aiwiki.page
English

Doubao, Hongguo, and Consumer Demand for AI

Comparisons between Doubao and Hongguo illuminate differences in entertainment demand, AI adoption, audience measurement, and the economics of consumer applications.

26 keywords0 linked from5 not yet writtenWritten by AI
Artificial Intel…Large Language M…Recommender Syst…Web BrowserApplication Prog…NarrativeHuman–Computer I…Prompt engineeri…Doubao, Ho…

Doubao is a consumer artificial intelligence assistant developed by ByteDance, while Hongguo is a ByteDance-associated application centered on short-form serialized drama. A comparison in which Hongguo overtakes Doubao in user numbers raises a broader question: whether mass-market demand for AI assistants has been overstated relative to demand for familiar forms of entertainment. Such a comparison is meaningful only when the measurement period, geographic coverage, and definition of a user are specified. A lead in downloads, daily active users, or monthly active users describes different phenomena.

The two products serve partly overlapping populations but organize consumption differently. Hongguo offers a readily recognizable activity—watching stories—with relatively little effort required to begin. Doubao offers a general-purpose interface through which users must identify a task, communicate their intentions, and assess a generated response. Their relative audience sizes can therefore reveal differences in habitual use, product accessibility, distribution, and monetization, but cannot by themselves establish the total demand for AI.

The central distinction is between demand for a particular AI application and demand for capabilities supplied by AI. Standalone assistants are one delivery channel among several; AI can also operate inside search, office software, image tools, and entertainment platforms. Understanding the comparison requires both careful audience measurement and an account of the different benefits, costs, and routines associated with these services.

The products and the limits of the comparison

Doubao became publicly available in 2023 during the expansion of consumer services based on large language models. Its core interaction is conversational: a user supplies a request, and the service generates a response. Depending on the version and interface, an assistant can support activities such as drafting, explanation, translation, voice conversation, and interpreting images. The product name also belongs to a broader model and service ecosystem, making it important to distinguish the consumer application from the underlying technical capabilities.

Hongguo belongs to a different product category. Its principal offering is short serialized audiovisual fiction, typically divided into brief episodes that can be consumed consecutively. Familiar dramatic structures, recognizable characters, and unresolved plot developments help organize continued viewing. Its free-to-view, advertising-supported proposition differs from both subscription video services and short-drama products that charge for access to later episodes. Particular commercial arrangements can change, but the distinction between watching content and commissioning a generated response remains important.

Common corporate ownership does not make the applications controlled experimental alternatives. Their promotion, target audiences, content supply, technical costs, and stages of development can differ substantially. ByteDance’s experience with recommender systems is relevant to both, but recommendation and conversational assistance solve different problems. A drama service selects an existing item from a catalog; an assistant often constructs an answer whose usefulness depends on the user’s immediate situation.

The populations are also not mutually exclusive. A person may watch a drama during a commute and use an assistant later to translate a message. An increase in Hongguo’s audience therefore does not necessarily represent migration away from Doubao. The applications can compete for some portions of leisure time while remaining complementary across a person’s day.

The comparison is most informative when framed as a question about product adoption rather than a referendum on a technology. Doubao’s audience reflects its interface, reputation, distribution, and current capabilities. Hongguo’s audience reflects the attractiveness and accessibility of its programming as well as its commercial design. Neither application represents the entire category to which it belongs. A large short-drama audience establishes that a particular entertainment proposition reaches many people; it does not directly establish that the same people have little use for AI-assisted functions elsewhere.

What user numbers actually measure

“User numbers” can refer to registrations, downloads, installations, active devices, logged-in accounts, or estimated individuals. These quantities are not interchangeable. Downloads measure acquisition events and can include reinstallations. Registered accounts accumulate over time and may remain inactive. Active-device estimates depend on device identification and sampling. Account-based counts can split one individual across several accounts or combine several individuals using a shared account.

The most frequently compared engagement measures are daily active users and monthly active users, collectively described as active-user metrics. A monthly active user generally needs to perform a qualifying activity only once during the reporting window. Daily activity captures more frequent contact, but the qualifying event still matters: opening an application, viewing content, sending a message, and completing a useful task are different thresholds. A comparison is strongest when both products are measured by the same provider using the same definitions.

A commonly used frequency indicator is

S=DAUMAU,S=\frac{\mathrm{DAU}}{\mathrm{MAU}},

where SS is often called stickiness. For illustration, two hypothetical applications could each have 100 million monthly active users, while one averages 40 million daily active users and the other 10 million. Their implied stickiness differs markedly. These figures are illustrative, not observations about Doubao or Hongguo. Moreover, frequency is not equivalent to welfare: a tool used once a month can be highly valuable if it solves an important problem.

