What GEO Is and Why It Matters

GEO (generative engine optimization) is the work of improving how often, and how accurately, your brand shows up in answers written by AI tools such as ChatGPT, Gemini, Perplexity and Google's AI Overviews and AI Mode. Lesson 1 of GEO 101 covers what the term means and where it came from. It also covers what the original research did and didn't show, and how GEO lines up against the SEO work you already do. You'll also see real numbers from MentionX scans on how often AI answers name a brand at all.

In this lesson

• Why this matters: the shortlist you never saw
• GEO in one sentence
• Where the term comes from
• The three outcomes: mentioned, recommended, cited
• How GEO relates to SEO
• From MentionX data: how often answers name a brand
• Worked example: explaining GEO at Ledgerline
• What GEO is not
• Common mistakes in week one
• Exercise: the 60-second explanation
• Quiz
• Sources

Anything marked "Our view" is our opinion from running AI visibility scans, not documented engine behaviour. Ledgerline, the company in the worked example, is fictional.

Why this matters: the shortlist you never saw

A finance director who wants to automate supplier invoices used to search Google, open six tabs and skim vendor pages. More and more, that first step looks different. She types "what's the best accounts payable automation for a 300-person company on NetSuite?" into ChatGPT or Gemini. A few seconds later she has a written answer with four or five named vendors, a sentence on each, and sometimes a handful of source links.

That answer is a shortlist. If your product isn't on it, you didn't lose a ranking. You lost the chance to be considered at all, and nothing in your analytics will tell you. No impression was logged and there was no click to miss. A blue-link search leaves a trace in Search Console. A conversation with an assistant mostly doesn't.

That's the core reason GEO exists as a discipline. It's not because SEO stopped working; we'll come back to how much of it still applies. It's because a growing share of buyer research now ends in a generated answer, and you need a way to see, measure and influence what those answers say about you.

GEO in one sentence

Here's the definition we use across this course:

GEO is the practice of measuring and improving how AI answer engines mention, recommend and cite your brand for the questions your buyers ask.

Every word in that sentence is doing work:

• Measuring comes first. You can't improve what you haven't checked, and AI answers vary from run to run (Lesson 5 shows how much). So GEO starts with a repeatable set of questions, not a one-off screenshot.
• AI answer engines means any tool that writes an answer rather than returning a list of links: ChatGPT with search, Gemini, Perplexity, Microsoft Copilot, and Google's AI Overviews and AI Mode.
• Mention, recommend, cite are three separate outcomes, covered below and in depth in Lesson 3.
• The questions your buyers ask keeps the work honest. Showing up for questions no buyer asks is a vanity metric.

Where the term comes from

"Generative engine optimization" was formalised in a research paper by Pranjal Aggarwal, Vishvak Murahari and colleagues from Princeton University and IIT Delhi. It was first posted in November 2023 and published at the ACM KDD conference in 2024 (GEO: Generative Engine Optimization). It's worth knowing what the paper did, because it gets quoted loosely.

What they built. The authors described a "generative engine" as a system that takes a user query, rewrites it into simpler search queries, retrieves sources, and has a language model write an answer with inline citations. To test it, they built a benchmark called GEO-bench with 10,000 queries drawn from nine sources, including real Bing and Google query datasets and Perplexity's trending questions.

What they tested. They took one source among the top five search results for each query and rewrote it in nine different ways. Examples included making it more authoritative in tone, adding statistics, adding quotations, citing sources, simplifying the language, and stuffing in keywords from the query. Then they measured how much of the generated answer drew on that source.

What they found. Three changes helped most: adding quotations from relevant sources, adding statistics, and citing sources. These lifted the source's visibility by roughly 30 to 40% on the paper's position-adjusted word-count metric. Keyword stuffing, the classic old-school SEO move, did slightly worse than doing nothing. The gains were largest for sources that ranked lower in the original search results. A source ranked fifth gained 115% from citing sources, while the top-ranked source lost visibility on average. They saw similar patterns when testing on Perplexity, with gains of up to 37% on one metric.

What it didn't show. The engine in the main experiments was GPT-3.5 answering from the top five Google results, not ChatGPT or Google's AI Overviews as they work today. The metric measures how much of an answer is attributed to a source, not whether a buyer chose your product. And the content edits were made by a language model, then tested immediately, with no crawling or indexing delay. So treat the paper as strong evidence for a direction (evidence-rich, well-sourced content gets used more), not as a promise that adding three statistics will lift your share of AI answers by 40%.

The three outcomes: mentioned, recommended, cited

Ranking has one dimension: you're position 3 or position 9. AI answers have at least three, and they don't move together.

A brand can be mentioned without being recommended ("Some teams use Ledgerline, but it lacks a mobile app"). It can be recommended without being cited, because the model may name you while linking a review site. And your blog can be cited for a definition while a competitor gets the recommendation. Lesson 3 gives you scoring rules for each, including the awkward edge cases.

Our view: recommendation is the outcome closest to revenue, and the one most teams undercount. A mention in a list of ten is not the same as "for a NetSuite shop your size, start with Ledgerline."

How GEO relates to SEO

There's a lot of noise claiming GEO replaces SEO. Google's own documentation says close to the opposite. Its guide to AI features states that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." It also says a page must be indexed and eligible to show with a snippet to appear as a supporting link (Google Search Central). The same page says you don't need new machine-readable files, "AI text files" or special schema to appear.

So for Google's AI features, SEO is the entry ticket. Other assistants work in a similar way when they search the web: they retrieve pages through a search index, then write from what they retrieved. Lesson 2 explains that pipeline.

