Generative engine optimization (GEO): what the research found and what to do

What is generative engine optimization?

A search engine returns a list of links and lets you choose. A generative engine writes the answer itself: it searches, reads several pages, and composes a reply in its own words, usually with a few sources linked. ChatGPT with search, Perplexity, Google AI Overviews and Claude with web search all work this way.

Generative engine optimization, or GEO, is the work of making your content one of the sources such an answer is built from, and making sure it is represented correctly when it is. The goal is different from classic search engine optimization. In a list of ten links, position is everything. Inside a written answer there is no page two: either the assistant used your page and named you, or it did not.

For a business, GEO comes down to three questions about its website. Can the AI bots get in and read the text? Do they understand who the business is? And do they find answers clear enough to quote? The rest of this guide takes them in turn, after a look at where the term comes from and what the evidence behind it says.

Where the term comes from

The name comes from a research paper, “GEO: Generative Engine Optimization”, by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, from the Indian Institute of Technology Delhi, Princeton University and independent researchers. It was posted on arXiv on November 16, 2023 and accepted at KDD 2024, a major data mining conference.

The authors built a benchmark they called GEO-bench: about 10,000 search queries from nine existing datasets, across many subjects. For each query they took the web pages a search engine returned, rewrote one of them using a given method, and measured how much more or less that page was used in the answer a generative engine wrote. The engine in the main experiments was built on GPT-3.5. They then repeated part of the test on Perplexity, a public AI search engine.

They measured visibility two ways. The first, Position-Adjusted Word Count, counts the words of the answer that come from a source, giving more weight to words near the top. The second, Subjective Impression, asks a language model to rate how relevant, influential and prominent a source looks in the answer.

What the research found

The paper tested nine ways of rewriting a page. The table shows the ones that helped most and the classic SEO tactic that did not, with the change against the unchanged page, computed from the averages in the paper’s Table 1 on the test set.

Three findings stand out for a business site.

Evidence beats repetition. The methods that worked added something a reader can check: a number, a quotation, a source. Repeating keywords, the oldest trick in search, made the page slightly less visible on the main measure.

Smaller sites gained the most. The authors split results by where a page ranked in the original search results. Pages ranked fifth gained the most: Cite Sources raised their visibility by 115.1%, while the same method lowered the visibility of first-ranked pages by 30.3% on average. The paper’s own reading is that GEO can help smaller content creators compete.

Results vary by subject. The paper reports that what works differs between domains, and it argues for domain-specific methods. On Perplexity, the gains were smaller but went the same way, up to 37% on one measure.

The paper’s headline, “up to 40%”, is a maximum in a controlled test, not a typical result, and the paper does not claim more than that.

What the findings mean for a business website

A research benchmark is not your website, and an engine built on GPT-3.5 in 2023 is not ChatGPT today. The lessons that carry over are the general ones, and they match what the AI companies themselves say about readable, useful pages.

• Put the numbers in the text. Prices or price ranges, how long a job takes, how many clients you serve, the warranty in years, the area you cover in miles. An assistant answering “how much does a kitchen remodel cost in Denver” can quote a page that says it, and cannot quote one that says “contact us for a quote”.
• Quote people with names and roles. A licensed electrician explaining when a panel must be replaced, a dietitian on your team explaining a menu choice. Never invent a quotation or a customer: a fake testimonial is the fastest way to lose the trust you are trying to build.
• Say where facts come from. “The 2024 building code requires…”, “according to the manufacturer’s spec sheet…”. A claim with its source is easier to use than a claim without one.
• Write plainly. Short sentences, one idea each, the answer first and the detail after. Fluency helped in the study, and it helps the people reading the page too.
• Do not stuff keywords. Writing “best plumber in Austin” five times in a paragraph did nothing for visibility in the study and reads badly to buyers.

Note what these steps have in common: each makes a page more useful to a human reader as well. GEO done well does not need a second version of your site written for machines.

A worked example: one service page, before and after

Here is how those lessons change a common kind of page. The roofing business, the towns and every figure below are invented for illustration; the pattern is what matters.

Before. “Welcome to Summit Roofing, the trusted roofing experts! We offer top-quality roof replacement at competitive prices. Our experienced team is committed to excellence and customer satisfaction. Contact us today for a free quote!”

