GEO: getting into AI answers

GEO (generative engine optimization) is work to get pages cited in AI search answers. What probably helps: an answer in the first sentence, facts in a list.

When a text answers instead of a list of links

Ask a modern search engine “how do I find where my property boundary runs”. Instead of the familiar list of ten links you’ll often see a ready answer written as one coherent text, with small numbers that lead to sources. That’s how “AI Overviews” and “AI Mode” work in Google, how Copilot works in Bing, how ChatGPT search, Perplexity and other AI assistants with search work, and, in Russia, Yandex’s “Neuro”.

How they work is clear in outline. First the system finds suitable pages (by the same principles of crawling, index and ranking that you studied in the lessons on SEO). Then a language model, a program trained on huge amounts of text, words an answer from the pages it found and adds links. This design has a name: retrieval-augmented generation, or RAG (an article is in the sources). Google’s own guide uses the same words, and adds that its systems may fan a question out into several related searches to find more pages.

Two facts follow for you.

• If you are cited, you gain trust. The person sees that the source gave the answer and can go to you.
• If you aren’t cited, you are not in that answer at all. Ordinary results had places from first to tenth, but here there’s one answer, with three or four sources in it.

Another consequence: visits to the site may become fewer. The person got the answer and left (this is called “zero click”). So the value of getting into an answer isn’t only in clicks but also in the project’s name being heard.

What GEO is, and what is known about it

GEO (generative engine optimization) is the name for work so that your pages are used and cited in such answers. The term was introduced in 2023 by researchers led by Pranjal Aggarwal, in a paper presented at the KDD conference in 2024 (a link is in the sources). On a test set of questions they tried several techniques and reported that some raised the “visibility” of a page in answers, such as adding quotations, references to sources and statistics, while others did not, such as keyword stuffing. Remember this is a result on their set and their system. The paper doesn’t guarantee what will happen with real systems.

Now what no one knows except the companies themselves: how a particular system chooses its sources. Those rules are closed and change. So there’s a lot of noise around GEO: “secret techniques”, paid “rankings” and promises of “guaranteed inclusion in the answer”. Treat them like the promises of a “guaranteed top 1” from the lesson on SEO.

What do the companies themselves say? As of 28 October 2026, Google’s documentation says that to appear in AI Overviews and AI Mode there are no additional requirements and no special optimizations: a page must be indexed and eligible to be shown in Search with a snippet, and you don’t need new machine-readable files, AI text files or special markup. A newer Google guide, “Optimizing your website for generative AI features on Google Search” (published in 2026), says that from Search’s point of view optimizing for generative AI search is still SEO, and what it asks for is valuable, non-commodity content. It even names the terms “AEO” (answer engine optimization) and “GEO” and advises caution with third-party services selling them. This is the soberest starting point: GEO is SEO plus clarity. Claims about what companies do change quickly, so reread the sources on the day you work.

What probably helps

These are techniques worth using whether or not anyone confirms them. They make a page better for people and for machines, so you lose nothing even if it turns out they change nothing.

1. The answer in the first sentence. You know the inverted pyramid: the main thing first. The page’s first sentence answers the question of its heading. “A property’s boundaries are set by the recorded deed and plat, and a licensed land surveyor marks them on the ground”, not “Sooner or later many homeowners wonder”.
2. A heading that is a question where the page answers a question. That’s the wording the person types.
3. Facts in a list or a table. Prices, times, documents, addresses. A fact is pulled from a list more easily than from a paragraph, and a person checks it by eye.
4. An author and a date. Who wrote it, with what qualifications, when it was updated. Where an error is costly, as in law, this matters most: laws change, and an outdated answer does harm.
5. Links to sources (a law, a document, a study) and your own data: figures and examples others don’t have. They are what makes a source worth having.
6. Markup from the last lesson, matching what is visible on the page.
7. Access for bots. If the bot is blocked, citing is impossible.
8. Mentions of you elsewhere: professional communities, directories, articles. The clearer it’s across the web who you are, the easier it is to name you.

What not to do: stuff the text with keywords, hide text from people (hidden, display:none, white on white), or multiply hundreds of identical pages with generated text. All of that is spam, and search engines have fought it for a long time.

A toy model

Open the diagram. It’s a simplified model: it shows the direction, not real weights. Switch on features of your page and watch the answer change and see the moment your page starts to be cited. See how the deceitful techniques and a blocked bot act.

