✦ Key takeaway
In September 2026, we re-ran the same panel as our July study: more than 100,000 sources cited by ChatGPT, 20+ industries, 100% French-language queries. In three months, the hierarchy of sources changed dramatically. Reddit, the most cited domain in July with 43% of the Top 10, has dropped out of the ranking. Government websites (.gouv.fr) now lead on their own, with 60.8% of the Top 10. Google and Que Choisir are rising sharply, and four new domains have entered the ranking: PagesJaunes, Frandroid, Vogue and Les Numériques. The Top 10 now accounts for 15.4% of all citations, up from 7.2% in July. The long tail still represents the vast majority, but concentration is increasing.
In July, our study “How does ChatGPT choose the sources it cites in France?” revealed a hierarchy dominated by Reddit and government websites. Three months later, we measured the cited sources again on exactly the same panel. The result is clear: the sources cited by generative AI do not form a fixed ranking. They change fast, and sometimes abruptly.
Methodology: the same panel, two months later
To ensure a reliable comparison, we kept the July scope unchanged: more than 100,000 cited sources, 20+ industries, 100% of queries written in French, and ChatGPT as the conversational engine.
Citations are still aggregated by root domain: support.google.com and developers.google.com count towards google.com, and all government websites are grouped under gouv.fr. Brand websites are excluded from this Top 10, in order to focus on third-party sources: institutions, platforms, media and review sites.
The new Top 10: gouv.fr takes the lead on its own
Domain | Citations | Share of the Top 10 | Share of all citations |
gouv.fr (government websites) | 7,108 | 60.82% | 9.38% |
google.com | 1,580 | 13.52% | 2.09% |
quechoisir.org | 1,308 | 11.19% | 1.73% |
pagesjaunes.fr | 364 | 3.11% | 0.48% |
frandroid.com | 279 | 2.39% | 0.37% |
wikipedia.org | 266 | 2.28% | 0.35% |
trustpilot.com | 244 | 2.09% | 0.32% |
vogue.fr | 201 | 1.72% | 0.27% |
lesnumeriques.com | 189 | 1.62% | 0.25% |
leboncoin.fr | 148 | 1.27% | 0.20% |
Total Top 10 | 11,687 | 100% | 15.43% |

“TOP 10 most cited sources in France - ChatGPT - September 2026”, Source: Search Ai
With 7,108 citations, government websites alone account for nearly 61% of the Top 10 and 9.4% of all citations in the panel. As in July, this figure hides a wide variety of sources: service-public.gouv.fr, economie.gouv.fr, ecologie.gouv.fr, entreprendre.service-public.gouv.fr and legifrance.gouv.fr are all distinct sources.
The same goes for Google, 70% of citations come from support.google.com. In other words, ChatGPT does not cite the search engine itself, but mostly Google's help documentation.
July vs September: the big shifts
Domain | July rank | September rank | Share of Top 10 in July | Share of Top 10 in September |
gouv.fr | 2 | 1 | 34.5% | 60.8% |
google.com | 3 | 2 | 5.5% | 13.5% |
quechoisir.org | 8 | 3 | 1.5% | 11.2% |
pagesjaunes.fr | - | 4 | - | 3.1% |
frandroid.com | - | 5 | - | 2.4% |
wikipedia.org | 5 | 6 | 4.3% | 2.3% |
trustpilot.com | 6 | 7 | 1.9% | 2.1% |
vogue.fr | - | 8 | - | 1.7% |
lesnumeriques.com | - | 9 | - | 1.6% |
leboncoin.fr | 9 | 10 | 1.4% | 1.3% |

“Evolution of the TOP 10 sources cited by ChatGPT - July vs September 2026”, Source: Search Ai
Three trends stand out. First, institutional sources are growing much stronger: gouv.fr moves from 2nd to 1st place. Second, Que Choisir posts the biggest gain in the ranking, climbing five places from 8th to 3rd. Finally, six out of ten domains appear in both studies: gouv.fr, google.com, quechoisir.org, wikipedia.org, trustpilot.com and leboncoin.fr form a relatively stable core.
Reddit, Le Monde, YouTube, Le Parisien: the big exits
Domain | July rank | Share of Top 10 in July |
reddit.com | 1 | 43% |
lemonde.fr | 4 | 5.1% |
youtube.com | 7 | 1.5% |
leparisien.fr | 10 | 1.2% |
Reddit is the most striking case. The most cited source in July, with 43% of the Top 10, the platform received only 0.10% of the total in September . Community content, which seemed to be one of ChatGPT's two main citation patterns, has almost disappeared from our panel.
General news media are also declining. Le Monde, 4th in July, now appears to only 0.06% of citations. YouTube follows the same path with 0.05%. Le Parisien is a special case: it drops out of the Top 10 even though its share of all citations remains fairly stable. It was not cited less; other sources simply grew faster.
Our data does not allow us to attribute these shifts to a single cause. What it does show is that a strong position in ChatGPT's answers can be reversed within a few weeks.
The newcomers: reviews, local search and specialist media
The four domains entering the Top 10 have one thing in common: they meet very concrete needs.
✦ PagesJaunes (4th): this business directory reflects a strong local focus. It echoes our July study, where the word “france” appeared in 41.8% of ChatGPT's sub-queries.
✦ Frandroid (5th) and Les Numériques (9th): two tech review sites, confirming the weight of comparison content in this sector.
✦ Vogue (8th): fashion and beauty bring a lifestyle media outlet into the ranking.
Together with Que Choisir and Trustpilot, review and comparison sites now play a central role. ChatGPT favours sources that help people decide: product tests, reviews, official information and local listings.
The Top 10 carries twice the weight: from 7.2% to 15.4%
In July, the Top 10 accounted for only 7.2% of citations. In September, it accounts for 15.4%, more than double. As a result, the long tail falls from 92.8% to 84.6% of citations.

