Research after research?
Ecosia is not a new tool, it has been around since 2009. It is a non-profit search engine, based on Bing and Google, whose advertising revenue funds reforestation and renewable energy projects around the world. Many have already heard of it, some have been using it for years. What has changed recently is that Ecosia has integrated two AI features: “Overviews” at the top of search results and an AI Chat in conversational mode.
I spent part of the week researching from Ecosia to observe these two features. My first instinct was to check the quality of the results and sources and see if their relevance was comparable to what you get elsewhere, since Ecosia aggregates Bing and Google. The results seemed very convincing to me. But what kept me more busy was the way the AI fits into the experience.
Overviews are displayed at the top of the results in the form of summaries with links to the sources. The first thing I noticed is that they are enabled by default but can be disabled in the settings. This is a design decision that is rarer than it seems: most engines that have integrated AI summaries do not offer this option, or make it difficult to find. Here, it is in the settings, accessible and assumed. You may want to search without a summary and it is expected.
What Ecosia offers, Google offers in 200 countries but not yet in France, the group indicated that the “AI Overviews” have not been launched in France because of “legal uncertainties”, without further details. No deployment date has been confirmed, or even if the AI Overviews would be different. These uncertainties could be linked to related rights: since 2019, France has transposed the European Copyright Directive strictly, with reinforced protections for press publishers on the reuse of their content. Ecosia, relying on Mistral Small and a European infrastructure, could find itself in a different regime and offer its overviews where Google would not yet be ready?
What is at stake behind the issue of overviews is also that of traffic. In the countries where Google has deployed them, the decreases in click-through rates observed range from 20 to 40% on average for affected queries, and over 60% for very informative queries, as detailed by Thomas Skowronski in a column published on Minted to be found in the shared reads: the user gets his answer in the results page, without clicking on any link. The study by the Media Review / INA shared in Letter 37 showed this from a close angle: chatbots redistribute traffic to the media according to their own editorial logic, far from being neutral. If summaries replace clicks, the visible surface of sources is even smaller, in addition to the filtering already operated by ranking algorithms. Before overviews, the first page displayed a lot of links. With a synthesis, only three to five sources are put forward. The question is not new, but it is changing scale.
Le The AI Chat works like a conversational search assistant: you ask a question in natural language, it searches the web, you refine, you relaunch, you can add documents. Here too, something broader is happening. Since search engines have existed, we have been the ones who have learned to speak their language: the keyword, the quotation marks to force an exact expression, the plus or the dash of the “minus something” to integrate or exclude a term. We adapted our way of formulating to what the machine could process. With the AI Chat integrated into the search engine as in Ecosia, this adaptation is reversed: you can formulate as you think and it is the model that does the work towards the sources. I think this is not trivial: we no longer search in the same way, and we may no longer think in exactly the same way as we do when we are no longer constrained by the format of the keyword. It seems to me that the keyword requires more synthesis and reformulation.
Part of what we lose seems to me more difficult to discern. We already lose the page itself: its design, its editorial angle, its own way of staging a subject… and even somewhere the thinking of its authors or bearers… Will we finally still need web pages if the trend continues or simply databases? Would a return to the display of HTML texts finally become sufficient to feed AI responses?
We also lose serendipity, that moment when we come across something we weren’t looking for, that has arisen in the list made up by the algorithm, this subject that we didn’t even know existed. The chatbot responds to what we ask of it. It doesn’t show what we hadn’t planned to look for.
Apart from the two AI features, what I also remember about Ecosia is the coherence of the ethical approach. It has been integrated into the technical choices: model, hosting, default settings, data policy. Most of the time, these choices are invisible to the user, buried in pages of documentation that no one reads. Here, they are highlighted and explicitly claimed.
The two AI features are based on Mistral Small hosted on a European infrastructure. Mistral is a European, French model, and choosing to back up its AI features when you are a search engine whose mission is explicitly ethical, is part of the overall coherence. The choice of Mistral Small and not a heavier model is also a choice of sobriety, consistent with an approach that claims frugality.
These are not spectacular commitments, but they are precise and meaningful commitments.
The real question that is finally asked is not “Is Ecosia better than Google?”, it uses part of the search architecture anyway, but rather a question that is perhaps a little more annoying for us: when we choose a tool, do we also really choose its values? And when this same tool frees us from the keyword to let us speak normally, do we even notice that something has changed in the way we search and in what we may no longer be looking for?
Echoing
→ Letter 36, What remains when the tool has done the work?
→ Letter 34, Is this information plausible?