GEO claims that held up
What we are solving
Section titled “What we are solving”A GEO playbook — GEO meaning getting your product named in AI assistants’ answers — landed in my chat in October 2026. It was full of percentages. Some came from studies, some from people selling GEO services, and the text did not say which was which.
I did not want to plan a quarter of work on numbers I could not trace back to someone who had actually measured them. So I checked every one.
- I pulled 124 factual claims out of the playbook and looked for each one’s primary source: the study, the official documentation, the court ruling, or the seller’s own page.
- A claim counted as confirmed only when that source was open to anyone and said the same thing. Where it said almost the same thing, the correction is in the table.
- That left 64 confirmed claims. The playbook repeats itself, so after merging the repeats there are 50 rows below, grouped by topic.
- Every row links to the source I read. Use them as facts to plan on, and treat everything else in such playbooks as ideas to test on your own site.
What did not work
Section titled “What did not work”My first pass left a dozen numbers marked “not checked”. The sources were spoken in videos, posted on Reddit and later deleted, or sitting on pages that had moved.
I went back through video subtitles, Reddit archives and the Wayback Machine. Most opened.
One figure was right but sat in a different report than the playbook named, and it measured channels, not videos. I only caught that on the second pass. Figures that exist only in a vendor’s webinar stay out of this page, because nobody outside can repeat them.
Verify
Section titled “Verify”Pick three rows and open their links. The number should be on the source page, with the same scope written next to it, and if the scope there is narrower than in the row, the row is wrong. Open an issue and I will fix it.
These figures age fast: the data in most rows covers mid-2025 to autumn 2026. If a link no longer shows the number, trust the updated date at the top more than the row.
How an assistant builds an answer
Section titled “How an assistant builds an answer”| What the source shows | Source |
|---|---|
| Claude takes 79.2% of its web citations from Brave’s top 10, 34% from Google’s top 10, and 36.4% from third-party “best X” lists. | Blyskal and Singh, State of AEO 2026 |
| Only 11.9% of URLs cited by AI assistants rank in Google’s top 10 for the user’s own question. Sample: 15,000 long queries, July 2025. | Ahrefs |
| ChatGPT turns a question into 1–7 search queries of 5–7 words; Google’s AI Mode into 10–40. | Mike King, Josh Blyskal |
| Claude adds the current year to 94% of its sub-queries, ChatGPT to 17%. | Blyskal and Singh |
| ChatGPT’s cited results overlap with Google’s results at about 19%. | Josh Blyskal at BrightonSEO |
| Through the API, Claude Sonnet 4.6 searched the web on 825 of 1,000 questions, DeepSeek-v4 on 584, GPT-5.3 on 140. The last one is the API model, not the ChatGPT app. | Max Planck Institute, arXiv 2609.19244 |
| Between under 1% and 5% of cited URLs come from the model’s memory; the rest come from search. Studied: ChatGPT, Claude, Grok, DeepSeek. | arXiv 2609.19244 |
| Links per answer on average: ChatGPT 4, Claude 3.5, DeepSeek 4.9, Grok 0.7. | arXiv 2609.19244 |
| AI crawlers other than Google’s do not execute JavaScript. OpenAI, Anthropic and Perplexity were checked. | Vercel and MERJ |
| It follows that text inside a PDF, an image or a script-loaded widget is invisible to those crawlers. | Vercel and MERJ |
| OpenAI’s goal is to answer 80% of ChatGPT queries from its own search index. The playbook said traffic; the ruling says queries, and the date has slipped. | US v. Google, Doc 1436, p. 143 |
Whose sources each assistant uses
Section titled “Whose sources each assistant uses”| What the source shows | Source |
|---|---|
| Of 253 frequently cited domains, all five AI systems cite only 23. Based on 5.31 million citations, June 2026. | Vercite |
| Brand websites make up 44.7% of ChatGPT citations, 59.8% in Google’s AI Overviews and 28.9% in Perplexity. | Otterly, February 2026 |
| After 7 May 2026 the homepage’s share of ChatGPT referrals rose from 26–32% to about 60%. Profound saw 28% go to 62%. | Similarweb, Profound |
| Wikipedia is ChatGPT’s most cited domain. 67% of its 1,000 most-cited pages are out of a brand’s reach: Wikipedia 29.7%, other companies’ homepages 23.8%, app stores 6.6%. | Ahrefs |
| Third-party “best X” lists make up 43.8% of pages ChatGPT cites. | Ahrefs |
| Reddit’s share of ChatGPT citations fell from 3.83% on 1 August to 0.24% on 14 August. | Pierview |
| In summer 2026 Perplexity produced 71% of all Reddit citations: 31,640 of 44,488. | Pierview |
| Google’s AI Mode takes half its citations from 71 domains; YouTube is the top external one. | Vercite |
