Le Monde has access to exclusive data provided to it by OpenAI and which reflects how articles perform in ChatGPT, explains Justine Werbrouck, marketing and strategy director at M Publicité.
JDN. For its GEO offer, Le Monde highlights the fact of measuring in a deterministic manner the themes that work in GEO. What do you mean by that?
Justine Werbrouck. Today, there are a number of tools on the market (Peec AI, Profound, Meikai, etc., editor’s note) which allow advertisers, as well as agencies and publishers offering GEO offers, to carry out a brand audit in conversational engines. These methods (which are based on feedback via API, browsers, panels) provide probabilistic and non-exhaustive data. However, on all advertising platforms, such as Google or Meta, advertisers are used to dashboards with deterministic data. At Le Monde, our data is deterministic because we know how our articles “live” in the responses of conversational engines. In addition, this data is, for the moment, exclusive.
Concretely, what exclusive data does OpenAI send you to allow you to manage your GEO offer?
I can tell you, for all the articles in Le Monde, how they live in the conversational engine. We can also know how Le Monde articles perform in ChatGPT responses and correlate them with specific themes that interest our advertisers. It is therefore deterministic data because it is provided by OpenAI and from responses formulated by ChatGPT users. The goal for us is then to produce content for our advertisers based on these lessons so that they can be reflected in the ChatGPT responses. Once we produce brand content for our advertisers based on these analyses, we can then know how the article lives in ChatGPT via OpenAI data. We will know if it works or not. Our agreements with OpenAI date from 2024. It has been two years since our data teams were able to identify patterns that work and those that do not.
What data do you provide to advertisers who subscribe to your GEO offer?
We aggregate the deterministic data we have access to to provide trends. We select 20% of the articles from Le Monde which perform best in ChatGPT to carry out a statistical and semantic analysis. For the statistical aspect, we look for correlation links in order to understand what the articles that perform well have in common and the links between the articles and specific themes with regard to our client’s brief, its brand, its sector, its competitors. For example, we cross-reference high-performance articles with the BBC’s “user needs” (get informed, act, get inspired, understand), to find out what needs these articles meet.
Thanks to the linear reduction statistical method, we can determine whether trends consistent with the advertiser’s brief emerge. At the same time, semantic analysis comes into play using AI: we look at, on the top 20% of articles that appear the most, which lexical and thematic fields come back. Finally, as this data is based solely on the editorial content of Le Monde, we cross-check all our conclusions from this first audit with the data from the Profound audit.
All this allows us to learn lessons in substance and form. For a given sector, we will know the most promising themes and angles among those on which the advertiser wishes to position itself. We will also look at how to title, construct and sequence paragraphs, place (or not) figures, sources, quotes, etc.
Have you been able to test the impact of your GEO in terms of mentions obtained for you and for your partner advertisers?
Our campaigns are still in progress, we are in the production phase and we do not yet have a report to share. These processes last at least three months while the first deals we signed extend over the longer term. We have around fifteen advertisers and the first campaigns were activated a month ago.
Is your production offered to all readers of Le Monde?
The articles that we produce are hosted on the Le Monde website. They are also mediated with advertising, which pushes them within the World and social networks, so that they can be read by humans and not just by AIs.
