AI 引用調査：AI が最も多く引用しているプラットフォームはどこか（301025 件の回答を測定）

The short answer

Across 301025 AI answers collected between August 2025 and August 2026, Trustpilot is the most cited platform of the panel, with a citation rate of 2.7%. The ranking changes completely from one model to another: Wikipedia is cited in 3.43% of ChatGPT answers and 0.03% of Claude answers. Only 6 platforms are cited by all five models.

There is no single way to be cited by the AIs. There are five citation diets, and a platform that carries one model can be almost invisible on another.

Key figures

• 301025 AI answers analysed between August 2025 and August 2026
• 1820738 citations extracted from those answers
• 131349 distinct domains cited at least once
• 5 models measured separately: ChatGPT, Gemini, Perplexity, Claude and Grok
• 79% of the questions asked in French, the rest in English

Which platforms the AIs cite most

• Trustpilot: 2.7%
• YouTube: 2.35%
• Wikipedia: 1.62%
• Reddit: 1.48%
• Amazon: 1.12%
• TripAdvisor: 1.08%
• Microsoft: 1.03%
• Shopify: 0.91%
• Facebook: 0.85%
• LinkedIn: 0.71%
• Salesforce: 0.5%
• HubSpot: 0.47%

The same platforms, model by model

• Trustpilot: 2.7% on ChatGPT, 4.33% on Gemini, 1.21% on Perplexity, 1.57% on Claude, 1.8% on Grok
• YouTube: 0.04% on ChatGPT, 3.69% on Gemini, 3.46% on Perplexity, 0% on Claude, 2.76% on Grok
• Wikipedia: 3.43% on ChatGPT, 1.38% on Gemini, 0.19% on Perplexity, 0.03% on Claude, 1.08% on Grok
• Reddit: 0.33% on ChatGPT, 2.5% on Gemini, 1.51% on Perplexity, 0% on Claude, 4.11% on Grok
• Amazon: 0.9% on ChatGPT, 1.96% on Gemini, 0.4% on Perplexity, 0.49% on Claude, 3.03% on Grok
• TripAdvisor: 1.09% on ChatGPT, 0.73% on Gemini, 1.48% on Perplexity, 0.19% on Claude, 0.83% on Grok
• Microsoft: 1.14% on ChatGPT, 1.89% on Gemini, 0.19% on Perplexity, 0.14% on Claude, 0.09% on Grok
• Shopify: 0.96% on ChatGPT, 1.54% on Gemini, 0.35% on Perplexity, 0.08% on Claude, 0.09% on Grok

How to read these numbers

A citation rate of 2% means the source was cited twice for every 100 measured answers. Rates are computed per model against the number of answers of that model, so a model with a smaller sample never distorts a column. A source cited several times inside one answer is counted once per company block, which keeps a single verbose answer from dominating a month.

The distribution is worth as much as the ranking. The top three sources hold 45% of the citations counted in this table, 7 platforms sit above 1%, and Trustpilot and YouTube are close enough to be treated as one tier. A market where three sources hold most of the citations rewards depth on those three rather than presence everywhere.

What changed over 13 months

The largest rise of the period is Trustpilot at plus 2.34 points between the first and the last quarter measured. The largest fall is Wikipedia at minus 7.94 points. Three platforms moved by more than one full point, which is why a ranking older than a quarter is a poor basis for a decision.

What this means in practice

The practical consequence is that a platform budget decided on an average is a budget decided on a number that describes none of your buyers. If your market asks Gemini, video and review platforms carry most of the citations. If it asks ChatGPT, encyclopedic and institutional sources still dominate. The right question is not which platform the AIs prefer, it is which model your buyers actually open before they choose a supplier.

How to use this study

1. Ask your last 20 customers which assistant they used before contacting you, then read the column of that model rather than the average.
2. Claim and complete your profile on the two platforms that lead that column before producing any new content.
3. Re-measure after 60 days: platform positions in this dataset move by more than one point in a single quarter.

How this study was measured

1. We run a panel of buyer questions every month, the questions people type before choosing a supplier.
2. Every question is sent to five AI: ChatGPT, Gemini, Perplexity, Claude and Grok.
3. Every answer is parsed, every cited source is extracted and normalised to a domain and a platform.
4. The citation rate is the number of citations of a source per 100 measured answers, computed per model so an uneven number of answers per model never distorts a column.
5. The month is closed, the dataset is rebuilt, and the public files are republished on data.sorank.com.

Limits of this study

1. This panel is not a neutral sample of the web. The questions come from the sectors our users operate in.
2. The panel leans French: 237930 answers come from questions asked in French and 63095 from questions asked in English.
3. Model coverage is uneven: 87008 answers from ChatGPT, 87004 from Perplexity, 87000 from Gemini, 20011 from Claude and 20002 from Grok. The two smallest samples are read as direction rather than as precise values.
4. A citation is not a click. This study measures what the AIs cite, not what a reader does next.

The dataset behind this study

Every number on this page comes from the public Sorank dataset. The underlying figures, the monthly detail and the filters are available here:

Open the source data on data.sorank.com

Data updated on 2026-08-26.

How to cite this study

This study is free to quote, in an article, a deck or a report. Please cite the source and the measurement period, because the numbers move every month.

Sorank (2026). AI Citation Study: Which Platforms the AIs Cite Most (301025 Answers Measured). Sorank GEO studies, measured on 301025 AI answers collected between August 2025 and August 2026. Available at https://www.sorank.com/data/ai-citation-study-platforms

Related studies

The same panel, read from another angle:

• ChatGPT Citation Study: What ChatGPT Cites
• Trustpilot AI Citation Study
• Wikipedia in AI Answers: Decline Study
• All GEO studies

Measure your own citations

This study describes a market. Sorank measures your site: which questions mention you, which mention someone else, and what to publish to change that, tracked weekly across the same five AI.

Start with 3 days free. A payment card is required to open the trial.