Answer Engine

Anyone who searched for something on the internet in the past received a list of links. You clicked through several pages and pieced the answer together yourself. An Answer Engine takes this step off your hands. It reads the sources itself and writes a finished answer from them in complete sentences. Below it, there are usually small references to the websites the information came from. The German term for this would simply be Antwortmaschine, but the English term has become the established one.

What this does to the internet’s business model

A large part of the internet finances itself through clicks. A news site or a recipe blog earns money when people visit the page and see advertising there. Answer engines intervene exactly at this point. They use the content of these pages, but the user often no longer clicks through. Experts call this a zero-click search, meaning a search without any click on a result at all.

For Google, this is a delicate situation. The company has earned money through ads next to search results for decades. If its own AI displays the answer directly, the number of clicks decreases. At the same time, Google cannot afford to fall behind competitors like Perplexity or OpenAI here. This dilemma regularly comes up in financial news when Google’s advertising revenue is discussed.

On the other hand, a new market is emerging. Publishers are now negotiating licensing agreements so that their articles may be used by AI providers. Others sue instead for copyright infringement. Which path will prevail is still an open question.

The path from question to finished answer

At their core, Answer Engines work with a method called Retrieval Augmented Generation, or RAG for short. The process has three steps. First, the question is translated into one or more search queries. Then the system retrieves the matching passages from the web or from a database. Finally, a language model is presented with these texts together with the question and formulates the answer from them.

The decisive point is the middle step. The language model does not answer from memory but based on texts it has just read. This reduces the risk of so-called hallucinations, meaning freely invented claims that sound convincing. However, it does not eliminate them. If the sources are poor or the model summarizes them incorrectly, nonsense still ends up in the answer.

That is why the source citations below the answer are not decoration but the most important part. They are the only way to verify whether the answer holds up. A good answer engine can also be recognized by the fact that it admits when it has found nothing solid.

From Perplexity to Google’s AI Overview

The best-known standalone example is Perplexity, a service that was built as an answer engine from the very start. Google displays its variant as an AI Overview above the normal results. Microsoft has built the technology into Bing, and ChatGPT also searches the web on request. Voice assistants on mobile phones are also increasingly working according to this principle.

Within companies, the same thing happens on a smaller scale. There, answer engines do not search websites but internal manuals, contracts, or support tickets. An employee asks in plain language and receives the answer along with a reference to the correct document.

In marketing texts, one increasingly encounters the abbreviation AEO for Answer Engine Optimization. It refers to the attempt to write content in such a way that AI systems cite it. This is the successor to classic search engine optimization. How well this actually works is disputed among experts.