AIs focus on the first part of your content (and then they become tired)

Kevin Indig is an SEO specialist who has worked for highly influential companies throughout his career, such as Shopify, Airbnb, and Snapchat, and who regularly publishes content of great interest from Growth Memo. On this occasion, he has conducted a study based on 1.2 million ChatGPT query results, with the objective of analyzing how AI models like this one decide which sources to cite and which to disregard.

And well…the conclusion is clear: AI is an impatient reader. If you want your content to be its reference source, you should be aware of this.

The 30% rule: the important information goes first

Indig’s study reveals the existence of a model he refers to as the “ski slope.” Thus, almost 44% of AI citations come from the first third of the content. As the model continues reading, its attention diminishes substantially, with 31.1% coming from the middle portion of the content (30–70%) and 24.7% of citations stemming from the final third.

In other words, if you want ChatGPT to cite the answer from one of your articles in response to a query, you should include it in the initial paragraphs. As the author explains, “capture the ‘Who, What, Where’ from the beginning. If your key idea is in the introduction, the likelihood of citation is high.”

Viewed in this light, we might say that AI rewards efficiency, not literary flair. Interestingly, although AI prefers the beginning of the page, within a paragraph, it tends to extract information from the middle portion (53% of citations).

The “clarity tax”

One of the noteworthy aspects of the study is the concept of the Clarity Tax, which could be described as the price creators must pay in the form of simplicity. Therefore, in order for ChatGPT to understand and validate your content, you should opt to use headings (H2, H3) that are straightforward questions, in addition to implementing structured data: pages that use it are quoted 1.7 times more than others.

Moreover, AI appears to favor direct definitions. Phrases that begin with “X is…” or “X means…” have almost double the likelihood of being cited compared to lengthy and ambiguous explanatory paragraphs.

Additionally, the study highlights the importance of eliminating empty marketing language. AI looks for entities (names, dates, key concepts) and pertinent data, once again disregarding what it considers filler content. Specifically, there are five best practices to follow:

1. Definitive language: When a user asks, “What is X?”, the model seeks the strongest vector route, which is almost always a direct sentence structure such as “X is Y.”
2. Conversational question-and-answer structure: Cited text is twice as likely (18% versus 8.9%) to contain a question mark. In the author’s words, an example of a losing structure would be “The history of SEO”> It began in the early 1990s…</p> By contrast, an example of a winning structure (78%) would be “When did SEO begin?” > SEO began in… (Direct answer)”
3. Entity richness. A sentence containing three entities provides more information than a sentence without any, so a structure such as “The best tools include Salesforce, HubSpot, and Pipedrive” is rewarded over “There are many good tools for this task.”
4. Balanced sentiment. AI is not seeking boring Wikipedia text, nor pure opinion. It is looking for the expert opinion of an analyst, preferring phrases that explain how a fact is applied, instead of simply mentioning a statistic. Does this remind you of the principles of EEAT in Google SEO?
5. Simple writing. AI appears to favor simple subject-verb-object structures with sentences that are short or not too long, as it is easier for the model to extract information from them.

The author explains these ChatGPT behaviors by noting that LLMs, such as this one, have trained on journalism and academic articles, in which the most highly weighted information typically appears at the top. On the other hand, although modern models can read up to 1 million tokens in a single interaction (approximately 700,000–800,000 words), “their goal is to establish the framework as quickly as possible and then interpret everything else through that framework.”

In any case, and this is a personal opinion, if you intend to put these concepts into practice, you should consider both what you may gain… and what you may lose. After all, the CTR from citations you may obtain from ChatGPT is volatile and questionable and, ultimately (and as Google reminds us every time it discusses GEO), you are writing (at least for the time being) for human readers, who may well appreciate other characteristics that are more free-form and less rigid in your writing style—quite different from what an AI values.

Image: Gemini