Back to Blog
  • AI SEO
  • AI Search Visibility
  • Citation-Ready Content

Generative Engine Optimization Strategies: What the Research Tested and What Holds Up

Hayden BondUpdated 11 min read
Generative Engine Optimization Strategies: What the Research Tested and What Holds Up
Many generative engine optimization strategies trace back to one 2023 lab study. It found that adding quotes, statistics and sources helped a page show up in its test AI's answers, and that keyword stuffing hurt. But it ran on GPT-3.5, and a newer test found those rewrites didn't move a page up the list of cited sources. If you're paying for GEO work, that's worth knowing.

What are generative engine optimization strategies?

Generative engine optimization (GEO) is the work of getting your business named or cited when an AI tool like ChatGPT writes an answer.
I covered what the acronyms mean in what is answer engine optimization, and how GEO compares with regular SEO in GEO vs SEO.
Search for GEO advice and you'll find the same three tips again and again: add statistics, add quotes, cite your sources. Much of it points back to one paper. I wanted to know what that paper really tested.

What the original GEO paper tested

In November 2023, Pranjal Aggarwal and researchers from Princeton and IIT Delhi posted "GEO: Generative Engine Optimization." It was later accepted at KDD 2024, a data science conference. It's the paper that gave GEO its name.
They built a test set of 10,000 questions. For each one, their setup pulled the top five Google results and had GPT-3.5 Turbo write an answer from them. Then they rewrote one of the five pages in nine different ways and measured how much of the answer came from it.
The keyword stuffing made things worse. The page got 8.3% less visibility than if it had been left alone. Adding quotes from credible sources did best, at 40.9% more. Adding statistics gave 30.6%, smoother writing 28%, and citing sources 27.5%. A more authoritative tone made no real difference, the authors wrote.
Two findings get less attention. When the team rewrote all five pages at once, the pages lower in Google's results gained the most. With sources cited, the page in fifth place gained 115.1%, while the top page lost 30.3% on average. And the best tactic depended on the subject. Statistics helped most on law and government questions. Quotes helped most on people and society, and on history.
I think the core of it holds up. A page that backs up what it says gives an AI tool something to use. That's the idea behind our citation-ready content work.

Why the lab results may not carry over to today's AI tools

I read these numbers as a lab result. Three things make me careful with them.
The model is old. GPT-3.5 Turbo is several generations behind. By August 2026, ChatGPT's default model for free users was GPT-5.6.
The setup was small. The test engine only ever read five pages, and all five came from Google's top five. Real AI tools run their own searches and read more than that. David Konitzny at Peec AI, a tracking tool, found ChatGPT pulling up about 24 sources per question after that update, up from about 12.
The Perplexity check wasn't live search. The paper says the tactics also worked on Perplexity, where quotes helped about 22%. Keyword stuffing did about 10% worse on one of the paper's two scores, though its other score went up. But Perplexity doesn't let you choose its sources, so the team uploaded the page text as files, on 200 questions. That tests whether Perplexity uses your page once it has it. It doesn't test whether Perplexity's search finds your page in the first place.
I wouldn't expect a 40% lift from adding a few quotes to your service page.

What a newer test on newer AI models found

In 2025, a team led by Haritz Puerto at Parameter Lab built a new test called C-SEO Bench. It was accepted at NeurIPS 2025, an AI research conference. They tried eight of the GEO paper's methods, plus two of their own, on four newer models from OpenAI and Anthropic, including GPT-4o mini and Claude 3.5 Haiku. Depending on the set, each question came with about five to ten pages, for both general questions and product recommendations.
They measured something different from the first paper. The GEO paper measured how much of the answer's wording came from your page. C-SEO Bench measured how early the AI tool cited your page. By that measure, none of the eight rewrites taken from the GEO paper gave a significant gain. Adding statistics moved a page down in 19 of the 24 settings they tested. What did help was what they called traditional SEO, which in their test meant getting a page placed higher in the list the model reads.
The authors say the results of the two papers "do not contradict each other." So I wouldn't call the first paper wrong. A page that fills more of the answer just isn't the same as a page that gets cited first.
They also tested what happens when more pages use the same rewrite. They ran it on their own two methods, the only ones that showed any significant gain, for shopping-type questions. The more pages used them, the smaller each page's gain got. The authors called it "congested and zero-sum."
That doesn't surprise me. I've watched the same thing happen in regular SEO for years.

