If AI becomes widespread, it will provide options, compare them in tables, and even make decisions for us, so won't the importance of awareness and recall, as we've known them, decrease?"

Therefore, in this article, I will organize how the mechanisms of recall, CEPs (Category Entry Points), and mental availability (MA) are being expanded by the emergence of AI as a "mediator."To start with my personal conclusion: the structure and importance of recall itself do not change. The only thing that changes is that there is now one more entity doing the recalling. However, that one point is large enough to shift the center of gravity of entire operations, depending on the category.

Since around 2025, when generative AI began to spread explosively, another layer of "mediator" has entered the space between consumers and purchasing. If a consumer asks, "Tell me about a hot spring inn where I can go with my family," ChatGPT or Gemini will organize the candidates. If they ask, "What kind of insurance should I get for my child's first policy?" an AI agent will narrow down the options and provide an answer. "Rufus" is now resident in the Amazon search bar, and "AI Overviews" line the top of Google search results.

I wouldn't go so far as to say we are moving from recall in the mind to recall on the screen, butit is certain that a mediator called AI has created a new layer between "human minds" and "purchasing behavior"The map of recall is beginning to expand its contours from a map that was completed solely within the human mind to a map where both humans and AI perform recall.

Here, I will clarify the relationship between the two terms that will appear repeatedly in this article: "generative AI" and "AI agents." Generative AI refers to the foundational technology itself that generates text and images, represented by ChatGPT, Gemini, and Claude. On the other hand, AI agents refer to applications built on that generative AI foundation that answer questions on behalf of the user, organize candidates, and sometimes even execute purchases or reservations.

The entity that a consumer is talking to when they say, "Tell me about a hot spring inn where I can go with my family," is strictly speaking not the generative AI itself, but an AI agent (a chat assistant, shopping assistant, etc.) that incorporates generative AI. In this article, I will use "generative AI" or "LLM" when emphasizing the foundational technology aspect, and "AI agent" when emphasizing its role as a mediator.

1. AI has entered the "layer of recall"—the tectonic shift as of early summer 2026

First, let's confirm what is happening now with numbers. However, please note that the following figures are as of early summer 2026. Since the AI landscape changes every 3 to 6 months, please keep the timing in mind.

ChatGPT's weekly active users have more than doubled in one year, from approximately 400 million in February 2025 to approximately 900 million in February 2026, with an annualized revenue (ARR) of $10 billion and over 50 million paid subscribers [1]. Google's Gemini app has also surged from approximately 450 million monthly active users in early 2025 to approximately 750 million in early 2026 [1]. Amazon's AI shopping assistant "Rufus" has been deployed to 300 million customers, and Amazon itself explains that the service has contributed approximately $12 billion in incremental annualized revenue [2]. Google's AI Overviews serves 2 billion users monthly across more than 200 countries, and consumer surveys have begun to show that over a quarter of respondents say they "hardly used Google search result pages during their most recent product purchase" [3]. According to Adobe Analytics, during the 2024 holiday season (November-December), retail site visits via generative AI grew significantly compared to the same period the previous year (a temporary figure reported as "+1,200%" in their report) [3].

Japan is no exception. According to the 4th "Survey on Consumer Awareness Regarding AI" conducted by Dentsu in November 2025 (targeting 3,000 people aged 15-69), the digitalization of the purchasing process continues to progress, with the usage ratio of digital touchpoints increasing from the previous year at every stage: +4% in the awareness phase, +6.4% in the comparison/consideration phase, and +3% in the purchase phase [4]. The usage experience rate of generative AI has doubled in just one year, from 15.6% in June 2024 to 30.3% in June 2025. The ChatGPT usage rate among teenagers is 87.7%, and 85.9% for those in their 20s. The entry point for information contact is definitely leaking from search engines to conversational AI [4].

The market forecast side cannot be ignored either. Gartner predicts that by 2026, search volume for traditional search engines will decrease by 25%, with that volume flowing to generative AI chatbots [5]. Looking at B2B, Gartner estimates that by 2028, organic search traffic for B2B companies will decrease by approximately 50%, and the main information gathering channels for decision-makers will shift to large language models [5]. The Edelman-LinkedIn "2025 B2B Thought Leadership Impact Report" (survey of 1,934 US B2B decision-makers / March-April 2025) also highlights the existence of "hidden buyers" who have a strong influence on purchasing decisions but are difficult to visualize; 63% of them engage with thought leadership content for more than one hour per week, and 81% responded that "high-quality thought leadership highlights issues we were not aware of" [6].