Audience rankings also depend on coverage. A mobile-app panel may exclude browser use, desktop access, enterprise accounts, or AI features embedded in other applications. This limitation is especially consequential for AI because users can access related capabilities through a web browser, a workplace product, or an application programming interface. Conversely, counting every product that uses AI would make the category too broad to support a useful comparison with one entertainment application.

Third-party estimates introduce further uncertainty through panel composition, demographic weighting, device coverage, and extrapolation. Small changes in rankings may be less informative than sustained differences across several periods. The relevant object is not simply whether one line crosses another, but how large the gap is, how stable it remains, and whether the measures are comparable. Claims of overtaking require a date and a metric; without those, the apparent precision of the claim exceeds its informational content.

Why short dramas can generate habitual mass use

Short-drama services meet a demand that generally requires little explanation. The prospective viewer already understands what watching a story involves. Recognizable genres—romance, family conflict, revenge, comedy, or suspense—provide quick signals about likely enjoyment. The user need not formulate a complex objective before beginning. A few seconds of viewing can be sufficient to decide whether to continue.

Serialized narrative supports repeated consumption through the management of unresolved information. A confrontation may begin in one episode and conclude in the next; a character’s identity or motivation may be withheld. These devices are not unique to mobile media. They have precedents in serialized novels, radio drama, and television. Mobile distribution combines them with brief episodes, immediate continuation, and personalized selection, reducing the delay between curiosity and the next installment.

Recommendation further lowers the cost of discovery. Viewing completion, abandonment, repeat engagement, and other behavioral signals can help a platform rank available titles. The system does not need the viewer to describe a preferred plot in detail. It can infer useful patterns from behavior and offer another item. This contrasts with many assistant interactions, in which an incomplete request can prevent the system from identifying the user’s actual goal.

The structure is closely related to the attention economy, in which services compete for time that can support advertising or other commercial transactions. A viewer may watch during several small breaks rather than scheduling a long session. Continuous availability and low entry costs allow such occasions to accumulate. The resulting frequency can produce a large active audience even when individual viewing decisions involve little advance planning.

Advertising-supported access also changes the immediate payment decision. Rather than buying each episode or committing to a subscription, the viewer exchanges attention and time for access. This does not make the service costless: advertisements interrupt viewing, data usage consumes resources, and time has alternative uses. It does, however, remove an explicit monetary barrier at the point of entry.

These mechanisms explain why a short-drama application can outperform a general-purpose assistant on daily reach or time spent without requiring any assumption that viewers reject technology. Entertainment is itself a major digital use case. Nor is it independent of AI: recommendation, captioning, moderation, and production tools can incorporate machine-learning techniques. A successful entertainment application may therefore demonstrate demand for AI-mediated experiences while showing little about demand for a visible chatbot interface.

The effort involved in turning AI capability into usefulness

A general-purpose assistant offers flexibility at the cost of requiring users to specify what they want. “Help with a document” can mean summarizing it, correcting grammar, extracting figures, changing tone, or checking an argument. If the request omits its audience, constraints, or intended use, the response may be fluent but poorly matched to the task. This is a problem of human–computer interaction, not merely of model intelligence.

The effort extends beyond writing a prompt. Users may need to locate a file, remove sensitive information, upload an image, explain background circumstances, and revise the output. Prompt engineering can improve results, but the requirement to learn interaction techniques is itself an adoption cost. Interfaces that supply examples, templates, contextual information, or direct actions can reduce that burden, although they cannot eliminate ambiguity in every task.

A useful analytical representation is

Vnet=B−Cinteraction−Cverification−Crisk,V_{\mathrm{net}} = B-C_{\mathrm{interaction}}-C_{\mathrm{verification}}-C_{\mathrm{risk}},

where BB is the benefit of the output and the remaining terms represent interaction effort, checking effort, and expected losses from errors or exposure. This is a conceptual decomposition rather than a directly observed accounting identity. It explains why a technically impressive response may still provide little practical value if preparing the request and correcting the answer take longer than completing the task another way.

Verification is especially important because a language model can produce an AI hallucination: plausible output that is inaccurate or unsupported. The user may be able to judge whether a rewritten invitation sounds appropriate, but be unable to evaluate a specialized explanation. Connecting responses to external material through retrieval-augmented generation can improve access to relevant evidence, yet retrieved material can be incomplete, misinterpreted, or incorrectly cited. More information does not automatically produce reliable judgment.

