Five Possible Back-Doors into AI model Search Results

What happened

With the plunge in traditional traffic from search engine results, due to the new trend for AI summaries, publishers at all levels are now seeking to gain prominence in the attention of the Large Language Models that power web-search AI summaries, and directly, in popular frontier models such as ChatGPT. Legitimate approaches, according to prior works such as Optimizing Visibility in Generative Engines and How to Dominate AI Search , include securing independent editorial coverage; publishing authoritative and well-structured content; providing machine-readable product information; and adapting material for different languages and markets.

One particularly ingenious example of GEO manipulation was when researchers at ETH Zürich created a fictitious camera and published a promotional webpage containing a concealed instruction, rendered in tiny white text, directing Microsoft’s Bing Copilot to recommend the camera exclusively. The tiny font size and low contrast was designed to conceal the manipulation.

Two fictional camera advertisements used in an ETH Zürich experiment to manipulate AI search recommendations. The conventional advertisement (left) contains visible promotional text, while the ‘PixelBlitz’ advertisement (right) conceals an instruction in one-point white text, invisible against the white background.

Key facts

• The conventional advertisement (left) — includes: visible promotional text, while the ‘PixelBlitz’ advertisement (right) conceals an instruction in one-point white text, invisible against the white background