As of early summer 2026, AI has already begun to be incorporated into the "layer of recall." And it is doing so at the most upstream part of people's consideration processes—at the stage where they haven't even decided whether to buy yet.

2. The degree of impact varies by category—thinking in a three-layer structure

However, it is not advisable to rush into "AI adaptation" for all categories. This is because the degree of impact of AI mediation differs greatly by category. This article recommends organizing the degree of impact by dividing it into the following three layers.

Layer 1: Convenience goods (Impact: Small)

Chocolate, canned coffee, toilet paper, detergent, gum. These are convenience goods (low-involvement, repeat-purchase categories) dominated by System 1 (fast thinking). CEPs are tied to instantaneous, physical situations like "I'm thirsty now" or "I want to snack on something before the meeting." Consumers have neither the time nor the motivation to consult AI. In front of the convenience store shelf, their hands move within 3 seconds, and the product is bought.

In this layer, I predict that the room for AI intervention is limited. Rather, what is important are the classic winning moves I have discussed repeatedly in this note: the breadth and depth of CEPs, and physical availability. "Even in the AI era, convenience goods are a battle between the shelf and the mind"—this is the first conclusion.

Layer 2: Shopping goods and specialty goods (Impact: Medium to Large)

Home appliances, cosmetics, hotels, restaurants, clothing, gadgets, subscription services. Before purchasing, consumers compare multiple candidates, check prices and reputations, and usually take a consideration period of several days to several weeks. Information asymmetry is moderate. CEPs create contexts where situations and conditions are complexly intertwined, such as "a dryer I want to replace during the rainy season" or "a hot spring inn where I can take a 3-year-old child."

In this layer, the value of AI agents as "organizers" stands out clearly. If you ask ChatGPT, "a laptop under 200,000 yen that is strong for video editing," it will return a candidate list and a summary of reviews at the same time. If you ask Amazon Rufus, "candidates for a 3-year-old's birthday present," it will filter for age-appropriateness and line up several options.What consumers are handing over to AI is not a search query, but an "extremely contextual situational statement."One could even say that they are verbalizing the CEPs themselves and throwing them at the AI.

There are two strategic implications for the second layer. First, whether or not your company is included in an AI agent's response will directly determine entry into and exit from the recall set.Second, the maintenance of primary information sources that AI references—official websites, reviews, professional media articles, IR materials, and industry white papers—will begin to carry the same level of importance as traditional advertising investments.

Third Layer: High-Involvement/Information Asymmetry Products (Impact: Maximal)

Housing, automobiles, life insurance, non-life insurance, medical care, education, financial products, and large-scale BtoB implementations. The consideration period ranges from several weeks to several years. There are multiple decision-makers and many stakeholders involved. Information asymmetry is extremely high, and consumers and buyers are forced to walk through a jungle of technical jargon and complex conditions in these categories.

This is the layer that benefits most from AI agents. 'What is the difference between loan types when buying a house?' 'Which is better for a dual-income couple in their 30s with children, term insurance or whole life insurance?' 'What are the commonly overlooked items when buying a used condominium?'—AI agents provide structured answers to these questions in seconds. In the BtoB space, we have entered an era where questions like 'Major options for cloud ERP' or 'MES comparison for mid-sized manufacturers' are being thrown at ChatGPT and Gemini.

In the third layer, AI intervention goes beyond being an 'information organizer' and significantly influences the narrowing down of purchase candidates itself. The map of recall changes shape here. It becomes necessary to capture both the top-of-mind recall in the consumer's head and the top of the AI agent's response list.

