What Is GEO (Generative Engine Optimization)? A Guide to Getting Your Brand Featured in AI Search

Since 2012 I have optimized websites for Google rankings, and for most of that time the unit of success was simple: position and click. Then AI Overviews, AI Mode, ChatGPT and Gemini started answering questions directly. Consequently, clients now ask me a new question: how do we get named inside those answers? The discipline behind that goal is generative engine optimization, or GEO. In this guide I explain what it is, what actually works according to public research, and how a small business can start within thirty days.

What is generative engine optimization (GEO)?

Generative engine optimization is the practice of making your brand appear as a cited source or a named recommendation inside answers produced by AI search experiences such as Google AI Overviews, AI Mode, ChatGPT, Gemini and Perplexity. It builds on classic SEO and adds a citability layer: passage level answers, verifiable statistics, named sources and consistent brand signals.

The term comes from a research paper, not from an agency brochure. In 2023 researchers from Princeton, Georgia Tech, IIT Delhi and the Allen Institute coined it and tested which content changes raise visibility in generated answers. I will return to their findings later, because they separate real tactics from wishful thinking.

My honest position after fourteen years in the field: GEO is not magic that replaces SEO. Instead, most of what we did for Google rankings still applies. What changed is the measurement unit, citation and brand mention instead of click, and the content format, which now rewards clear answers at the paragraph level.

What is the difference between generative engine optimization and SEO?

Classic SEO aims to rank a page and win a click; generative engine optimization aims to get a passage cited and a brand named inside a generated answer. Both rely on the same crawl, index and quality systems. However, the unit of work, the success metric and the measurement tools differ. The table below shows the practical contrast I use in client proposals.

In short, the left column is the foundation and the right column is the extension. Therefore, if your site has crawl errors, thin pages or no authority, GEO tactics will not rescue it. That is why my SEO consulting work always starts with the basics before we talk about AI citations.

How do generative engines build an answer: what are query fan-out and RAG?

Most AI search products use a pattern called retrieval augmented generation, or RAG. First, the system retrieves relevant documents from an index. Then a language model writes an answer grounded in those documents and attaches links. Consequently, your page must be retrievable before it can ever be cited.

Google adds a second technique. According to its AI features documentation, AI Overviews and AI Mode use query fan-out: they split one question into subtopics and run several related searches at the same time. As a result, the answer can show a broader and more diverse set of links than a single classic results page.

This detail matters for your content plan. For example, a question like "best CRM for a small agency" may fan out into pricing, integrations, support and migration. Therefore, a page that covers one subtopic clearly and cites sources can enter the answer even if it never ranked first for the head term. In practice, I now write pages around subtopic clusters rather than a single keyword.

Do AI search results really reduce clicks?

Yes, for informational queries the loss is large and well documented. Ahrefs analyzed 300,000 keywords from Search Console data, half with AI Overviews and half without. In its updated study using December 2025 data, the presence of an AI Overview correlated with a 58 percent lower click through rate for the page in position one. Position two lost 50.8 percent and position three lost 46.4 percent. The first version of that study, from April 2025, had found a 34.5 percent drop.

Pew Research Center reached a similar conclusion from user behavior. In March 2025 it tracked 68,879 Google searches from 900 US adults. When an AI summary appeared, users clicked a traditional result 8 percent of the time, compared with 15 percent when no summary appeared. Clicks on links inside the summary made up only 1 percent of all visits.

However, the picture is not uniformly bleak. Semrush studied more than 10 million keywords and found that the share of queries triggering an AI Overview rose from 6.49 percent in January 2025 to 24.61 percent in July 2025, then fell to 15.69 percent in November 2025. Meanwhile, within those queries the informational share dropped from 91.3 to 57.1 percent. The commercial share climbed from 8.15 to 18.57 percent, and the transactional share rose from 1.98 to 13.94 percent. In other words, AI answers are moving toward the queries that make money, so being the cited source there is worth real revenue.

Who should care about generative engine optimization first?

Not every business needs to move at the same speed. Based on the data above and what I see in client accounts, I rank urgency like this:

1. Publishers and informational sites, because they already lose the most clicks and need citations to stay visible at all.
2. Service businesses with comparison heavy buyers, such as agencies, software vendors, clinics and consultants, because commercial queries increasingly trigger AI answers.
3. E-commerce brands in categories where shoppers ask "which one should I buy", because AI Mode and ChatGPT now recommend products by name.
4. Local businesses last, since map results and reviews still dominate their surface, although AI answers do quote review themes.

