What is generative engine optimization (GEO), and when is it worth doing?

Generative engine optimization (GEO) is the work of making your pages and your business more likely to be used, quoted and cited when AI assistants such as ChatGPT, Perplexity, Claude and Google's AI Overviews write an answer. The term comes from a 2023 research paper by Aggarwal and colleagues (Princeton University, IIT Delhi and others), which found that adding quotations, statistics and source citations to a page made it more visible inside AI-generated answers. GEO is worth doing when your buyers research a considered purchase and ask an AI assistant for a shortlist, for example B2B software, professional services and higher-priced products. For local, quick-decision businesses and brand-new sites, standard SEO covers most of it for now. Google itself says that optimizing for its AI features is still SEO.

GEO in one paragraph

Someone asks ChatGPT which bookkeeping software suits a 10-person design agency. The assistant searches, reads a handful of pages and writes one answer. It may cite three or four sources and name a few brands. GEO is what you do so that your pages are among the ones it reads, your passages are the ones it quotes, and your business is described correctly when it comes up. Most of that work overlaps with SEO: pages have to be crawlable, indexed and useful. What changes is the target. In SEO you aim for a position on a results page. In GEO you aim to be used inside an answer, and that answer may never send you a click.

GEO, AEO, LLMO, AI SEO: several names, mostly one practice

As of early 2026 there was no agreed academic definition, and these labels were used interchangeably, as Wikipedia's entry on generative engine optimization notes. AEO (answer engine optimization) is often used more broadly, covering featured snippets and voice answers as well as AI answers. GEO points at generative assistants specifically. The label matters less than the question behind it: which answers do you want your business to show up in?

Where the term comes from

The 2023 paper

The name comes from “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande. The first version went up on arXiv on 16 November 2023 and lists Princeton University, Georgia Tech, the Allen Institute for AI and IIT Delhi. The paper was accepted at KDD 2024, the ACM SIGKDD conference held in Barcelona in August 2024.

What it actually tested

The researchers built a benchmark called GEO-bench: 10,000 queries drawn from 9 sources. About 80% were informational, 10% transactional and 10% navigational. For each query they took the top 5 Google results and had gpt-3.5-turbo write an answer that cited them. Then they rewrote one of those five sources using one of nine methods (a language model did the rewriting) and measured how much more of the answer drew on that source. Each result was averaged over 5 runs.

They measured visibility two ways. The main metric, Position-Adjusted Word Count, counts how many words in the answer are attributed to a source, weighted toward the top of the answer. The second, Subjective Impression, is a model-graded score for how prominent and relevant the source seems to a reader. They also ran a smaller check on Perplexity.

This was a research benchmark built on a 2023 model, not a test of live ChatGPT results. Keep that in mind whenever you see the paper quoted.

What raised visibility

Source: Table 1 of version 3 of the paper (28 June 2024).

The paper reports that quotation addition, statistics addition and cite sources gave a 30 to 40% relative gain on the main metric and 15 to 30% on Subjective Impression, with the best methods reaching 41% and 28%. Rewrites for fluency and easier reading gained 15 to 30%. On Perplexity, adding quotations raised the main score by 22%.

Results varied by topic. Statistics worked best for law and government questions and for opinion questions. Quotations worked best for people and society, explanation and history questions. The best pair of methods (fluency plus statistics) beat the best single method by more than 5.5%, measured on a 200-example subset.

What didn't help

Keyword stuffing, a staple of old-school SEO, gave little to no improvement and scored below the baseline. On Perplexity it came in 10% below baseline. Rewriting a page in a more persuasive, authoritative tone gave no significant gain either. The assistant cared about verifiable substance, not confident wording.

What “up to 40%” means

You will see “GEO boosts visibility by up to 40%” on a lot of pages. That figure is a relative gain in a word-count score inside generated answers. It is not 40% more traffic, leads or sales, and it says nothing about rankings.

It is also partly zero-sum. With the cite sources method, the source ranked 5th gained 115.1% in visibility on average, while the top-ranked source lost 30.3%. An answer has limited room, so when one source gets more space, others get less. If you already rank first, competitors doing GEO work can take space from you.

