How to Optimize Content for LLMs and AI Search
What to change in the writing of a page so AI assistants use it as a source, what the data says about each tactic, and which kind of page shows up most in Google’s AI answers across two DoGeo projects.
Published 18 October 2026
Short answer
Optimizing content for LLMs means writing every section of a page so that an assistant such as ChatGPT, Gemini or Perplexity can lift it out on its own and cite it. In practice it comes down to five writing changes: answer within the first 40 to 60 words, one idea per paragraph with the subject named, an explicit definition, comparisons in a table and every figure with its source. What helps least is what gets sold most: Google says its AI answers require no special markup, and turning an article into an FAQ is associated with less influence on the answer, not more.
What is content optimization for LLMs
Content optimization for LLMs is the work of writing and ordering a page so an AI assistant can take a passage from it that answers a question on its own, and cite it.
The difference from writing for Google lies in what gets evaluated. A classic search engine ranks whole pages; an assistant such as ChatGPT, Gemini or Perplexity splits the pages it finds into passages, compares each one with the question and keeps the ones that answer it best. A paragraph that only makes sense after reading the one before is of no use to it, however good the article around it.
Each answer is built from few searches and quite a few sources. In a DoGeo study of 1,230 Gemini runs (October 2026), when the model searched the web it issued about 2 sub-queries per answer and used about 10 sources. That is a single model, Gemini with Google Search, but it shows the scale: a page is not competing for a slot in a list, it is competing with other sources to supply a sentence.
How each assistant decides which sources to look at before citing is a separate question, covered in how AI models choose their sources. Content optimization starts after that: with the page already found, what makes it quotable.
How to optimize content for LLMs step by step
Five writing changes, in this order, that you can ask of whoever writes your pages without touching the design or the technology of the site.
- 01 / ANSWER
Answer the heading’s question within the first 40 to 60 words
Each heading should read like a question someone actually asks, and the first paragraph under it has to answer it with no preamble. That is the passage an assistant finds easiest to extract. If the answer only arrives after three paragraphs of context, the assistant takes it from another site that gives it sooner.
- 02 / PARAGRAPHS
Write one idea per paragraph and name the subject in every section
A paragraph that mixes two topics half-matches two questions and wins neither; splitting it in two is the cheapest change there is. Within each section, the product, the company or the concept is named in full at least once: “the tool” or “this process” means nothing in a passage that travels alone.
- 03 / DEFINITION
Define the core concept with an “X is…” sentence
Every page about a concept needs a section of its own that defines it with an explicit marker: “X is…”, “X is defined as…”. It is the kind of sentence an assistant reuses when someone asks what something is, and the easiest one to attribute to a named source.
- 04 / TABLES
Put comparisons in tables and the year in comparison titles
A comparison written as prose has to be rebuilt; in a table with the same attributes on every row, it can be lifted whole. The current year belongs in the title of comparison pages, because some searches add it depending on the topic: in DoGeo’s Gemini study, between 11% and 15% of the recruiting sub-queries carried “2025” or “2026”, against 1% of the GEO ones.
- 05 / SOURCES
Pair every figure with who measured it and when
A figure with an owner can be cited; one without an owner cannot. The format that works is the figure, the entity behind it and the date in the same sentence. First-party data counts for more than third-party data because nobody else has it: a share of customers, a measured lead time or a count from your own Search Console.
Which content tactics work with AI assistants and which do not
Figures, definitions and comparisons are what is most associated with an assistant using a page; a question-and-answer format, by contrast, works against it.
| Page element | Difference in influence | What to do with it |
|---|---|---|
| Numbers and statistics | +61.55% | Replace qualitative claims with sourced figures |
| Definition markers | +57.33% | One “X is…” section for each core concept |
| Comparison content | +55.28% | Tables with the same attributes on every row |
| How-to content | +41.20% | Numbered steps when the page explains how to do something |
| Question-and-answer format | −5.74% | Do not turn the article into a list of FAQs |
That April 2026 study analysed 602 prompts run on ChatGPT, Gemini and Perplexity and 18,151 cited pages that could be downloaded. Pages in the most influential quartile averaged 10.59 headings and 1,943.3 words; pages in the bottom quartile, 0.85 headings and 169.8 words.