After. “Summit Roofing replaces asphalt shingle roofs in Denver, Aurora and Lakewood. A typical single-family roof of 2,000 square feet takes two days and costs between $14,000 and $19,000, depending on the pitch, the layers to remove and the shingle line. Every job carries a 10-year workmanship warranty on top of the manufacturer’s. The license number and insurance certificate are on the About page.”

Nothing in the second version is clever. It simply replaces every claim a reader cannot check with one they can.

• Numbers instead of adjectives. “Competitive prices” became a price range and what moves it; “experienced” became a warranty in years. This is the Statistics Addition finding, applied to a business.
• Places instead of a slogan. Three named towns answer “who replaces roofs near me” in a way “the trusted experts” never can.
• A source for the trust claim. The license and insurance are pointed to, not asserted. That is the spirit of Cite Sources.
• No keyword repetition. “Roof replacement” appears once, where it belongs.

The after version would also convert better with a human reader, which is the point: the buyer asking an assistant and the buyer reading the page want the same facts. What it should never do is invent a quotation, a customer or a statistic to look more quotable. A made-up figure that an assistant repeats is worse than no figure at all.

The three layers of GEO: access, identity and content

The paper studied one layer, the content of a page. On a real website, two layers sit underneath it, and if either is broken the best content in the world is never read.

Fix them in that order. Access first, because a site the bots cannot open gains nothing from better writing; how each assistant finds pages covers the bots one by one. Identity second, because an assistant that cannot tell who you are may attribute your facts to a competitor with a similar name, or leave you out to be safe. Content last, and continuously, because that is where the paper’s findings apply and where the work never quite ends. For the content layer in depth, see our guide to answer engine optimization.

How GEO relates to SEO

GEO does not replace search engine optimization; it is built on top of it. Google says plainly that there are no extra requirements to appear in AI Overviews or AI Mode: a page must be indexed and eligible to show in Search with a snippet, and no special schema or AI text file is needed. So everything that makes a page work in Google, from a sensible title to a page that loads fast, also counts for Google’s AI answers.

Where GEO adds work is in three places that classic SEO rarely touched.

• Other bots. ChatGPT, Claude and Perplexity send their own search bots, which can be blocked separately in robots.txt or by a firewall, and which do not run JavaScript.
• Identity across the web. An assistant composes an answer from many sources. If your name, address and services are written differently on your site, your profiles and directories, it may not join them into one business.
• Answers rather than pages. A page that ranks well for a keyword can still be useless to an assistant if the answer to the buyer’s question is buried under three paragraphs of introduction.

In practice the two rarely conflict. Clear, factual, well-structured pages help both.

What GEO cannot do

It is worth being as clear about the limits as about the methods, because the field is young and claims run ahead of evidence.

• Nobody controls what an assistant answers. The answer depends on the model, the wording of the question, the user’s location and history, and the moment. The same question can name different businesses an hour apart.
• The AI companies do not publish how they choose sources. They document their bots and how to block them. They do not document how a page is picked among those a bot can read.
• Recrawl schedules are unknown. OpenAI says a robots.txt change can take about 24 hours to reach its systems. How soon a rewritten page is read again is not published by any of the four big assistants.
• There is no full report of mentions. Google counts AI Overviews and AI Mode traffic inside the overall Web data of Search Console, without a separate AI view, and Bing reports citations in its own AI answers. ChatGPT, Claude and Perplexity offer site owners no equivalent.

So treat any promise of a guaranteed placement in AI answers with care. What GEO can do is remove the reasons an assistant cannot use your site, and give it material worth quoting. Whether and when it does is something to measure over time.

What the MWY checker does for GEO

The MWY checker scores your website on 24 items in the three layers of this guide: can AI read it, does AI know who you are, and can AI quote it. It reads the site the way the AI search bots do, as raw HTML with no JavaScript, and it obeys robots.txt. The rules are fixed, so the same site gets the same result every time, and each item says whether it is certain or only a signal.

The free check shows your score, the score of each layer and the five items that matter most on your site. Pro shows all 24 with the fix for your platform, snippets filled in with your details, and a plan ordered by impact and effort; it also asks ChatGPT, Gemini, Claude and Perplexity 20 buyer questions about your business every week and shows who gets named. See what Pro includes.

The checker does not rewrite your content, and it does not promise that an AI will name your business. It shows what stands between your site and the assistants, and how to remove it.