“Weight of the TOP 10 in all ChatGPT sources in France - July vs September 2026”, Source: Search Ai
This concentration is largely driven by gouv.fr, which alone accounts for 9.4% of citations. Yet the conclusion of our July study still holds: more than eight out of ten citations happen outside the Top 10, across hundreds of different domains.
Query fan-out: ChatGPT is starting to target its sources with “site:”
To find information, ChatGPT breaks each question down into sub-queries: this is known as query fan-out. In our September panel, 3.0% of these sub-queries contain the “site:” operator. In other words, ChatGPT does not always just search for a topic: it sometimes restricts its search to a specific website, for example “site:impots.gouv.fr IFI 2026 barème” or “site:peugeot.fr”.
In most cases, it is not the user who asks for a specific website. Only 37% of the prompts concerned contained an explicit request such as “which site” or “where can I find”. In the remaining 63%, the question was simply about a topic, for example “Who is subject to the French real estate wealth tax?”, and ChatGPT decided on its own to restrict its search to impots.gouv.fr. The phenomenon is still marginal, but it is worth monitoring: it shows that, for certain questions, ChatGPT already knows which sources it considers relevant before it even starts searching.
By classifying the 139 prompts concerned, five main types of questions emerge:
Type of question | Prompts | Share of prompts |
Looking for resources (courses, videos, forums, blogs) | 37 | 27% |
Regulations, taxes, administrative procedures | 32 | 23% |
Purchases, comparisons, reviews | 31 | 22% |
Brand mentioned in the question | 27 | 19% |
Technical product advice | 12 | 9% |
Total | 139 | 100% |
✦ For regulatory or tax questions, ChatGPT locks its search onto official sources. service-public.fr is the domain most often targeted with “site:”, which also helps explain gouv.fr's dominance in the Top 10.
✦ When a brand is mentioned in the question, ChatGPT goes straight to its official website: peugeot.fr. If that website does not clearly answer the question, the brand leaves the answer to other sources.
✦ For purchase questions, ChatGPT targets comparison sites and retailers, adapting to the market.
For brands, the challenge is twofold: becoming the source ChatGPT chooses to target in their industry, and making sure their own website precisely answers the questions people ask about them.
What these results mean for your GEO strategy
First, AI visibility must be measured continuously. A one-off snapshot is not enough: Reddit went from 1st place to a marginal presence in two months. Only regular tracking of your strategic prompts can detect these shifts in time.
Second, do not rely on a single platform. A strategy built solely on Reddit in July would have lost most of its impact by September. Diversifying your presence remains the best protection.
Third, review sites, directories and specialist media are gaining ground. Taking care of your presence and reviews on Que Choisir, Trustpilot or PagesJaunes, and targeting the leading media in your industry, is becoming a concrete GEO lever.
Finally, the long tail still dominates. Even with a more concentrated Top 10, 84.6% of citations happen elsewhere. Identifying the sources actually cited in your own industry remains the starting point of any strategy.
Track how cited sources evolve in your industry
The sources cited by ChatGPT change fast. Want to know which ones matter in your industry today, and how they evolve from one month to the next? Request a Search Ai demo: we will show you a personalised analysis of your visibility in generative AI.
About the author

Antonia
R&D & IA, CPO Search Ai
Antonia is Chief Product Officer (CPO) of Search Ai, a Generative Engine Optimization (GEO) platform for measuring, managing, and optimizing brand visibility on generative AI engines such as ChatGPT, Gemini, Perplexity, ... With over 10 years of experience as an R&D engineer, she continues to work at the intersection of R&D and business, transforming AI advancements into concrete functionalities for marketing teams.