For example, a boutique law firm in Istanbul that publishes explanatory guides sits in group two. On the other hand, a neighbourhood bakery sits in group four and should keep its energy on reviews and a clean website. Above all, be honest about which group you belong to before you spend money.

Does Google require special markup for AI Overviews and AI Mode?

No. Google states this plainly in its AI features documentation: there are no additional requirements or special optimizations to appear in AI Overviews and AI Mode, and no special schema.org markup or "AI text file" is needed. The AI features rely on the core Search ranking systems. In practice, this means files like llms.txt have no official standing with Google.

That said, Google did publish a separate guide for generative AI features. It asks for original expert or experience based perspective that goes beyond common knowledge, instead of commodity content that anyone could write. It also recommends good page experience, main content that is easy to distinguish from other page elements, and high quality images and video. Finally, it points to the generative AI performance report in Search Console for measurement.

So the answer is: no secret file, but a higher bar for substance. Therefore, my first step in every GEO project is a content inventory that marks each page as original, commodity or outdated.

How should you configure robots.txt to appear in ChatGPT search?

ChatGPT search shows your site only if its search crawler can read it. OpenAI explains in its bots documentation that it operates three separate crawlers with three separate purposes. OAI-SearchBot indexes sites for ChatGPT search and does not feed model training. GPTBot collects content for model training. ChatGPT-User handles requests that a user triggers inside a conversation.

The practical steps are short:

1. Open your robots.txt and check whether a blanket rule blocks all bots.
2. Allow OAI-SearchBot explicitly, because that is the crawler behind search visibility.
3. Decide separately about GPTBot, since it only affects training, not search.
4. Leave ChatGPT-User alone unless you have a specific reason to block user initiated fetches.
5. Wait about 24 hours, which is the time OpenAI says a robots.txt change takes to propagate.

A minimal example looks like this: "User-agent: OAI-SearchBot" followed by "Allow: /", then optionally "User-agent: GPTBot" followed by "Disallow: /". If you prefer not to edit the file by hand, my free robots.txt generator builds the rules for you. Also, remember that blocking GPTBot is a business decision about training, not an SEO decision.

Does blocking Google-Extended remove you from Gemini?

Google-Extended is a robots.txt token, not a crawler with its own user agent. According to Google's crawler documentation, it controls whether crawled content may be used for training and grounding the Gemini models in Gemini Apps and Vertex AI. Google adds a sentence I quote to every client: Google-Extended does not affect a site's inclusion in Google Search and is not used as a ranking signal.

So what happens if you block it? Your pages still appear in Search, AI Overviews and AI Mode, because those features run on the Search index. However, the Gemini app may lose the ability to ground its answers in your content. For a brand that wants to be recommended inside Gemini conversations, that is a real trade off.

The decision matrix below summarizes what I recommend to most clients. Naturally, a publisher with licensing concerns may choose differently.

How do you write citable content: what is passage level optimization?

Passage level optimization means writing each section so that a single paragraph can stand alone as a complete, accurate answer. Generative engines do not quote whole pages; instead, they lift a few sentences. Therefore, the paragraph directly under each question heading carries most of the weight.

The formula I use has four parts:

• A question heading that matches how people actually phrase the query.
• A direct answer of 40 to 60 words that starts with the definition or the method, not with a warm up sentence.
• One number or one named source that makes the answer verifiable.
• A short example, ideally from your own experience, that proves you have done the thing.

Readability matters here more than in classic SEO, because a model will pick the cleanest passage among several similar ones. For instance, I run every draft through my readability checker and rewrite any paragraph above 80 words. Additionally, I avoid pronouns like "it" or "this" at the start of the answer paragraph, since a lifted passage loses its context.

Why do statistics, sources and quotations improve GEO results?

Because the researchers who defined GEO measured it. The KDD 2024 paper by Aggarwal and colleagues, available on arXiv, tested nine content modifications on GEO-bench, a set of 10,000 queries. Adding statistics, citing sources and including quotations raised visibility in generative engine answers by up to 40 percent. Keyword stuffing, by contrast, did not help and in some settings hurt.

