Almost Nobody Asks ChatGPT for "Recommendations" — A Quantitative Analysis of 11,349 Real Logs

Who This Article Is For

If you're creating content with the goal of getting cited in AI search results — read on.

Are you designing FAQs and content pages around queries like "What's the best ___?" Are you writing SEO articles targeting "best ___" and "top 10 ___," assuming those same patterns will work for AI search?

The data says otherwise.

This article presents findings from a quantitative analysis of 11,349 real ChatGPT conversation logs (including 2,500 Japanese queries) across 7 dimensions. We show, with numbers, how users actually talk to AI.

---

Study Overview

All analysis is based on open data. A caveat: enterprise users and paid-tier users are underrepresented in these datasets.

---

Finding 1: 78% of Japanese Users Don't Ask "Questions"

The first discovery is that the query structure of Japanese users is fundamentally different from English users.

78% of Japanese users don't use interrogative forms. Instead, they use imperative and request patterns — "~shite" (do this), "~oshiete" (tell me), "~kudasai" (please do) — rather than forming questions.

The usage rate of question marks ("?") is 85% in English but only 46% in Japanese.

What This Means for AEO

Even if you design content in a Q&A format — "What is ___?" or "How do I ___?" — it won't match the actual input patterns of Japanese users. AI selects content to cite based on whether it's an appropriate answer to the user's input. Designing content without knowing the input patterns is like shooting arrows without knowing where the target is.

---

Finding 2: "Best" and "Recommend" Are Barely Used

This is the most important finding.

Out of 2,500 Japanese queries, "osusume" (おすすめ, meaning "recommendation") appeared in only 7 queries.

Moreover, most of those 7 weren't even recommendation queries. They appeared in contexts like "Write a blog post that includes a section on recommended manufacturers" — as part of content generation tasks, not product discovery.

Total recommendation intent: 4.4% in English vs. 0.56% in Japanese.

The "Best/Recommendation" Trap

Many AEO articles and SEO strategies target "best" or "recommendation" keywords. But the data shows users simply don't use these words when talking to AI.

What they actually say:

• "Tell me about ___"
• "What's the difference between ___ and ___?"
• "Compare ___ for me"

If you build your strategy around "best" and "top 10," you're missing virtually all real query patterns.

---

Finding 3: What Users Ask About Differs Dramatically by Language

Japanese users interact with AI in a qualitatively different way than English users:

• Romance queries are 3x higher: 5.8% in Japanese vs. 1-2% in English
• Coding is 1/3 or less: English users treat ChatGPT as a coding tool; Japanese users use it more for creative and consultation purposes
• Translation is 7x higher: 4.2% in Japanese vs. 0.5-0.7% in English

AEO Implications

If you're optimizing for AI search in the Japanese market, don't copy English-market case studies directly. Japanese users primarily "consult" and "request" rather than "research," and content design must reflect this difference.

---

Finding 4: "Prompt Engineering" Is Largely a Myth

The most-used "technique" isn't a technique at all. Polite language (please / kudasai) leads at 16%.

For Japanese users specifically, polite expressions jump to 32-39% — roughly double the English rate. Meanwhile, English-style "prompt engineering" techniques like "act as" are virtually nonexistent in Japanese queries.

The vast majority of users don't engineer their prompts. They ask in ordinary language, the way they'd speak to a person.

---

Finding 5: Query Length Is Determined by Topic

Creative and implementation tasks get long, detailed instructions. Consultation and recommendation queries are short — often a single sentence.

Romance: 61 characters. Travel: 69 characters. Advice: 72 characters.

AEO Implications

For short consultation queries, your content must supply the context the user didn't. When a user types a 61-character input like "tell me about ___," your content needs enough information density and structure for the AI to construct a proper answer — otherwise it won't be cited.

---

Finding 6: Japanese Users Name Specific Products

Cross-country specificity comparison (top 5 countries on WildChat, based on IP geolocation — treat as directional):

Japan has the highest product-naming rate at 51% across 12 countries analyzed.

Furthermore, Japan has the highest constraint rate at 58% — users tend to add temporal (37%) and purpose-based (31%) qualifiers, asking in specific, detailed ways.

AEO Implications

Japanese users ask with specific product names — "Tell me about [Product Name]'s ___." Since the probability of brand/product names appearing in queries is high, whether your company owns the content that AI can cite when answering about that product becomes critically important.

---

Summary: Rethink Your AI Search Optimization Assumptions

This analysis revealed 6 facts:

1. 78% of Japanese users don't use question forms — Reconsider FAQ-style content design
2. "Recommendation" queries account for only 0.28% — Stop optimizing for queries that don't exist
3. Japanese users primarily consult and create — Don't directly apply English-market case studies
4. Prompt techniques are barely used — Users ask in ordinary language
5. Consultation queries average 61-72 characters — Content structure must compensate for minimal user input
6. Japan is a "name the product" culture — Own the information about your own products

Content design for AI search citations should be reverse-engineered from data on how users actually ask. Assumptions and extensions of "common search keywords" won't produce content that AI chooses to reference.

---

Data Sources and Caveats

• All open data: Only publicly shared conversations. Enterprise use, private consultations, and paid-tier users are underrepresented. Results may differ from closed datasets
• Rule-based classification: No NER used. Specificity and topic classifications are approximate. About half of all queries fall under "Other." LLM-based clustering would improve precision
• Japanese data contamination: Of 2,500 Japanese queries in WildChat, 41% were translation tasks and 12% were spam. Clean analysis is based on 1,053 queries
• Country data limitations: Based on IP geolocation (includes VPN/proxy). "Japanese language" does not necessarily mean "from Japan"

The purpose of this analysis is to capture the broad landscape of how users talk to AI. The value lies in discovering qualitative patterns, not in the precision of individual numbers.