These are correlations, and the study says so itself: it does not treat it as established that adding a comparison will increase the citations of a specific page. What it does describe is what the pages assistants already use look like.
An academic study from 2024 of 10,000 queries reached a similar conclusion by modifying pages. Citing sources, adding quotations and adding statistics improved by 30% to 40% a visibility metric that measures how much of the source’s text reaches the answer and in what position. Repeating keywords, the classic SEO tactic, brought little to no improvement.
Which kind of page appears most in Google’s AI answers: Kulturo and HeyTalent
In two DoGeo projects, the pages that appear most in Google’s generative AI features are the ones that settle a specific question with information of their own: calculators, pricing and legal questions.
| Project | Kind of page that appears most | Share of that kind of page | Share of the top 10 pages |
|---|---|---|---|
| Kulturo (kulturo.tech) | Calculators for net salary, severance pay and net self-employed income | 54.9% between them; a single calculator, 37.3% | 82.0% |
| HeyTalent (heytalent.app) | A page that answers a pricing question | 29.8%; with a legal question and a list of tools, the top 3 add up to 56.1% | 72.3% |
At Kulturo, according to Google Search Console, the calculators account for 54.9% of impressions by page in the generative AI features, and a single one of them for 37.3%. A calculator returns a result that is not written down on any other site, which is exactly what an assistant cannot answer from memory.
At HeyTalent, according to Google Search Console, a single page answering a pricing question accounts for 29.8% of impressions by page. The top three (that one, a legal question and a list of tools) reach 56.1%, and the top ten 72.3%.
These are correlations from two sites over one quarter: they do not prove that a calculator or a pricing page causes the appearance. What they show is that impressions concentrate on a few pages, and that those pages answer a precise question rather than covering a topic in general. The full figures are in the Kulturo case study results and the HeyTalent case study results.
What role structured data and JavaScript play
A smaller one than is usually sold: Google says its AI answers need no special markup, and what they do need is for the text to be on the page.
Structured data is code, usually written with the schema.org vocabulary, that describes in machine language what a page says: who wrote it, what it sells, what it costs. In its documentation on AI features in Search, last updated on 10 December 2025, Google writes: “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.”
The same Google documentation asks for two things that do affect content: that structured data matches the visible text on the page, and that important content is available in textual form. A price that only appears in the markup, or a table that only exists as an image, meets neither.
JavaScript matters for the same reason. If a page’s text only appears after a script runs, a robot that does not run it receives an empty page. You can check whether AI crawlers can read your site in a minute and without giving an email; whatever needs changing on the crawling side is work for whoever runs the technical side, not for whoever writes.
How to check whether optimized pages start appearing in AI answers
Three sources give real data at no cost: the AI report in Bing Webmaster Tools, the one in Google Search Console and a log of AI robots in the site’s firewall.
The AI performance report in Bing Webmaster Tools counts how many times a site is cited and how many different pages. According to that report, Kulturo gathered 8,700 AI citations in three months, with an average of 10 pages cited, and HeyTalent 14,200 citations with an average of 19 pages. For a newly optimized page, the signal is that it joins that list of cited pages.
Google Search Console has a generative AI features report, still in beta, that gives each page’s impressions in Google’s AI answers. They are impressions, not clicks or citations, but they show whether a specific page starts being shown after the change.
Neither report says whether ChatGPT, Claude or Perplexity have read a page. For that, a rule is created in the site’s firewall or CDN that logs, without blocking, the requests of each AI robot, identified by its user-agent and, where the provider publishes its ranges, verified by IP. Those logs show which pages each robot visits and when.
Each kind of robot tells you something different. The search ones (OpenAI’s OAI-SearchBot, Anthropic’s Claude-SearchBot and PerplexityBot) say which pages enter each assistant’s index. The ones that open a page when a person asks (ChatGPT-User, Claude-User and Perplexity-User) are the closest signal that the assistant has read it in a conversation. The training ones, such as GPTBot or ClaudeBot, only speak to training.
This is what DoGeo sets up for each client. The log does not say what the assistant wrote in its answer: for that you still need a fixed set of prompts repeated every month.
Keep reading
Can your pages be cited today?
DoGeo reviews which sections of your site answer a question on their own, which ones mix topics or bury the answer, and which first-party data you have but are not using. It takes four details: your name, your company, your website and a work email.