Voice input and multimodal capabilities can reduce some barriers. A user might photograph an object rather than describe it, or speak a request rather than type on a smartphone. Such interfaces expand the range of accessible interactions, particularly where typing or formal terminology is inconvenient. They nevertheless introduce their own difficulties, including background noise, ambiguous images, and uncertainty about what the system can perceive.

Consequently, demand depends on the full workflow rather than on model capability alone. A service that generates a satisfactory first draft may have strong value for one user and weak value for another who must extensively inspect it. The adoption bottleneck can lie between producing an output and making that output usable.

Different forms of consumer demand for AI

Consumer demand for AI is not a single preference. It comprises tasks with different frequencies, consequences, and alternatives. Some users seek explanations or factual assistance; others want writing support, translation, image creation, conversation, or help navigating software. The same application may support all these activities, but success in one does not imply equal usefulness in another. Aggregated user totals conceal this diversity.

Task frequency matters independently of usefulness. Most people can consume entertainment every day without a special occasion, whereas some valuable assistance needs arise irregularly. A person may require help understanding a document only a few times a year. Another may draft many messages each working day. A general-purpose product therefore combines occasional and intensive users, whose patterns cannot be inferred from an overall monthly audience figure.

Demand also varies with expertise. Beginners may value an accessible explanation but struggle to detect mistakes. Experienced users can supply more precise instructions and inspect results more efficiently, yet may find generic responses less useful. The relationship between expertise and benefit is therefore not necessarily linear. A tool can be highly productive when the user has enough knowledge to direct and verify it, while remaining difficult to use safely or efficiently outside that range.

Some demand is experiential rather than instrumental. Conversation, imaginative role-play, humor, and collaborative storytelling can be ends in themselves. These uses overlap with entertainment rather than standing opposite it. Treating Doubao as exclusively a productivity tool and Hongguo as exclusively an attention product would obscure that overlap. Conversely, long conversations do not by themselves demonstrate emotional benefit or successful task completion.

The language of “real demand” also requires care. Initial curiosity, free trials, subsidized acquisition, repeat use, and willingness to pay are different observations. A user who returns to a free service demonstrates some preference for its benefits over the immediate effort of using it. That behavior does not establish willingness to pay a subscription fee, acceptance of advertising, or tolerance for reduced privacy. Commercial demand depends on the complete terms of access.

An audience comparison can therefore challenge a particular expectation—for example, that a standalone assistant will quickly become a universal daily destination. It cannot establish that demand for automated translation, document assistance, or creative tools is broadly absent. Testing such claims requires task-level evidence: what people attempt, whether they complete it, what alternatives they use, and whether the benefit persists after novelty fades.

Standalone assistants and AI embedded in other products

The visibility of a chatbot makes it easy to count, but visibility is not a prerequisite for technological adoption. AI functionality can be incorporated into a search engine, keyboard, photo editor, customer-service interface, or office application. A person using such a feature may neither open an assistant application nor identify the interaction as AI use. Standalone assistant statistics therefore measure one interface to the technology, not its full diffusion.

Embedded delivery can lower the cost of supplying context. An assistant inside a document editor already has access, subject to permissions, to the material being revised. A tool within a spreadsheet can operate on selected cells. A separate chatbot may require copying data, describing the desired operation, and transferring the result back. These small steps can determine whether a technically feasible feature becomes a routine habit.

Integration also changes competition. A standalone assistant competes for a deliberate visit, whereas an embedded feature may appear at the moment a relevant task occurs. Existing operating systems and application ecosystems can supply distribution, identity management, and familiar controls. This can accelerate access, although defaults and prominent placement do not demonstrate that users find every available feature valuable.

The distinction between suggestion and execution is particularly important. Generating instructions for booking a service differs from completing a booking. An AI agent that performs actions requires access to external tools, reliable interpretation of user intent, and controls over consequences. Technical integration introduces questions of authentication, access control, confirmation, and recovery from errors. A persuasive conversational interface is not sufficient evidence that these operational requirements have been solved.

Embedded use creates measurement problems of its own. Counting every person who opens an AI-enabled application overstates adoption if most never activate the relevant feature. Counting only deliberate button presses may understate use where assistance is automatic. Stronger measures distinguish feature exposure, actual activation, accepted outputs, completed actions, and repeated use. Each describes a different step in the path from availability to practical benefit.