In the Japanese market, the most clear demonstration of this third-layer movement is the real estate brokerage LIFULL. The company launched the ChatGPT plugin version of 'LIFULL HOME'S' in 2023, and in 2024, it jointly developed 'AI ANSWER Plus (beta version),' a real estate transaction consultation AI utilizing generative AI, with Nomura Real Estate Solutions [7]. Furthermore, in March 2025, it launched its internal generative AI infrastructure 'keelai,' and in December of the same year, it released the conversational property proposal agent 'AI Homes-kun,' and has announced the 'LIFULL AI' concept, an integrated agent across all services, for 2026 [7]. In the category of housing, which has the longest consideration period and the highest information asymmetry, AI is beginning to take over everything from 'extracting conditions -> suggesting candidates -> generating comparison tables.'

In the insurance industry, Sompo Japan Insurance introduced the generative AI chat 'MS-Assistant' for customer and accident response in 2023, and in a 2025 Accenture survey, 74% of insurance customers answered that they 'have no resistance to talking to an AI chatbot for routine transactions.' This ratio has risen significantly from 43% in 2021 [8]. In BtoB as well, Gartner predicts that organic search traffic for BtoB companies will decrease by 50% across the industry by 2028 [5].

Immediate-need products like convenience goods rarely pass through AI. Shopping goods have an increased probability of passing through AI. High-involvement/information-asymmetry products are moving in a direction where the consideration process cannot function without AI. This is the assessment as of early summer 2026.

Supplement: The 'Possibility' that the Landscape for Convenience Goods Will Also Change Around 2030

However, I will add the caveat that this is the landscape as of early summer 2026. The world of convenience goods also has the potential to be repainted around 2030. When major e-commerce platforms like Amazon, AI agents, and IoT appliances (smart refrigerators, replenishment agents, healthcare-linked meal suggestions) connect and reach a certain level of penetration as an 'automated replenishment purchasing platform,' some purchasing behavior for convenience goods may shift from in front of the shelf to 'initial settings for AI.' People who fix their brand preference with 'I always use this mayonnaise' and people who completely delegate to AI with 'I'm not picky about seasonings, so just send me whatever is cheap and highly rated at the time'—consumer settings may become polarized. The problem is that when 'left to the AI' with the latter setting, what the AI ultimately chooses is either the top-of-mind/high-recall brand, the incumbent brand with accumulated review volume and ratings, or the product that has a coupon issued and is the lowest price. Brands that have not achieved top-of-mind or high-recall status by 2026 should be careful, as even if they try to make a comeback with end-cap displays or discounts in stores after entering the 2030s, that 'moment at the store' itself may have shrunk.

Convenience goods currently rarely pass through AI—that is likely correct as of 2026. But as a long-term trend, convenience goods may be heading toward a future where the FMOT (First Moment Of Truth) itself, which is directly linked to purchasing rather than 'choices' or 'comparative consideration,' shifts to AI.

Why Do First-Mover Advantages and Concentration of Power Easily Occur in the Second and Third Layers?

Let's take it a step further. Why is it that in the second and third layers, first-mover advantages and concentration of power easily occur when mediated by AI? The principle lies on both sides of AI's 'learning method' and 'answering method.'

The Learning Side—LLM foundation models are trained on massive amounts of text from news, books, and professional media, centered on Common Crawl. Brands that have accumulated articles, reviews, papers, and citations over the years have a first-mover advantage in 'mention volume,' 'contextual consistency,' and 'resolution of co-occurring words' within the training data. History and volume are transcribed directly into the machine's memory. This is not a gap that latecomers can close in a short period.

The Answering Side—AI agents are under design pressure to avoid disappointing users by making uncertain choices. Recent studies have reported that output tendencies similar to the ambiguity aversion presented by Ellsberg in 1961 and the status quo bias mentioned in behavioral economics are also observed in LLM responses [16]. Major company reassurance factors such as 'X years of track record,' 'X0,000 companies introduced,' and 'industry standard' are easily optimized as reasons for recommendation. Furthermore, because summarization models have the property of choosing the 'most consistent explanation' in their output, they have a structure where incumbent players with converging co-occurring words are picked up more advantageously than emerging players whose co-occurring words are dispersed.

First-movers have the advantage in both learning and answering—this dynamic acts repeatedly in the second and third layers. On the other hand, the first layer (convenience goods) rarely passes through AI in the first place, so these dynamics are almost irrelevant. There are category differences in the fact that 'concentration occurs due to AI' itself.—First, let's make sure we don't misjudge this point.