The logic is intuitive once you think like a model. A generated answer needs support, so passages that already contain a figure and a source are cheaper to use than vague claims. Similarly, a quotation from a named expert reads as evidence. Consequently, the writing habits of good journalism turn out to be the writing habits of good generative engine optimization.

In practice, I add three things to every important page: at least one dated statistic with a link, one attributed quotation, and one clearly labelled example calculation. Also, I remove the fluff those elements replace, because the total length rarely needs to grow.

Why has brand mention become as important as a backlink?

A language model learns which brands belong to which topics from text across the web, not only from links. When ten independent pages describe your company as "an Istanbul based Google Ads specialist", the model associates your name with that phrase. In addition, retrieval systems often surface brand names that co-occur with the query terms, even without a hyperlink.

That changes off page work. Previously, I asked partners for a link; now I ask for a consistent description. Specifically, the company name, the category, the location and one distinctive claim should read the same on your site, your directory listings, your social profiles and the articles that mention you. Inconsistent naming splits the signal.

This is also where visual and verbal identity meet search. A clear brand identity gives writers and models a stable set of words to repeat. For example, if your tagline changes every quarter, every mention describes a slightly different company, and no single description reaches the threshold of confidence.

Which content formats get cited most in AI search?

From my own prompt testing and from patterns in the studies above, four formats appear again and again in generated answers. Naturally, results vary by topic, so treat this as a starting ranking rather than a law.

Notice the pattern: every strong format is structured, and every weakness is a lack of substance. Therefore, when I plan a content calendar for a client, I pair each format with a source of original input, such as their own customer data, their own case notes or their own tests. Otherwise, the page becomes the commodity content Google warns about.

How do you strengthen E-E-A-T signals for GEO?

E-E-A-T stands for experience, expertise, authoritativeness and trust, and generative engines lean on the same signals when choosing sources. The good news is that most of them are pages you can build this week. First, publish a real author page with a photo, a biography and a history that a reader can verify. My own about page lists the years, the certifications and the way I work.

Second, show proof of work. A references page with named clients and described outcomes tells both a human and a model that the experience is real. Third, connect the author to every article with a visible byline and, where possible, links to profiles on other platforms.

Fourth, write from experience in the text itself. Phrases like "in the accounts I manage" or "when I tested this in March" are not decoration; instead, they are exactly the experience signal Google's guide asks for. Finally, keep dates honest and update pages when facts change, because a stale statistic undermines trust faster than a missing one.

Is schema markup still necessary for generative engine optimization?

Necessary, no; useful, yes. Google is explicit that no special markup is required for AI features. However, structured data still earns rich results in classic listings, and it removes ambiguity about who wrote a page, when, and what entity it describes. That clarity helps any retrieval system match your page to a query.

My recommendation is a small, accurate set rather than a large speculative one:

• Organization or Person markup with the same name and description you use everywhere else.
• Article markup with author, date published and date modified.
• FAQ or HowTo markup only where the visible content genuinely matches it.
• Product markup with real prices and availability for e-commerce pages.

You do not need a developer for this. For example, my schema generator produces valid JSON-LD for these types in a minute. Then validate the output and, above all, never mark up content that a visitor cannot see on the page, since that violates Google's guidelines and helps no engine.

Why do speed, mobile fit and content separability matter for AI search?

Because retrieval happens on rendered pages, and a slow or cluttered page gives the engine less to work with. Google's generative AI guide recommends good page experience and a main content area that is easy to distinguish from navigation, ads and sidebars. In practice, a crawler that cannot tell your answer from your cookie banner will not cite your answer.

Three technical checks cover most of the risk:

• Load the main content in the initial HTML rather than injecting it with JavaScript after the fact.
• Keep one clear article container with headings in a logical order, without duplicated blocks.
• Serve the same content on mobile and desktop, since most of Google's indexing uses the mobile version.

Many older sites fail the first check without anyone noticing. Consequently, when a client's pages never appear in AI answers despite strong content, I look at the template before the text. A clean rebuild through my web design service often fixes visibility problems that no amount of rewriting could.

How do you measure generative engine optimization success?

Measure four things, in this order: citations, referral traffic, branded search and manual prompt tests. Start with Search Console. Google reports clicks from AI features inside the Web search type of the Performance report, and its generative AI performance report separates that traffic where available. Track it monthly, not daily, because the volume is noisy.

