In March 2026, an Ahrefs study of 863,000 SERPs and 4 million URLs appearing in Google's AI Overviews found that only 38% of the pages cited by the AI also rank in the top 10 classic organic results, down from around 76% in July 2025 (Ahrefs). Ranking well on Google no longer guarantees being cited by the AI. These ten rules cover GEO, generative engine optimisation: how a piece of content gets cited by ChatGPT, Perplexity, Gemini or AI Overviews rather than merely ranked. Here are ten golden rules for 2026.
What are the 10 golden rules of GEO in 2026?
1. Citation matters more than ranking
A page can dominate Google and remain invisible in ChatGPT, Perplexity or AI Overviews, or the reverse. Organic ranking is still useful, but it no longer predicts citation in a generative answer.
2. Precise statistics remain the most powerful lever
Adding precise, sourced statistics to a piece of content remains the tactic that increases cited visibility the most, by around 41%. This is the founding mechanism of the field, established by Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI (KDD 2024) and confirmed at larger scale in 2026 on more than 12,500 queries and 8,000 domains, with results cross-checked by BrightEdge, Semrush and HubSpot (GEO Benchmark Study 2026).
3. Citing your sources works, without overstating its reach
Adding inline references to external sources remains one of the most effective tactics, with gains of up to 115%. That figure applies specifically to content that started lower in the rankings, and does not extend to every page that adds a source (a nuance documented in 2026).
4. A quote from a named expert makes a real difference
A quote from a named expert carries far more weight than an anonymous claim. The measured gap is clear: 4.1 AI citations on average for a page that contains one, against 2.4 for a page without, close to +71% (2026 data).
5. The answer is decided in the first 200 words
In position 1, a page is cited by generative engines 33% of the time; in position 10, only 13%. AI systems favour the statistics, quotes and structured data present in the first 200 words (2026 data). Content that delays its answer loses most of its chances of being cited.
6. Schema markup alone earns nothing
JSON-LD helps an engine parse a page; it does not make the page more citable on its own. Across 1,885 pages that added it in 2026, no notable citation gain was measured (Ahrefs study, 2026).
7. Optimised content does not replace brand authority
Generative engines first choose the brand to recommend from what they already know, then look for a source to cite that justifies the choice (six independent tests, 2026). If this mechanism holds at larger scale, optimising content will never be enough to create a citation where brand authority does not yet exist. Generative engine optimisation cannot substitute for that: reputation has to exist before the content, and the content alone will not create it.
8. Separate training bots from search bots in robots.txt
GPTBot or ClaudeBot feed model training; OAI-SearchBot, PerplexityBot or Claude-SearchBot feed real-time answers. Blocking them all by reflex is expensive: 41% of B2B sites were still blocking at least one major AI bot in the first quarter of 2026, a choice that would cost between 18% and 34% of potential citations on the engine concerned (2026 data).
9. An llms.txt file complements robots.txt without replacing it
Proposed by Jeremy Howard in 2024 and popularised by Mintlify among others, the llms.txt format gives AI agents a structured index of the pages to favour. It spread in 2026 as a complement to robots.txt, which remains the file that allows or blocks access (technical reference, 2026).
10. Most fashionable “conversational” tactics do not work
The majority of the “conversational SEO” tactics in circulation have no measurable effect, some are harmful, and the raw relevance of the source is what keeps working (2026 measurement).
| Rule | What the figures say |
|---|---|
| 1. Citation before ranking | Only 38% of pages cited by AI are also in Google's top 10 (Ahrefs, March 2026) |
| 2. Precise statistics | Up to +41% cited visibility: the most powerful lever |
| 3. Inline cited sources | Up to +115%, mostly for content starting low |
| 4. Named expert quote | 4.1 AI citations on average against 2.4 without (+71%) |
| 5. The first 200 words | 33% citation in position 1 against 13% in position 10 |
| 6. Schema alone | No gain measured across 1,885 pages (Ahrefs 2026) |
| 7. Brand authority | The brand is chosen before the source that justifies it |
| 8. Differentiated robots.txt | Blocking everything would cost 18% to 34% of potential citations |
| 9. llms.txt as a complement | An index for AI agents: robots.txt keeps control of access |
| 10. “Conversational” tactics | No measurable effect: source relevance comes first |
AIxH's view
These ten rules reward verifiable proof and authority already built, and give nothing for presentation alone. As a GEO agency in Luxembourg, this is exactly the logic of our SEO & GSO service: a method that separates what genuinely earns visibility in generative engines from what belongs to technical folklore. If you want to know which of these ten rules your site already follows, that is the first question we ask.
“Of ten GEO tactics doing the rounds, seven are technical folklore. The three that remain, sourced figures, verifiable proof and real authority, are exactly the ones that require work. That is why they work.”