So, is there no path to victory for emerging and small-to-medium players? To conclude first, there is. If you fight head-on with major companies using 'category words' (e.g., 'hair dryer,' 'microwave,' 'laptop,' 'real estate brokerage'), a frontal breakthrough with big words will be extremely disadvantageous for latecomers, but building a local battle in a unique category defined by your company or specific context statements (CEPs)—there is a path to victory within this asymmetric strategy. Designing CEPs is effective as an asymmetric strategy precisely in the AI era. This will be discussed in detail in Section 5.

3. What Does It Mean to Have AI 'Remember'?

Here, we return to the central concept of this paper. Top-of-mind recall is the brand that is remembered first the moment a certain situation (CEPs) arises. When adding an AI agent as a mediator, the definition is expanded as follows.

Top-of-Mind Recall (Expanded Version): The brand that comes to mind first in a given situation, either in the consumer's head or at the top of the response list of the AI agent the consumer uses.

Consumers' minds are formed as they have always been: by advertising, PR, in-store experiences, word-of-mouth, and personal experience. On the other hand, the 'mind' of an AI agent is made up of training data, search indexes, and evaluation logic embedded by vendors. Getting an AI to 'recall' your brand is nothing less than ensuring your company's information remains in this AI's 'mind' in a reliable form.

Ayelet Israeli and colleagues at Harvard Business School have shown that when using LLMs for marketing research, brands with a higher volume of product reviews and primary information lead to more stable LLM response quality [9]. Reisenbichler et al. demonstrated in the Journal of Marketing Science the conditions under which product content using Natural Language Generation (NLG) influences search rankings and conversions [10]. Meanwhile, several independent studies have reported that when LLMs make brand recommendations, they exhibit cognitive biases such as anchoring and decoy effects, and that answers fluctuate significantly even when the same question is repeated [11].

The practical implications drawn from this can be organized into three points.

AI reads 'source material' The brands that an LLM includes in its responses depend heavily on how much information about that brand has been accumulated in its training data and search index, and the quality of that information. Reviews, professional articles, IR materials, industry white papers, official FAQs, and news releases are more effective as 'recall material' for AI than advertising creatives.
AI searches using 'situational sentences' AI agent queries are moving away from the keywords of the search engine era toward natural language 'situational sentences.' The task of organizing CEPs as situational sentences becomes the gateway to 'AI Recall Optimization' (sometimes called GEO = Generative Engine Optimization or AEO = Answer Engine Optimization) in the AI era.
AI fluctuates LLM recommendations have instability. It is not uncommon for the same question to be asked twice and receive different answers. That is precisely why 'AI Recall Monitoring,' which measures multiple times across multiple AIs to grasp which brands are structurally more likely to surface, is being added as a new KPI.

There is one more distinction I would like to emphasize in this article. When we talk about getting an AI to 'recall' something, we actually need to discuss two separate layers. This is because the speed of execution and the way of fighting are different.

Learning Data Layer (slow layer)—This is the set of texts that an LLM ingests when training its foundation model. It includes Wikipedia articles, news reports, industry white papers, cited books, and accumulated articles from professional media.Whether your company exists here in the 'correct context' and in 'sufficient quantity' is a battle that takes years. It will not move with a one-off press release.Long-term PR, raising the quality and quantity of media coverage, disseminating primary information worthy of being posted on Wikipedia, contributing to industry white papers, and publishing cited primary research—these layers build the assets of the learning data layer.

Inference-Time Layer (fast layer)—This is the layer of primary information that the AI refers to via web search or RAG (Retrieval-Augmented Generation) the moment a consumer asks a question. It includes the latest information added after the foundation model was trained, structured product pages on your company's site, the latest IR, reviews, FAQs, and third-party evaluation content.This can be influenced by structuring (machine-readable heading design, FAQ format, explicit descriptions of uniqueness) and update frequency. This is a battle fought on a monthly or quarterly basis.

The learning data layer requires time, while the inference-time layer requires structure and update frequency. The former is the greatest barrier for incumbents, while the latter is the closest thing to a counter-attack point for newcomers.When implementing AI recall monitoring, I recommend measuring these two layers separately. Even with the same 'brand appearance rate,' looking at the learning data layer and the inference-time layer separately reveals what you should invest in and at what speed.

