Internet Advertising Systems: Technical Principles, Delivery Architecture, and Compliance

1. Scope and Definitions

Internet advertising refers to commercial advertising delivered through websites, web pages, internet applications, and other online media in the form of text, images, audio, video, or other formats to directly or indirectly promote goods or services. The key test for whether content is an ad is whether it has a commercial promotional purpose, not which type of page it appears on or which medium it uses.

Under current rules, internet advertising generally includes five categories: linked text, image, or video promotional ads; email ads that promote goods or services; paid search ads; ads in commercial presentations; and other commercial ads that promote goods or services through internet media. Information that must legally be provided to consumers in commercial presentations is displayed in accordance with relevant laws and regulations. For developers, this classification directly determines which delivery channels a system must integrate and which types of data it must handle.

2. Development and Regulatory Context

In October 1994, a US magazine focused on technology and culture launched an online edition whose homepage featured banner ads from more than ten companies. This is commonly viewed as the starting point of online advertising. Over the following three decades, online advertising evolved from fixed-size static banners into data-driven programmatic delivery centered on automated bidding, with repeated changes in formats and billing models.

On the regulatory side, in 2023 China’s market regulator revised and issued the Measures for the Administration of Internet Advertising, effective May 1, 2023. The measures detail rules for ad labeling, advertiser responsibilities, platform review obligations, and formats such as pop-ups and push notifications. When designing advertising features, developers should treat compliance requirements as part of the product requirements rather than as a check to be added before launch.

The source cites 2025 internet advertising business revenue of RMB 1,357.43 billion, up 34.6% year on year and accounting for 66.2% of total advertising revenue. Continued scale growth means ad systems face rising pressure on concurrent requests, data storage, and service stability.

3. Main Ad Formats

(a) Banner Ads

Banner ads generally refer to image-based ads placed in fixed areas at the top, side, or middle of a page. Common forms include full-width banners, vertical banners, and large-format banners. Creative assets often use common image formats such as GIF, JPG, and PNG. Scripts can also be used for click interactions, or rich-media technology can be used to enhance dynamic presentation.

Three implementation points matter. First, asset size directly affects first-screen load speed, so images should be compressed and individual asset size should be limited. Second, dynamic assets require control over frame rate and resource usage to avoid slowing page rendering. Third, a static fallback version should be provided for low-performance devices or unstable network conditions.

(b) Email Ads

Email ads offer strong targeting, low per-touch cost, and content that is not limited by layout space. However, they require recipient authorization. Technically, the system must provide a working unsubscribe mechanism and manage sending frequency, recipient lists, and bounce status. Before sending, the system should perform deduplication and validity checks to avoid being classified as spam due to bulk sending.

(c) Paid Search Ads

Paid search ads are bid and displayed based on user query terms. Advertisers bid on keywords, and the platform combines bid, relevance score, landing page quality, and other factors to determine placement and order. Developers need to understand keyword matching methods, the role of negative keywords, and how landing page relevance to keywords affects delivery performance.

(d) Linked Text, Image, and Video Ads

These ads are embedded in article body content or content feeds and appear in native form, making them the category most easily confused with ordinary content. Design should ensure that ad labels are clearly recognizable and avoid layouts that mislead users into clicking. This is also a key area in compliance review.

4. Delivery Technical Pipeline

An ad impression typically goes through the following six stages:

1. Ad request: When a user visits a page or app, the client sends a request to the delivery system with information such as ad slot, device type, time period, and page context.
2. Candidate filtering: The system filters ad campaigns that meet targeting conditions such as geography, time period, and interest tags.
3. Bidding and ranking: Multiple candidate ads are scored using factors such as bid and predicted click-through rate, and the highest-scoring ad wins.
4. Frequency control: The system checks how many times the user has seen the ad. If the limit is exceeded, it skips that ad and selects the next candidate.
5. Creative return and rendering: The system returns the asset URL and tracking parameters, and the client renders and displays the ad.
6. Measurement and callback: The system records impression, click, and conversion events for billing, settlement, and performance analysis.

