ChatGPT serves ads. Here's the full attribution loop
A meticulous technical breakdown reveals how OpenAI is serving contextual ads and tracking user behavior within ChatGPT, complete with a proprietary attribution SDK. This finding confirms the long-feared 'enshitification' of AI services, sparking widespread dismay among users. It reignites fervent discussions about the increasing necessity and benefits of embracing local and self-hosted LLMs as a privacy-preserving alternative.
The Lowdown
A detailed investigation by Buchodi uncovers the full attribution loop for ads now being served within ChatGPT. OpenAI has developed a sophisticated, two-part advertising platform that injects structured ad units into conversation streams and tracks user interactions via a custom SDK on merchant websites.
- Ad Injection: Ads are delivered as distinct
single_advertiser_ad_unitobjects within the SSE stream of ChatGPT's backend API, alongside model output. - Contextual Targeting: Ads are contextually relevant to the ongoing chat topic, with examples including Grubhub for Beijing trip planning and Gametime for NBA playoffs.
- Attribution System: A complex four-token Fernet-encrypted system tracks ad impressions and clicks. The
oppreftoken, in particular, is stored in a 30-day cookie by an OpenAI-hosted JavaScript SDK (OAIQ) on merchant sites to tie post-click activity back to OpenAI. - OpenAI Control: OpenAI hosts advertiser creatives on its
bzrcdn.openai.comdomain and uses in-app webviews for clicks, maximizing its data collection on user navigation. - Blocking: The article identifies
bzrcdn.openai.comandbzr.openai.comas domains to block for ad event prevention, along with inspecting__opprefand__oaiq_domain_probecookies.
This analysis clearly demonstrates that OpenAI is building a robust advertising and tracking infrastructure within ChatGPT, leveraging its platform for monetization and data collection from user interactions.
The Gossip
Enshitification Eruption
The overwhelming sentiment among commenters is one of profound disappointment and resignation. Many users lament the arrival of ads and tracking as the inevitable 'enshitification' of ChatGPT, marking the perceived end of a 'golden age' for AI services. The move is widely seen as a betrayal of user expectations and a corporate push for monetization at the expense of user experience.
Local LLM Liberation
This development significantly strengthens the argument for local and self-hosted AI models. Commenters enthusiastically discuss the growing viability and appeal of running LLMs locally, citing increased privacy, control, and freedom from commercial pressures (like ads and price hikes) from centralized providers like OpenAI. While acknowledging potential trade-offs in performance or features compared to the latest hosted models, the sentiment leans towards self-hosting as a necessary and increasingly practical alternative.
Ad-Blocking Blueprint
A thread of discussion focuses on the technical feasibility and future challenges of blocking these integrated ads. Initial observations suggest that because ads are currently served as distinct events, they might be relatively easy to filter. However, users express concern that future iterations could see ads injected more subtly directly into the model's main response, making detection and blocking significantly more difficult.