In conclusion, while there are many factors, the work of getting an AI to 'recall' your brand is an extension of long-standing PR, reputation management, content marketing, and IR; it is not an entirely new job. What is different is that the reader is not only human but also a machine, and that machine can only know the brand through training data and search indexes.

4. Reinterpreting Mental Availability through a dual structure of humans and intermediaries

Is the concept of mental availability presented by Sharp and Romaniuk still valid in the AI era? Regarding this, Romaniuk herself continues to speak out. In an interview, she stated, 'AI accelerates the verification of creative, the personalization of messages, and the prediction of consumer behavior. But for AI to truly move brand value, a powerful interpretive framework is needed. Mental availability will be that framework' [12]. The Ehrenberg-Bass Institute is also conducting research using AI models to extract key CEPs from consumer stories [12].

This article's position aligns with this. The definition of mental availability remains unshaken even in the AI era. The only difference is that the 'subject that recalls' now includes not only humans but also AI agents. Mental availability can be reinterpreted as a dual structure: 'recall probability in the human mind' and 'probability of appearing at the top of the AI agent's response list.'

Mental Availability (Human): The probability of your company being recalled via CEPs, recall ranking, and appearance rate in the consideration set
Mental Availability (Intermediary): The probability that an AI agent includes your company in its response to a situational sentence, appearance ranking in the response list, and response stability

The same applies to the physical availability side. In addition to physical shelves and sales floors, e-commerce paths, and the ease of getting internal approval, a layer is added: 'whether purchase links or inquiry paths function within the candidate list presented by the AI agent.' Amazon Rufus and Google's Shopping AI are beginning to embed purchase links directly into their responses, and here too, the classic question of 'can they buy it after being recalled?' is repeated in a new space.

Wertenbroch et al. theorized the tension between consumer autonomy and algorithmic mediation in the 2020 issue of Marketing Letters [13]. Previous research by André, Carmon, Wertenbroch, and others points to the trade-off that while AI-driven automation of choice enhances consumer well-being, it may also impair the sense of autonomy [14]. The optimal point for how much decision-making the intermediary should take on differs by category. For convenience goods, full automation (replenishment purchasing) is acceptable, but for high-involvement products, 'AI as an organizer' is preferred, with the final decision left to the human. The three-layer structure presented in Section 2 is valid here as well.

AI's answer is not the 'final answer'—it works by multiplying with human recall ranking

Here, I would like to emphasize a point that is easily overlooked. I do not believe that just because being recalled by AI is becoming the new main battlefield, the importance of recall in the human mind—the ranking of aided recall, unaided recall, and top-of-mind recall—is decreasing. On the contrary, I suspect that under AI mediation, a mechanism where human recall ranking has an even greater effect may be at work.

Let's look at the mechanism step by step. Suppose a consumer asks an AI a question using a situational sentence: "I want to replace my dryer before the rainy season starts. I have a family of five and I'm concerned about electricity costs." Suppose the AI presents five brands as candidates. At this point, the following processing occurs in the consumer's brain.

First, screening for aided recall. Among the five presented, the consumer instantly sorts the brands they recognize (brands they can recall with aid) from those they have never seen (brands they cannot recall with aid). The latter (brands not aided-recalled = unknown brands) are at a significant disadvantage at that point, even if recommended by AI. The risk aversion of "it's scary to buy an unknown brand" does not disappear even with an AI recommendation.

Second, cross-referencing with unaided recall and top-of-mind recall. Suppose that among the few brands that were aided-recalled, the AI's recommendation list includes brands that the consumer had previously unaided-recalled as "When it comes to dryers, it's X" or brands that hold the top-of-mind position. At this point, confirmation bias and the affect heuristic work strongly. "I knew this was the best choice (it matches my top-of-mind recall)" and "If the AI reaches the same conclusion, then this must be correct"—the "endorsement effect" from the AI acts to reinforce the consumer's existing recall.

Third, convergence of choice. The consumer chooses the brand that matches their unaided recall or top-of-mind recall almost reflexively. I suspect that the five brands in the AI's recommendation list are rarely compared "fairly".

In other words, purchasing choices mediated by AI are generally determined by the following three-layer filtering.