The following pseudocode illustrates only the selection logic and omits exception handling and concurrency details:

function select_ad(request, campaigns, freq_store):
    eligible = filter campaigns where campaign matches request.context
    ranked = sort eligible by (campaign.bid * campaign.predicted_ctr) descending
    for campaign in ranked:
        key = request.user_id + ":" + campaign.id
        if freq_store.count(key) < campaign.daily_cap:
            freq_store.increment(key)
            return campaign
    return NULL

In production environments, two issues require special handling. First, frequency counts are shared state across multiple service instances, so atomic increments must be used and expiration must be set; otherwise counts can become inconsistent or grow without bound. Second, the timing of count writes must be consistent with impression confirmation. If rendering fails after the asset is returned but an impression is still counted, users will actually see the ad fewer times than recorded. Teams should decide when to write the count based on how much error the business can tolerate.

5. Technical Support for Targeted Delivery

Advertising precision mainly depends on three capabilities.

Data collection and profiling. The system collects behavioral data within the scope of user authorization and converts it into labels. This step should follow the principle of minimum necessity: collect only the fields required for delivery, clearly inform users of the purpose of data use, and allow users to withdraw authorization.

Targeting strategy. Targeting dimensions include geography, device, time period, interest tags, and page context. The finer the targeting conditions, the more precise the matched users, but the smaller the reachable audience. Product design must trade off precision against reach and validate the effects of different combinations through experiments.

Performance prediction. Click-through rate and conversion rate prediction usually use machine learning models, with training performed offline and inference performed online in real time. The most common engineering problem is feature inconsistency: the feature definitions used during training differ from those used during online inference, causing prediction bias. The solution is to consolidate feature computation logic into shared components used by both offline and online systems.

6. Performance Measurement and Billing

Common billing models include cost per mille (CPM), cost per click (CPC), and cost per action (CPA). Different models correspond to different event counting definitions, so developers need to establish a unified event model that ensures impression, click, and conversion events can be deduplicated, traced, and reconciled.

In implementation, each ad request can be assigned a unique identifier, and impression, click, and conversion events can all carry that identifier for association and deduplication. Click data also needs invalid traffic filtering, such as repeated clicks within a short period or visits that clearly come from automated programs. The length of the attribution window should be clearly defined in documentation; otherwise, reports from advertisers and platforms can easily diverge.

7. Compliance Points

1. Ad labeling: Ad content should be clearly marked as an ad in a prominent way so users can distinguish ads from non-ad information.
2. Content truthfulness: Content must not be false or misleading, and absolute claims must not be used.
3. Pop-ups and push notifications: Clear close controls should be provided, and users must not be misled into clicking.
4. Email: A working unsubscribe method should be provided, and unsubscribe requests should be handled promptly.
5. Personal information: The collection and use of targeting data must comply with personal information protection laws, and sensitive personal information must not be used for targeted push.
6. Protection of minors: Goods or services unsuitable for minors must not be advertised to them.

Compliance should be considered during system design. For example, field validation can be performed at the creative submission stage, automated review can be performed before delivery, and manual review can be retained for high-risk categories.

8. Engineering Practice Recommendations

• Module decoupling: Split delivery decisioning, billing and settlement, risk control and review, and reporting and analytics into independent modules so they can evolve and scale separately.
• Atomicity for budgets and frequency control: Budget deduction and frequency counting must use atomic operations, with timeouts and degradation strategies. Under resource pressure, it is better to under-deliver than to exceed the budget.
• Key metric monitoring: Continuously monitor request volume, fill rate, response latency, and impression-to-click ratios. Abnormal metric fluctuations should trigger alerts and support rapid rollback.
• Material review upfront: Complete compliance checks before assets go live, combining automated review with manual review for high-risk categories.