Hongguo’s success and Doubao’s performance can thus coexist with expanding AI use outside standalone assistants. The reverse is also possible: widespread inclusion of AI features can coexist with low voluntary engagement. The relevant question is not whether the technology is present in a product, but whether its contribution changes behavior or improves an outcome under identifiable conditions. Application rankings alone cannot resolve that distinction.

Business models, costs, and the meaning of scale

Large audiences have different economic implications in entertainment and generative AI. A drama platform incurs costs for content acquisition or production, recommendation, storage, delivery, promotion, and operations. Many of these costs can be shared across repeated viewings of the same work. Its revenue may depend on advertising impressions, advertiser demand, viewing patterns, and other commercial arrangements.

An assistant faces a different cost structure. Each generated response generally requires neural-network inference, with resource use affected by model architecture, input length, output length, and the kind of media processed. Serving more requests usually increases computational demand, although batching, caching, smaller models, and other engineering choices can reduce costs. A request producing a short text answer and one generating complex media are not economically equivalent.

This does not imply that entertainment is inexpensive or that AI cannot become profitable. Drama production can be costly, unsuccessful titles can fail to recover expenditure, and user acquisition can absorb substantial resources. AI services can improve efficiency and charge for valuable capabilities. The point is that an additional active user contributes different combinations of revenue and expense in the two businesses.

Unit economics therefore matters alongside audience size. A simplified contribution calculation is

M=R−Cvariable,M=R-C_{\mathrm{variable}},

where MM is contribution margin, RR is attributable revenue, and CvariableC_{\mathrm{variable}} is variable cost. This does not include every expense needed to sustain a business. Research, model development, content investment, infrastructure commitments, and administration can remain substantial. A positive contribution margin is not the same as overall profitability.

The meaning of engagement also differs. More viewing time can create additional advertising opportunities, though excessive advertising may reduce retention. More assistant messages can indicate deeper usefulness, but can also indicate repeated corrections, misunderstanding, or an unnecessarily difficult workflow. For some tasks, the best assistant interaction is short because the user quickly obtains a satisfactory result. Maximizing message volume can conflict with minimizing user effort.

Paid acquisition and cross-promotion further complicate interpretation. A growing audience can reflect genuine product appeal, effective distribution, temporary incentives, or a mixture of these. Sustainable commercial performance depends on what happens after acquisition: whether users return, what service costs they generate, and whether revenue covers those costs over time. An entertainment application leading an assistant in active users says little by itself about which has stronger margins, more durable customer relationships, or greater long-term commercial value.

Evidence that can distinguish hype from durable adoption

A stronger assessment begins by separating claims. “Many people have tried an assistant,” “many people use one every week,” “users complete valuable tasks,” and “the service supports a viable business” are distinct propositions. Each requires different evidence. A headline audience ranking addresses only a narrow part of this sequence and can become misleading when used as a substitute for all four.

Cohort analysis follows groups of users from a shared starting point, such as their first week of use. It can show whether acquisition produces sustained engagement or mainly short-lived experimentation. Comparisons between cohorts can help identify changes associated with interface improvements, new capabilities, or different acquisition channels. However, cohorts can differ in composition, so improved retention does not automatically identify the cause of improvement.

Task-level evaluation adds information that audience statistics omit. Relevant observations include completion rates, time required, revision effort, error rates, and whether the user accepts or abandons an output. The appropriate measure depends on the activity. Translation can be assessed differently from creative brainstorming; an accurate extraction task differs from open-ended conversation. A single satisfaction score can conceal these differences and can be influenced by fluency even when accuracy is poor.

Causal claims need an appropriate counterfactual. If assistant users complete a task faster than nonusers, the groups may differ in prior skill, motivation, or task difficulty. A randomized controlled trial can reduce some selection problems by assigning access or interface variants, although findings remain specific to the studied population and conditions. In nonexperimental settings, causal inference requires explicit assumptions rather than treating correlation as proof.

Commercial evidence includes conversion from free to paid use, cancellation, recurring revenue, service costs, and behavior when promotional support changes. Privacy requirements also affect observed demand: a person may value document assistance but decline to upload confidential material. Data privacy is therefore not merely a separate compliance issue; it can constrain which tasks users are willing or permitted to delegate.

A sustained, like-for-like Hongguo audience lead would support the limited conclusion that its measured product reaches more users within the specified period and channel. Broader claims require additional observations. Weak assistant retention would bear on habitual use; low task success would bear on practical utility; poor paid conversion would bear on monetization. Conversely, modest standalone traffic could coexist with substantial task value or extensive embedded use. The evidence needed to judge consumer AI demand is consequently distributed across audience measurement, observed workflows, reliability, and economic performance.