Layer 0: Is it included in the AI's response list (5–10 brands) in the first place?
Layer 1: Among the response list, is it aided-recalled in the consumer's mind?
Layer 2: Among the response list, is it positioned high in the consumer's unaided recall or top-of-mind recall?

Getting the AI to "remind" the consumer is the battle of Layer 0. The investments in the training data layer and the inference-time layer discussed in the first half of this paper are effective here. However, Layers 1 and 2 remain a "battle of recall in the human mind." I believe that the work of building up a consumer's aided recall, unaided recall, and top-of-mind recall for each CEP does not decrease in the AI era—no, it is even more important because of the AI era.

The implications of this mechanism are by no means small. Under AI mediation, brands that satisfy both "being recalled by AI" and "being high-recall brands in the human mind" have the potential to be overwhelmingly advantageous. AI recall and human recall work not by addition, but by multiplication. It is difficult to be chosen with only one. Even if you concentrate resources on AI recall monitoring, if the recall ranking in the consumer's mind is low, you will drop out during the selection process after being recommended by the AI. Conversely, even if you have captured the top-of-mind position in the consumer's mind, if you do not enter the AI's response list, you will not even be on the table for consideration.

Practically, this means that AI countermeasures and human recall countermeasures require operational design for "both," not "either/or." AI recall monitoring is added as a new metric, but traditional metrics such as human aided recall rate, unaided recall rate, and top-of-mind recall rate will not be abolished. Rather, the economic significance of the work to raise human unaided recall rankings may increase precisely to leverage confirmation bias after AI mediation.

5. Designing CEPs and CEPs is actually strengthened in the AI era

I see the design of CEPs and CEPs as being in a direction that is strengthened, rather than weakened, in the AI era. The reason is that queries to AI agents are becoming closer to "highly contextual natural language situational sentences."

With traditional search engines, consumers entered fragmented words like "dryer recommendation" or "insurance comparison." Even if CEPs were designed, most of them were worn down in the process of being translated into search queries. On the other hand, people often throw situations directly at AI agents, such as "I want a dryer to replace before the rainy season, I have a family of five, and I'm concerned about electricity costs." Here, the design of "situational sentences"—5W1H + Feeling—is reflected directly in the input information in the AI era.

This change is a tailwind for marketers. If consumers verbalize CEPs as they are, the resolution of the entry point that a company should address increases. The question is whether a company can incorporate the task of confirming "whether the company is included in the AI's response to this situational sentence" as a new KPI.

Implementation is surprisingly simple. Take the "entry list (with prioritized situational sentences)" created in advance and throw it directly into major AI agents (ChatGPT, Gemini, Google AI Overviews, Claude, etc.), and monitor the company's presence, appearance ranking, co-occurring competitors, and cited primary sources in the response list every month. This is the AI version of search ranking monitoring.

AI utilization in real estate, such as LIFULL HOME'S, and in the recruitment field is already moving in the direction of establishing their own intermediaries [7]. The issue of whether the intermediary is a third-party platform or a proprietary platform provided by the company will also create a major divergence in the next few years. However, this issue is a separate discussion about how to design the intermediary itself, and since it exceeds the scope of this paper, it is omitted here. For this paper, I recommend first adding "whether the company is recalled by third-party platform AI agents" to the standard KPIs for CEP operations.

I would like to state one more core point of this section. As shown in Section 2, AI mediation structurally tends to create first-mover advantages and concentration of power among the strong. The core of an asymmetric strategy against this is the design of CEPs. If you fight head-on with major players using "category words"—general terms like "hair dryer," "microwave," "laptop," or "real estate brokerage"—the outcome is likely to be decided by the difference in the volume of mentions and citations accumulated in the training data layer, making a frontal breakthrough by a latecomer extremely disadvantageous.

However, if you narrow your focus to proprietary categories defined by your company or specific situational sentences (CEPs), you can weaken the historical advantage of major players on that battlefield. This is because no one has yet poured a large amount of learning material into new words just defined by your company or questions tied to specific contexts. The design of target CEPs and the verbalization of situational sentences will become the primary weapons for emerging and small-to-medium players in the AI era. Instead of a frontal breakthrough with BIG words, construct a local battle with contextual CEPs defined by your company, and go after "both the training data layer and the inference-time layer" in that narrow battlefield. This is the prescription for the AI era in this paper.