Where GEO differs is in the goal and the measurement:

That last row matters most. A ranking is mainly about one URL. An AI answer blends many sources: review sites, Reddit threads, comparison articles, Wikipedia, YouTube. Your own site is one voice among them. A lot of practical GEO is about what other sites say about you, which is closer to PR and product marketing than to on-page SEO.

From MentionX data: how often answers name a brand

We wanted to give you a realistic picture of a "normal" AI visibility number before you run your own check. So we ran read-only aggregate queries over MentionX's own scan records.

Method. We counted every completed engine run with an extracted answer, from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 tracked projects. Each project tracks one target brand against named competitors. We counted an answer as a "mention" when our extraction found the target brand named anywhere in it. Prompts are tagged by intent when they're created. Engines were called through their APIs with web search or grounding switched on (OpenAI with web search, Gemini with Google Search grounding, Google AI Mode, Perplexity Sonar, Copilot with Bing grounding, and DeepSeek). Brands are anonymised, and no single customer's results are shown.

Results by prompt intent:

How to read this. The top two rows are close to 100% by construction: if the prompt names you, the answer will too. The rows that matter are the unbranded ones. When a buyer asks "what's the best X," the tracked brand appeared in roughly one answer in four. When the buyer only described their problem, it appeared in about one in seventeen.

Caveats. These are 18 projects, mostly B2B software, so they're not a sample of the whole market. The projects chose their own prompts. API results can differ from what a logged-in user sees in a consumer app. Treat the numbers as a reference point, not a benchmark you must hit.

Our view: the problem-solution row is the most useful number on this page. Most brand marketing assumes the buyer already knows the category. AI assistants are where people go before they know the category, and that's where most brands are close to invisible.

Worked example: explaining GEO at Ledgerline

Illustrative example. Ledgerline is a fictional company and the answers below are invented to show the method.

Ledgerline sells accounts payable automation to mid-market finance teams. Its marketing lead, Priya, has to get budget for a GEO pilot from a VP who thinks it's a buzzword. Here's how she builds the case in one afternoon, using only what this lesson covers.

Step 1: pick five buyer questions. She takes them from recent sales calls, not keyword tools:

1. "Best accounts payable automation for a company on NetSuite"
2. "How do I stop paying duplicate supplier invoices?"
3. "Alternatives to Bill.com for a 400-person company"
4. "Is Ledgerline good for multi-entity accounting?"
5. "Ledgerline vs Tipalti"

Step 2: run each in two assistants, logged out. That's ten answers. She notes, for each, whether Ledgerline is mentioned, recommended or cited.

Step 3: tally. Her invented results:

Step 4: write the three sentences. "When buyers ask AI assistants for AP automation without naming us, we showed up in 2 of 6 answers and were recommended once. For the 'duplicate invoices' problem, which is exactly what we solve, we weren't named at all. I'd like four weeks to build a proper prompt set, run a baseline, and fix the two or three biggest gaps."

Notice what she didn't say. There's nothing about "the future of search" and no 40% promise. Ten answers isn't a statistically solid sample, and she doesn't pretend it is; it's a reason to measure properly. That proper measurement is what Lessons 4 to 7 build.

What GEO is not

• Not a trick to fool models. The GEO paper found persuasive or authoritative wording didn't reliably help, and keyword stuffing hurt. Engines are built to ignore that kind of manipulation. Expect them to get better at it.
• Not a replacement for SEO. For Google's AI features, indexing and snippet eligibility are prerequisites.
• Not only about your website. Third-party pages (reviews, forums, comparisons, documentation) shape answers as much as your own pages.
• Not a one-time audit. Answers change between runs and over time, so you need repeated measurement.
• Not guaranteed. Google's documentation says plainly that indexing and serving aren't guaranteed even when every requirement is met. No one can promise placement in an AI answer.

Common mistakes in week one

1. Testing in your own logged-in account. Memory and past chats can colour answers. Lesson 4 sets up clean sessions.
2. Asking only branded questions. You'll see near-100% mention rates and conclude you're fine.
3. Taking one screenshot as the truth. Answers vary between runs; a single answer is an anecdote.
4. Counting every mention as a win. "Ledgerline is often criticised for slow support" is a mention. Read the sentence.
5. Starting with tactics. Rewriting pages before you have a baseline means you'll never know what worked.

Exercise: the 60-second explanation

Write three sentences you could say to your manager:

1. What GEO is, in your own words, using "mentioned, recommended, cited".
2. One real buyer question in your market where you suspect you're missing from AI answers.
3. What you'd measure first.

Then test it: run that buyer question once in ChatGPT and once in Gemini, logged out. Does the answer support your suspicion? Keep the note. You'll turn it into a full prompt set in Lesson 5.

Quiz

1. Which three outcomes does GEO aim for? Mentioned, recommended and cited.

2. In the original GEO paper, which change performed worst? Keyword stuffing, which did slightly worse than leaving the content unchanged.

3. According to Google, what extra technical requirements must a page meet to appear in AI Overviews? None beyond normal Search eligibility: the page must be indexed and eligible to show with a snippet.

4. Why should branded and unbranded prompts be scored separately? Branded prompts name you, so answers almost always mention you. Mixing them in inflates your mention rate and hides the real gap.

5. True or false: a brand mention is always good news. False. The context may be negative or may describe you inaccurately.

Next: Lesson 2: How AI engines build answers.

Sources

• Aggarwal, P., Murahari, V., et al. (2024). GEO: Generative Engine Optimization. KDD '24. arXiv:2311.09735
• Google Search Central, AI features and your website
• MentionX aggregate scan data, 22 August to 29 October 2026 (UTC), 2,154 engine runs, 18 projects. Anonymised and read-only. Method described above.