What the paper can't tell you

• It used GPT-3.5 in 2023. Nobody in that paper tested today's ChatGPT, Claude or Google's AI Overviews.
• Most queries were informational. Only 10% were transactional, which are closer to “which product should I buy” questions.
• It measured a visibility score, not clicks or revenue.
• Every rewritten page was already in Google's top 5. The methods made good pages more quotable. They didn't get unknown pages into the answer.
• The authors say they didn't test effects on search rankings, and that the methods may need to change as generative engines change.

GEO vs SEO: what overlaps and what's different

What Google says: for Google Search, it's still SEO

Google's guide to optimizing for generative AI search (last updated 10 July 2026) says that optimizing for generative AI search “is optimizing for the search experience, and thus still SEO.” Google's AI features run on its core ranking systems, using grounding (retrieval-augmented generation) and query fan-out, where one question is split into several searches.

The same guide is blunt about what you can skip for Google:

• llms.txt and similar files “will neither harm nor help” in Google Search.
• You don't need to chunk content into small blocks or rewrite pages for AI.
• No special schema is needed for AI features.
• Fake or inauthentic mentions don't help.

To be eligible at all, a page must be indexed and eligible for a snippet, and the site must be included in generative AI features in Search Console. Google also points to Merchant Center feeds for products and Google Business Profile for local businesses. It warns against tools that claim access to its internal metrics. More on the Google side in our guide on how to show up in Google AI Overviews.

Where the work really differs

Outside Google, the rules are less clear. ChatGPT, Perplexity and Claude don't publish which index they search or how they pick sources, and we haven't verified it either. What you can reason from is this: an assistant can only repeat what it finds. If the only page describing your business is your own, you give it little to work with. Reviews, listings, comparison articles and press coverage on other sites all add to what it can find.

There are also no positions to track. You test a set of prompts and count how often you appear, knowing the answer may change on the next run. And the unit is smaller: a single clear, sourced paragraph can get quoted even if the rest of the page is ordinary.

Which businesses benefit most

We found no reliable data on how many buyers ask AI assistants for shortlists, so this section is an argument, not a statistic. The logic: GEO pays off where people research before they buy and where an assistant's answer shapes who they contact.

Considered purchases and B2B

If your buyer spends days or weeks deciding, compares several options and has to justify the choice to someone else, they are likely to ask an assistant to summarize the field. Think accounting software, an IT managed service provider, a manufacturing supplier, a law firm for a specific type of case, or a $2,000 e-bike. Being named in that summary puts you on the list before a buyer ever visits a website.

Categories where buyers ask AI for a shortlist

Look at how people phrase questions in your category. If they type “best X for Y”, “alternatives to X” or “X vs Y”, they want a shortlist, and assistants are built to give one. Those are the categories where comparison pages, honest pros and cons, and clear facts about who you suit tend to get used.

A 15-minute test: ask the assistants yourself

1. Write down 5 to 10 questions a real buyer would ask, in their words. Include at least one “best X for Y” and one comparison.
2. Run each one in ChatGPT, Perplexity, Claude and Google (AI Overviews or AI Mode). Use a private window so your history doesn't color the answer.
3. Log the results in a sheet: date, assistant, prompt, whether you were named, which competitors were named, which pages were cited, and anything said about you that is wrong.
4. Repeat each prompt at least twice on different days. Answers vary between runs, so a single result tells you little.

If you are named often and correctly, GEO is mostly maintenance. If competitors are named and you aren't, and the cited pages are comparison articles or directories, you know where to start. Our step-by-step on checking whether ChatGPT mentions your brand goes into more detail.

Where GEO falls short

You can't control what the model says

You can make good material easy to find and quote. You can't decide what the assistant writes. The paper's own setup averaged 5 runs per result because answers vary. Anyone who promises a specific AI mention is promising something they don't control.

Measurement is noisy

For Google you get the Generative AI performance report in Search Console. For other assistants you have prompt samples, and samples drift: a different wording, location or day can change the answer. Treat the numbers as a direction over months, not a score you check daily.

New brands with no third-party mentions

The paper's methods only worked on pages already in Google's top 5. A new business with a new site, few reviews and no coverage elsewhere has little for an assistant to find. GEO formatting won't fix that. Getting indexed, ranking for some searches and earning genuine listings and reviews comes first.