| Inside Gemini’s 50 most-cited domains, Reddit has 28.5% and YouTube 14.6%. This is a share of the top 50, not of all citations. | Ahrefs, September 2026 |
| On commercial queries Reddit is 0.1% of Gemini’s citations. | Tinuiti via MediaPost, January 2026 |
| Inside Perplexity’s 50 most-cited domains, Reddit has 21.6% and YouTube 20.8%. | Ahrefs |
| A brand built from zero was visible in 30.6% of Perplexity answers, 8.3% in ChatGPT and 0.6% in Gemini. | SE Ranking, September 2026 |
| Claude and ChatGPT share 8% of cited domains. ChatGPT and Perplexity share 11% over 100,000 prompts, but that data is from July 2025. | Blyskal, Profound |
Citation, mention, customer
Section titled “Citation, mention, customer”| What the source shows | Source |
|---|---|
| People click a link inside a Google AI summary on about 1% of visits. Ordinary results get 8% of clicks with a summary on the page and 15% without. | Pew Research, July 2025 |
| A test brand got 42,435 citations in August 2026 and 100 site visitors over January–September. “Best X” lists named it in 14.2% of answers, a 6,000-word guide in 69.2%. The guide is one article. | SE Ranking |
| In the same test, external sites collected about 19,000 citations and the brand’s own site under 2,000. The 11 external sites belonged to the authors. | SE Ranking |
| Across 75,000 brands, web mentions correlate with AI visibility at 0.66–0.71 and YouTube mentions at about 0.74. Backlinks correlate very weakly. Correlation, not proof of cause. | Ahrefs |
| 35% of channels behind cited YouTube videos had under 10,000 subscribers. Based on over 100 million citations in 30 days. | Otterly |
When to expect results
Section titled “When to expect results”| What the source shows | Source |
|---|---|
| Half of cited pages are less than 13 weeks old. | Profound, September 2026 |
| Citations of a page halve in about 4.5 weeks: 3.4 in ChatGPT, 5.8 in Perplexity. Based on 3.5 million cases. | Scrunch and Stacker |
| A new short article got its first citation after a median 21 days; a how-to guide after 126.5 days. One campaign, one brand. | SE Ranking |
| Posts paid for at $14 showed up in AI search within an hour. One case, one platform. | Fudan University, arXiv 2610.11932 |
| An invented term was cited by 8 of 10 AI search systems within 7 days. In the worst case, 40 posts on the author’s own WordPress got 0 citations. | arXiv 2610.11932 |
| One article on a platform the engine favours did more than 20 posts on other platforms. The threshold differs by topic. | arXiv 2610.11932 |
Your own site
Section titled “Your own site”| What the source shows | Source |
|---|---|
| When the page title matches the question, ChatGPT cites the page 20.1% of the time; when it barely matches, 9.3%. | AirOps via DesignRush |
| Basic schema markup gave no lift in citations: 1,885 pages with it against about 4,000 control pages. | Ahrefs |
Google ignores llms.txt and needs no special schema for AI features. Of about 38,000 llms.txt files, 97% got no requests. | Google, Ahrefs |
| Hundreds of templated pages per query variant, including self-made “best X” lists, count as scaled content abuse. | Google spam policies |
| Bought links and private blog networks get neutralised. | Google spam policies, link spam update |
| Google’s own AI guide recommends keeping Merchant Center and the business profile current. | |
Google-Extended controls Gemini training and grounding. AI Overviews go through Googlebot, so the token does not decide whether you appear there. | Google’s crawlers |
| UTM tags are for counting visits in analytics. They do not affect whether a page gets cited. | Google Analytics |
| Search Console got a report on impressions in AI features in June 2026. Since 31 August 2026 every site has it. |
Crawlers and access
Section titled “Crawlers and access”| What the source shows | Source |
|---|---|
OpenAI runs OAI-SearchBot for search, ChatGPT-User for fetches on a user’s request, and GPTBot for training. robots.txt may not apply to ChatGPT-User. | OpenAI |
Anthropic runs ClaudeBot for training, Claude-SearchBot for search and Claude-User for user fetches. Anthropic warns that blocking by IP may not work reliably. | Anthropic |
Cloudflare’s managed robots.txt adds Disallow lines for GPTBot, ClaudeBot and Google-Extended on its own. A site can be closed to crawlers without the owner touching robots.txt. | Cloudflare |
False claims and trust
Section titled “False claims and trust”| What the source shows | Source |
|---|---|
| Press releases that cost $6 and $66 described a made-up ceremony, and Google’s AI Overview retold it as a real event. A demonstration, one case. | King of AEO |
| For a made-up brand, Gemini and Perplexity repeated other people’s inventions in 37–39% of answers, ChatGPT in under 7%. A Medium post beat the brand’s own FAQ. | Ahrefs |
| A Munich regional court held Google liable for false statements in its AI Overviews, case 26 O 869/26. It is a preliminary injunction from 28 May 2026 and not final. | The Decoder |