Anatomy of ChatGPT Ads: Conversational Tagging, Contextual Targeting, and 30-Day Tracking Cookies—OpenAI Knows Exactly What You’re Thinking

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1664 words

Technical researchers reverse-engineering the ChatGPT ad backend have discovered that OpenAI's advertising system is far more complex than previously understood: ad units are embedded directly into conversation streams, four types of Fernet-encrypted tokens track click attribution, and the OAIQ v0.1.3 SDK reports behavioral data from the merchant side—covering the entire path from search intent to actual purchase. (Previous coverage: WSJ: OpenAI misses user and revenue targets, CFO worries about data center bills before IPO) (Background: Amazon shopping assistant now supports real-time two-way conversation, Amazon AI shopping matrix upgraded) 720 hours. This is the lifespan of the tracking cookie in the browser after a user clicks a ChatGPT ad—a full 30 days, enough to cover the entire consumer journey from "planning a trip to Japan" to "looking up restaurants after landing." According to a reverse-engineering analysis of the ChatGPT ad backend by technical researcher Buchodi, OpenAI's advertising infrastructure is far more complex than the public realizes. Instead of inserting banners in the sidebar of the chat page, ads are injected directly into the conversation flow, paired with four types of Fernet-encrypted tokens, a merchant-side SDK, and an in-app browser that monitors user behavior after a click. This is a complete ad-tech stack. Every ChatGPT conversation is streamed via the backend /conversation endpoint using the SSE protocol. In plain terms, the server continuously pushes data streams to the browser, sending model responses as they are generated. Researchers found that ad units are embedded within this data stream as single_advertiser_ad_unit objects, containing advertiser brand identifiers, carousel card creative assets, and image assets hosted on the OpenAI ad CDN. When a user clicks an ad, the link does not open in an external browser but loads within ChatGPT's built-in WebView, allowing OpenAI to continuously track user navigation behavior post-click. Contextual targeting is also already operational. Researchers tested six conversation topics and confirmed that ads are indeed served based on conversation content: Axel promotions appeared when searching for flights, while Gametime ticket ads were displayed when discussing the NBA. What ChatGPT possesses is not just what keywords you searched for, but what you are currently thinking about. Each ad unit contains four tokens encrypted with the Fernet algorithm: - ads_spam_integrity_payload: A server-side verification mechanism used to filter invalid clicks. - oppref: A forward-attribution token stored in the user's browser cookie, valid for 720 hours (30 days). - olref: A reference token paired with oppref for internal OpenAI logging. - ad_data_token: A Base64-wrapped JSON containing an additional layer of encrypted data. Even without the decryption key, researchers can reconstruct the token's generation timestamp from the public header fields of Fernet, meaning the precise time of ad exposure can be tracked externally without decryption. Merchants have also deployed corresponding infrastructure. The OAIQ SDK version 0.1.3 automatically loads on the advertiser's website, reads the oppref parameter to create a local cookie, and reports user behavior events back to OpenAI's reporting server. This chain covers the entire path from "ad exposure in a conversation" to "user action on the merchant's website." In ad-tech jargon, this is called a "full attribution loop"—in plain terms, the ad platform can fully track which specific ad exposure ultimately led to a purchase. The core advantage of the OpenAI advertising system lies in the uniqueness of its contextual data. Google knows what you searched for, Facebook knows what you liked, but ChatGPT knows what you are thinking about: planning a trip, considering a job change, or researching a purchase decision. These are data points with far more intent than keywords, and they are the core metrics for contextual advertising effectiveness. Advertisers are not just pushing messages to a potentially relevant audience; they are appearing at the precise moment of decision-making. Earlier this year, The Wall Street Journal reported that OpenAI missed its user and revenue targets, with the CFO worrying about paying data center bills before an IPO. Advertising revenue is a growth curve that can be pulled without waiting for the next generation of GPT. Further reading: WSJ: OpenAI misses user and revenue targets, CFO worries about data center bills before IPO

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-04-29

Anatomy of ChatGPT's Advertising Mechanism: Conversations as Tags, Quadruple-Encrypted Tokens, and Merchant SDK Callbacks—The OpenAI Ad Stack Has Taken Shape

2026-04-23

OpenAI launches ChatGPT for Clinicians, assisting human physicians with 99.6% accuracy (free access in the US)

Similarity 170%關鍵字 openai/chatgpt同分類 zh

2026-04-21

OpenAI launches ChatGPT Images 2.0: Major evolution in text rendering, supports multi-image generation, but "Chinese generation" remains a hurdle

Similarity 170%關鍵字 openai/chatgpt同分類 zh

2026-04-20

Srinivas Narayanan, a key figure behind ChatGPT, has abruptly resigned from OpenAI, marking the third high-level departure in a week.