Regulated industries with strict claim rules

The methods that tested best were adding statistics, quotations and citations. In health, finance, legal and similar fields, those are exactly the kinds of claims that rules and compliance teams restrict. Every number and testimonial may need sign-off, and some may not be allowed at all. GEO can still help through clear, accurate explanations, but expect slower work and fewer of the tactics that scored highest.

What GEO work looks like in practice

1. Check eligibility in Search Console. Pages indexed, nothing blocking snippets, the site included in generative AI features.
2. Take a baseline with the prompt test above, so you can compare later.
3. Write answer-first pages. Put the direct answer in the first paragraph, back claims with sourced numbers, and quote your own people where they know something first-hand. This is the honest version of what the paper tested.
4. Publish comparison and alternatives pages for your category, including who you are not a good fit for. Shortlist questions pull from pages like these.
5. Get real listings and reviews. Google Business Profile, Merchant Center if you sell products, and the directories that matter in your industry.
6. Track a fixed set of prompts every week or month, in the same assistants.
7. Fix what assistants get wrong. If they quote an old price or a feature you dropped, update your own pages and listings so the correct fact is easy to find.

More on the assistant side in how to get your business recommended by ChatGPT.

This is what FoundWell runs every day inside your WordPress, Umbraco, Shopify or Wix site: articles, fixes, indexing checks and AI answer tracking (monthly on Starter, weekly on Growth and Complete), with pages written for gaps in AI answers on Complete. A person on our team reviews every article before it goes live, and in approval mode nothing is published until you approve it. Plans start at $250 a month with no contract. See our generative engine optimization services for what each plan includes.

Two things we're upfront about. Every plan includes schema and an llms.txt file. Schema helps pages qualify for rich results in Google; it isn't a special key to AI features. llms.txt is there for systems other than Google, and we haven't verified which assistants read it. The 10 directories are real business listings, not a trick to get mentioned. We don't promise rankings or AI mentions. The only guarantee is on Complete, and it is about Google clicks measured in Search Console.

When standard SEO is enough for now

Put GEO on hold, or treat it as a side effect of good SEO, if any of these apply:

• Your pages aren't reliably indexed yet, or Search Console shows crawl errors you haven't fixed.
• You run a local, quick-decision service such as a plumber, hair salon or takeout restaurant, where Google Business Profile and reviews do most of the work.
• Your business is new and nobody else has written about you yet.
• You have no time or budget to test prompts regularly, so you couldn't tell whether GEO work changed anything.
• The prompt test shows assistants already describe you correctly and name you alongside your competitors.

In those cases, the SEO basics are also what Google uses for its AI features, so the effort isn't wasted when you do move on to GEO.

If your buyers do ask AI for a shortlist and you'd rather not run the prompts, pages and fixes yourself, connect your site to FoundWell. It takes about ten minutes.

Frequently asked questions

Is GEO replacing SEO?

No. For Google, its own guide (updated 10 July 2026) says optimizing for generative AI search is still SEO, because its AI features run on the same core ranking systems. GEO adds work on top: tracking what assistants say, getting mentioned on other sites and writing passages that are easy to quote.

Does llms.txt help with GEO?

Not for Google. Google says llms.txt and similar files will neither harm nor help in Google Search. Whether ChatGPT, Perplexity or Claude read llms.txt hasn't been verified publicly, so treat it as a low-cost extra, not a ranking factor.

Can anyone guarantee that ChatGPT will mention my business?

No. Nobody controls what a language model writes, and the same prompt can give different answers on different runs. FoundWell doesn't promise AI mentions or rankings either. Be wary of any provider that does.

How long does GEO take to show results?

There is no reliable published timeline. Results depend on whether your pages are already indexed and ranking, how much is written about you elsewhere, and how often assistants refresh their sources. Track a fixed set of prompts over several months rather than expecting a set date.

What is the difference between GEO and SEO?

SEO aims to rank whole pages on search results and earn clicks, measured by rankings and clicks in Search Console. GEO aims to get your passages cited and your business named inside AI answers, measured by sampling prompts and, for Google, the Generative AI performance report. The foundations, indexing and useful content, are shared.

About the author

Peter Sinke, Founder

Peter Sinke builds websites at Datapad in the Netherlands and made FoundWell from the daily AI SEO routine he ran for Datapad's own site.