6. How to evaluate the "BtoA(AI)(toC)" discourse

Recently, the term "BtoA" has begun to circulate in the industry. Following BtoB and BtoC, the argument is that companies no longer reach consumers directly, but rather through AI. How should we evaluate this discourse? I will present my perspective in this article through three questions.

Question 1: Is B2A2C structurally new?

Yes. In the sense that one more intermediary has been added, this is a structural addition. In the B2C era, intermediaries were distribution channels (retailers, wholesalers, trading companies) and advertising media (TV, newspapers, magazines, search engines, social media). In the B2A era, "AI agents" are added to this. Sometimes a single AI handles both the mediation of commercial flow and the mediation of information (like Amazon Rufus).

Question 2: Does B2A2C overturn the theory of top-of-mind recall?

No. The structure of top-of-mind recall—"becoming the entity that is thought of first in a given situation"—remains unchanged. What changes are the number of entities doing the recalling (humans + AI) and the places where recall occurs (in the mind + in response lists). The structure of "CEPs → Recall → Consideration → Purchase" now has an additional layer of intermediaries inserted. The place where it is inserted may change, but the structure will not break.

Question 3: Should all companies implement B2A2C now?

It depends on the category. In terms of the three-layer structure in Section 2, the third layer (high-involvement, information-asymmetric products) has already entered the implementation phase as of 2026. The second layer (shopping goods/specialty goods) is beginning to enter the measurement phase. The first layer (convenience goods) can still afford to wait and see. While the term "B2A" is powerful, if you are pressured by its influence to invest without regard for category characteristics, you will end up with a distorted allocation: over-investment in convenience goods and under-investment in high-involvement products.

The "95-5 Rule," presented by Dawes et al. in collaboration with LinkedIn's B2B Institute and the Ehrenberg-Bass Institute, provides insight here as well [15]. In B2B, only about 5% of the total are actually entering the purchase consideration phase, while the remaining 95% are inactive. Being "remembered" by AI becomes a means to continuously re-expose one's company through intermediaries, not only to active considerers but also to the 95% who will eventually enter the consideration phase. AI agents cannot be ignored as a channel to reach "hidden buyers"—stakeholders who are strongly involved in purchase decision-making but are difficult to visualize—as shown in the 2025 Edelman-LinkedIn survey [6]. Humans are effective for short-term considerers, while intermediaries are effective for long-term inactive users and hidden decision-makers. The essence of B2A may lie right here.

In summary, B2A is a convenient term, but its essence is simply that "an intermediary has been added to the subject of mental availability." I believe that the map of recall remains effective even in the AI era—or rather, precisely because it is the AI era.

7. Summary of this article

I will summarize what has been discussed in this article into six points.

First, AI is already an intermediary that has entered the layer of recall. The expansion of Amazon Rufus to 300 million customers, the 2 billion monthly users of Google AI Overviews, and the doubling of generative AI usage rates in Japan—these figures indicate a tectonic shift as of the spring of 2026 [1][2][3][4].

Second, the degree of impact varies by category. AI intervention is limited for convenience goods, functions as an "organizer" for shopping/specialty goods, and becomes the center of the consideration process for high-involvement, information-asymmetric products. First, determine "where your company's category is positioned."

Third, the structure of top-of-mind recall is immutable. The only thing that changes is that an intermediary has been added to the subject doing the recalling. The wiring of CEPs → Recall → Consideration → Purchase flows in the same direction even after the emergence of intermediaries.

Fourth, both concepts of mental availability and physical availability are reinterpreted through a dual structure of "human + intermediary." Add AI recall monitoring to the recall KPI tree.

Fifth, the design of CEPs is actually strengthened in the AI era. This is because the questions that consumers throw at AI agents are not search queries, but natural language situational sentences. The verbalization of situational sentences leads directly to competitiveness in the AI era.

Sixth, while the discourse of "BtoA" is convenient, avoid both overreaction and underreaction. Wait and see for convenience goods, measure for shopping goods, and implement for high-involvement goods. Divide the intensity of your response across the three layers.