Your page

the first sentence answers the question facts are in a list or a table the author and the date of the update are given there are links to the law and the documents the text is stuffed with keywords part of the text is hidden from the person the search bot is blocked in robots.txt

Chance of being cited in this model: 20 out of 100.

You How do I find where my property boundary runs?

AI search

A property’s boundaries are described in its recorded deed and plat; a licensed land surveyor marks them on the ground. [1][2]

[1] a reference site, [2] a homeowners’ forum. Your page is not among the sources. The answer is buried in the text: it is easier to take a page where it is in the first sentence.

This is a toy model. Real systems are more complex and don’t reveal how they choose sources. We show only the direction: a clear answer, facts and honesty help, while deceit and a closed door hurt.

Check yourself: why does stuffing a page with keywords help neither SEO nor GEO?

Search engines recognized such text long ago and don’t count it as useful, and a language model picks a page from which a clear answer is easy to take. A stuffed text has no answer, and it’s hard for a person to read.

Which bots to let in

In robots.txt (the lesson on technical SEO) you can allow and forbid robots one by one. AI systems have robots with different jobs, and they are worth telling apart. By the documentation of OpenAI, Anthropic, Perplexity and Google:

• Training bots: GPTBot (OpenAI), ClaudeBot (Anthropic) and Google-Extended (a token that controls whether content Google crawled is used by some of its generative products, and has no effect on Google Search). If you block them, you signal that your pages shouldn’t go into model training.
• Search bots: OAI-SearchBot (ChatGPT search), Claude-SearchBot (Claude’s search), PerplexityBot, and the ordinary Googlebot and Bingbot. If you block them, your site won’t be shown in the corresponding search or in answers. OpenAI says that a site opted out of OAI-SearchBot won’t be shown in ChatGPT search answers, though it may still appear as a navigational link.
• “At the person’s request” bots: ChatGPT-User, Claude-User, Perplexity-User. They open a page when the person asks about it in the chat. For that reason OpenAI says robots.txt rules may not apply to ChatGPT-User the way they do to ordinary robots.

The decision is yours and depends on your aims. If you want to be found and cited, allow the search bots. If you don’t want to hand over your texts for training, block the training bots. An expert also has the question of copyright in their own texts. In the EU, the copyright exception for text and data mining (Directive 2019/790, Article 4) applies unless the rightsholder has reserved their rights in a machine-readable way, and a robots.txt rule is one way people do that; elsewhere the rules differ and are still being decided in courts and parliaments. If it matters to you, talk to a lawyer: this is general information, not legal advice.

Two honest caveats. A robots.txt file is a request, not a lock: companies say they honor it, but you can’t check that from outside. And bot names change: check them against the documentation on the day you read this, with links in the sources.

The llms.txt file

In 2024 the American researcher Jeremy Howard proposed an llms.txt file: a short description of a site in Markdown at its root, meant to help language models understand what is on the site and where. It’s a nice idea, and the description is at llmstxt.org.

The honest picture today: the big search engines haven’t confirmed that they read this file, and Google’s documentation says new special files aren’t needed. Making the file is cheap and does no harm. But don’t build a strategy on it. So it makes sense to write the decision down with a reason, as you’ll in the practice, and to come back to it if anything changes.

How to check that you are in the answers

• A series of questions. Write down twenty or thirty questions your audience asks (remember the “for whom” field). Ask them in several AI searches and record: were you mentioned, were you cited, what was said, was it right. Answers change from time to time, so one question isn’t enough.
• Sources of visits. In your counter (the part on metrics) see whether people arrive from AI assistants’ sites. Google also added generative AI performance reports to Search Console in 2026, so look in the consoles you verified.
• Wrong answers. If an AI says something untrue about you or your field, fix your pages so that the right answer is clear and has a source. Watch legal topics especially: an AI answer isn’t legal advice, and your site can say so outright.

Practice

Prepare two pages and the decisions.

1. Two pages. Each: a question as heading, a first paragraph of one sentence with the answer, an author and a date of the update, a list of three facts. If the brief has services, name them among the facts. No hidden text.
2. Decisions for bots. At least three rows: the bot, “Allow” or “Block”, and a reason.
3. The decision on llms.txt with a reason.

The check looks at form: a question mark, one sentence in the answer, a date in a time tag, the number of facts, no hidden text, reasons. Whether the answer is useful to a person from “for whom” is for you to decide: read only the heading and the first sentence and ask yourself whether that’s enough for a person to grasp the main thing.