Next, open your analytics referral report and filter for chatgpt.com, gemini.google.com, perplexity.ai and copilot.microsoft.com. Those sessions are direct evidence of citation. In addition, watch your branded search volume in Search Console, since a brand that gets recommended in AI answers is usually typed into Google afterward.

For campaigns that you promote alongside GEO work, tag the links so they do not pollute the referral data. My UTM builder keeps the parameters consistent. Finally, run a manual visibility test: ten queries across three engines, ChatGPT, Gemini and AI Mode, recorded in a simple table every month. Note whether your brand appears, whether a link appears, and which competitor appears instead.

What can a small business do for GEO in thirty days?

A focused month is enough to build the foundation. Here is the weekly plan I give to small teams, and it assumes roughly one working day per week:

1. Week one, technical: fix robots.txt for OAI-SearchBot, decide on GPTBot and Google-Extended, check page speed on mobile, and add Organization and Article schema.
2. In week two, content inventory: list every page, mark it original, commodity or outdated, and rewrite the top ten pages with question headings and direct answer paragraphs.
3. During week three, authority: publish or refresh the about and references pages, add dated statistics with sources to the ten rewritten pages, and align brand descriptions across directories.
4. Finally, week four, measurement: set up the referral filter, record the first manual prompt test, and note baseline branded search volume.

Here is an example calculation, clearly labelled as illustrative. Suppose a page ranking first for an informational query earns 1,000 clicks a month. If an AI Overview appears and the Ahrefs average of a 58 percent drop applies, clicks fall to about 420. However, if the same page becomes a cited source and branded searches rise, part of that loss returns as direct and branded traffic. The numbers are not a guarantee; instead, they show why measuring citations matters as much as measuring clicks.

What are the most common GEO mistakes?

I see the same errors repeatedly, and most of them come from treating GEO as a trick rather than a discipline. Check your own site against this list:

• Blocking all bots with one rule, which removes you from ChatGPT search overnight.
• Uploading an llms.txt file and expecting Google to act on it, when Google says no such file is needed.
• Stuffing the phrase "generative engine optimization" into every heading, a tactic the Princeton study found ineffective.
• Publishing AI written summaries of other people's articles, which is exactly the commodity content Google's guide discourages.
• Writing answer paragraphs that begin with "In today's fast changing digital world" instead of the answer.
• Measuring daily and panicking, since AI feature traffic swings widely from week to week.
• Ignoring brand consistency, so that three different company descriptions compete against each other.

Above all, the biggest mistake is abandoning classic SEO. For instance, a site with duplicate titles and broken canonical tags will not become citable because someone added quotations. Fix the foundation, then add the citability layer on top.

Is it a problem to use AI when writing content for AI search?

Using AI as a tool is fine; using it to mass produce pages is not. Google's spam policies name scaled content abuse, meaning many pages generated primarily to manipulate rankings without adding value, regardless of whether a human or a machine produced them. Meanwhile, its generative AI guide asks for perspective that goes beyond common knowledge, which a model summarizing the web cannot supply on its own.

My working rule has three parts. First, the experience, the data and the opinion must come from a person who did the work. Second, AI may draft, restructure, translate or shorten, but a human checks every fact and every source. Third, the final page must contain something that did not exist on the internet before, such as your own numbers, your own tests or your own client patterns.

Ironically, that rule also produces the content generative engines prefer to cite. A model looking for a source wants the original, not the tenth paraphrase. Therefore, the safest path with Google's policies and the most effective path for citations turn out to be the same path.

When does professional support for generative engine optimization make sense?

Bring in a specialist when three conditions meet. Your category already shows AI answers for commercial queries, your site has dozens of pages that need passage level rewrites, and you need a measurable result within a quarter. Below that threshold, the thirty day plan above is genuinely something you can execute yourselves.

When I take on a GEO project, I do not sell it as a separate product. Instead, it lives inside a structured SEO engagement: technical audit, content inventory, passage rewrites, authority pages and a measurement dashboard that reports citations next to rankings. I work without intermediaries, so the person auditing your site is the person writing the plan and answering your messages.

If you want an honest opinion about whether your site is ready, send me the domain through the contact page. I will tell you where the gaps are, even if the answer is that you should fix the basics first. Generative engine optimization rewards the sites that already do the fundamentals well, and that is a good reason to start with them.