The essence of marketing is not persuasion, but recall. The essence of recall is not brute force, but design. The essence of design is not measures, but sequence. Even with the addition of an intermediary, these three stages do not change. The only thing that changes is that one reader, the AI agent, has been added to the subjects of design and operation. Let us proceed with preparations calmly, yet steadily.

References

[1] OpenAI. (2026). Update on ChatGPT Usage (Feb 2026). / TechCrunch. (2026). ChatGPT reaches 900M weekly active users. / TechCrunch. (2026). Google's Gemini app has surpassed 750M monthly active users. / Business of Apps. (2026). Google Gemini Revenue and Usage Statistics.

[2] Tinuiti. (2026). Amazon Rufus AI: Optimization Strategies for 2026. / Amazon.com Investor Relations Materials (2025--2026).

[3] Adobe Analytics. (2025). Holiday Shopping Report 2024. / Stellagent. (2026). AI Shopping Assistants Compared. / AdventurePPC. (2026). Generative AI Shopping Assistants and the Future of Google's Ad Business.

[4] Dentsu Inc. "Survey on Consumer Awareness Regarding AI (4th Edition)" released December 23, 2025. / Dentsu Digital "Survey on Consumer Purchasing Behavior" November 2025.

[5] Gartner. (2024). Press Release: Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents. / Gartner B2B Marketing Outlook 2025--2028.

[6] Edelman & LinkedIn. (2025). 2025 Edelman-LinkedIn B2B Thought Leadership Impact Report: Invisible Influence -- The Rise of the Hidden Buyer.

[7] LIFULL Co., Ltd. Press Release: "LIFULL HOME'S and Nomura Real Estate Solutions Jointly Develop and Launch 'AI ANSWER Plus (Beta),' a Real Estate Transaction Consultation AI Service for General Users," 2024. / Company Press Release: "New Features Added to AI HOME'S-kun BETA LINE Version," 2025. / Company Press Release: "LIFULL Announces 'LIFULL AI,' an Integrated AI Agent for the 'Next-Generation Housing Search Concept'," 2025.

[8] Accenture. (2025). Insurance Consumer Study. / Mitsui Sumitomo Insurance "MS-Assistant" Release, 2023. / Juniper Research. (2024). AI in Insurance Market Forecast.

[9] Brand, J., & Israeli, A. (2023). Using LLMs for Market Research. Harvard Business School Working Paper 23-062. / Berman, R., & Israeli, A. (2022). The Value of Descriptive Analytics: Evidence from Online Retailers. Marketing Science.

[10] Reisenbichler, M., Reutterer, T., Schweidel, D. A., & Dan, D. (2022). Frontiers: Supporting Content Marketing with Natural Language Generation. Marketing Science, 41(3), 441--452.

[11] Cao, Y., et al. (2024). Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations. arXiv:2502.01349. / "Global is Good, Local is Bad?" -- Understanding Brand Bias in LLMs (EMNLP 2024).

[12] Romaniuk, J. (2023). Better Brand Health: Measures and Metrics for a How Brands Grow World. Oxford University Press. / Interview with Jenni Romaniuk (Marketing Science media, 2025) / Ehrenberg-Bass Institute. (2024). AI-Driven CEP Identification Working Paper.

[13] Wertenbroch, K., et al. (2020). Autonomy in consumer choice. Marketing Letters, 31(4), 429--439.

[14] André, Q., Carmon, Z., Wertenbroch, K., et al. (2018). Consumer Choice and Autonomy in the Age of Artificial Intelligence and Big Data. Customer Needs and Solutions, 5(1), 28--37.

[15] Dawes, J., Romaniuk, J., & Sharp, B. (2021). The 95-5 Rule: How Advertising Works. The B2B Institute (LinkedIn) / Ehrenberg-Bass Institute for Marketing Science. / LinkedIn B2B Institute. The 95-5 Rule. (2024 update).

[16] Ellsberg, D. (1961). Risk, Ambiguity, and the Savage Axioms. Quarterly Journal of Economics, 75(4), 643--669. / Samuelson, W., & Zeckhauser, R. (1988). Status Quo Bias in Decision Making. Journal of Risk and Uncertainty, 1(1), 7--59.