Wikipedia barely registers: what Google AI Overviews cite

The gist in 4 lines: across 1,000 French commercial queries analyzed on August 12, 2026, Google shows an AI answer (AI Overview) in 58.6% of cases. These answers cite 7.6 sites on average, and 68.4% of those citations point to company sites talking about their own offering. Wikipedia is cited exactly once out of 4,466 citations. And 42.3% of the cited sites do not appear in the classic top 10: getting cited by Google's AI no longer goes through the historical rankings.

Methodology

We built a panel of 1,000 commercial and comparative queries in French, spread across 10 sectors (100 queries each): SaaS software, B2B services (agencies, accountants, lawyers, HR), e-commerce and products, tradespeople and local services, finance and insurance, real estate, health and wellness, training, tourism and restaurants, marketing and tech. The queries reproduce what a buyer actually types: "best crm for smbs", "plumber hourly rate", "cheap car insurance", "payroll software comparison" (in French).

Each query was submitted to Google France (desktop results, French language) on August 12, 2026 via DataForSEO's SERP API, with AI Overviews loading enabled. For each query, we recorded whether an AI answer was present and the full list of cited sources. 999 queries out of 1,000 produced a usable result.

The 100 most cited domains, which concentrate 37.3% of all citations, were then verified one by one and classified manually: brand site talking about its own offering, comparison or affiliate site, media outlet, directory, institutional site, social network or video platform. The rest of the ranking relies on lists and automatic rules.

The lessons

1. AI answers are already the majority on buying queries

585 queries out of 999 (58.6%) show an AI Overview. But the gap between sectors is spectacular: 98% for SaaS software, 90% for marketing and tech, versus 33.3% in tourism and restaurants and 38% in health. Where Google monetizes differently (Shopping block on products, local listings on restaurants and hotels), the AI answer steps aside.

2. Google cites companies, not encyclopedias

Out of 4,466 recorded citations, 68.4% point to brand and company sites talking about their own offering: software vendors, firms, schools, insurers, tradespeople. Media outlets weigh 5.7%, institutional sites 1.2%. Wikipedia: 1 citation out of 4,466, that is 0.02%. On buying intent, the myth that "AI only cites Wikipedia and the press" does not hold.

3. YouTube is the most cited site in France

311 citations, meaning more than 53% of AI answers cite at least one YouTube video. That is nearly 5 times more than the second domain in the ranking (tool-advisor.fr, 63 citations). Video has become a first-rank source for AI answers, including on B2B topics like invoicing software.

4. Six out of ten AI answers cite an intermediary

60.3% of AI Overviews cite at least one comparison site, review site, marketplace or affiliate site. In two sectors, these intermediaries capture more than a third of all citations: finance and insurance (35.2%, driven by lesfurets.com, lecomparateurassurance.com, meilleurtaux.com) and tradespeople (34.8%, driven by travaux.com, habitatpresto.com, ootravaux.fr). Conversely, only 8.7% of AI answers cite direct brands exclusively: a brand standing alone in an AI answer is the exception.

5. 42% of cited sites are invisible in the classic top 10

42.3% of the domains cited by AI answers (1,887 out of 4,466) do not appear in the organic top 10 for the same query. Our August 2026 barometer already measured 43.8% on a general-purpose panel: the confirmation is clear on buying queries. The historical Google ranking and AI citations are two distinct competitions.

6. The long tail has taken over

1,855 different domains are cited at least once. 63% are cited only once, and the top 20 domains capture only 19.7% of all citations. In other words: there is no citation oligopoly. SMBs unknown to the general public (an accounting firm, a rental management agency, a regional ERP integrator) are cited alongside the big names.

7. Each sector has its own citation ecosystem

In e-commerce, the specialized press weighs 31.6% of citations (lesnumeriques.com, 01net.com, clubic.com): product reviews remain the reference. In B2B services, 82.6% of citations go to company sites. For tradespeople, a third goes through matchmaking platforms. There is no single citation strategy, there is one per sector.

The sector table

What this changes for an SMB or a brand

Getting cited is within reach. This is the most counter-intuitive result of this study. 63% of cited sites are cited only once, the top 20 captures only a fifth of citations, and 42% of cited sites are not even in the classic top 10. An SMB with a well-crafted site can get cited on its queries, without the authority of a major media outlet.

Content that answers wins. The cited company sites share one trait: pages that answer the question precisely (a price, an honest comparison, a buying guide), not showcase pages. This is exactly what we detail in our guide on structuring content to get cited by AI.

Watch your intermediaries. In finance and home services, being absent from the cited comparison sites and platforms means letting a third party tell the market's story in your place. Depending on your sector, the battle is fought directly or through these relays.

Video matters, even in B2B. With YouTube cited in more than one AI answer out of two, even a modest channel on your niche topics is a gateway to citations.

Methodology and limitations

• Measurement taken on August 12, 2026, single pass, desktop only. AI Overviews evolve fast: these figures are a snapshot, not a constant.
• 1,000 queries built by us (10 sectors x 100), commercial or comparative intent. 999 usable results.
• Data collected via DataForSEO's SERP API (Google France, French language), with no AI-generated results whatsoever: every figure comes from a real results page.
• Source classification by domain: the 100 most cited domains (37.3% of citations) were verified manually one by one, the rest is classified through lists and automatic rules. A domain's category does not tell the whole story of the cited page's content.
• Counts deduplicate domains within each answer (a domain cited 3 times in the same answer counts as 1).

A ForgR study, conducted by Gilles Helleu and Paméla Michel, co-founders.

The raw data is public, so every figure in this article can be recomputed: the 999 queries with the domains cited for each, the 1,855 cited domains with their category, the per-sector summary. The full dossier covers the method and the domain ranking in detail.