Which generative engine optimization strategies have already stopped working?

In my experience, a tactic works until enough people use it that the platform has to filter it out. Then it can turn into a liability.
SEO has been here before. Exact-match domains, where the web address is the search phrase, used to give sites a boost. In September 2012, Google's Matt Cutts announced a small change that would "reduce low-quality 'exact-match' domains in search results."
I think it's moving faster now. Listicles, the "best X" roundups that are sometimes written by a company that ranks itself first, are a good example. Lily Ray, an SEO consultant, calls them one of the formats "most heavily spammed for GEO over the past year or two." Konitzny looked at shopping questions in the prompts Peec AI tracks. When GPT-5.6 became ChatGPT's default in August 2026, listicles dropped from 15.47% of the pages ChatGPT pulled up to 8.19%. Product pages took the lead at 16.39%. Konitzny describes it as a change in how the new model looks things up, and those are pages ChatGPT read, not pages it cited. Still, a format a lot of GEO advice leaned on lost about half its share with one model change.

15.47% to 8.19%

Listicles' share of the pages ChatGPT pulled up for shopping questions in Peec AI's tracked prompts, before and after GPT-5.6 became the default in August 2026

Source: David Konitzny, Peec AI, August 2026 ↗
Google moved too. In May 2026, it made clear that its spam policies cover its AI answers. Its definition of spam now includes "attempting to manipulate generative AI responses in Google Search." Its guide to AI features also says you can ignore "AEO/GEO hacks" like chunking content, llms.txt files and fake mentions, at least for Google Search. I go through that list in GEO vs SEO.
You can buy mentions. You can't really buy people's respect. I mean, you can try.

Whose pages AI tools cite: yours or someone else's

In September 2025, a team at the University of Toronto compared what ChatGPT, Perplexity, Gemini and Claude cited with what Google showed. Most of their questions were about consumer products, like "best laptops for students." They sorted each source three ways: the brand's own site, earned media (reviews, publishers, other people writing about you) and social.
The AI tools leaned hard on earned media. It was about 9 in 10 citations for ChatGPT and Claude, and about 2 in 3 for Perplexity and Gemini. Google's mix was more balanced. Social posts were close to zero for ChatGPT but about 1 in 8 for Perplexity and Gemini.
Two more findings matter if you're a smaller business. On buy-now questions, brand sites moved up, most of all for ChatGPT. And big brands got named more. Asked about popular sodas, ChatGPT named major brands 56.3% of the time and niche brands 12.3%. Perplexity's split was 67.9% and 5.8%.

Share of soda brand mentions in AI answers

ChatGPT

Major brands

56.3%

Niche brands

12.3%

Perplexity

Major brands

67.9%

Niche brands

5.8%
Major brands
Niche brands

Chen, Wang, Chen and Koudas, University of Toronto, arXiv 2509.08919, September 2025. Fifty unbranded soda questions. The rest of the mentions were other brands.

An April 2025 study by Andres Algaba and others points the same way. GPT-4o, writing reference lists for 10,000 science papers, favored papers that were already highly cited. I take that as a hint, not proof.
The same caveats apply to the Toronto study. It used the developer versions of mid-2025 models, not the apps people use, and mostly "top 10" style questions. So I read it as a direction. What other people write about you counts for a lot, which is where our entity SEO work starts. In that study, your own site counted for more when someone was ready to buy. And on some AI tools, social search matters more than you'd guess.

How to tell if any of it is working

Picture someone asking ChatGPT where to eat tonight. It names a restaurant. They show up for dinner. Google Analytics sees none of this. No click. No session.
And the AI tool can get the details wrong. In March 2025, Klaudia Jaźwińska and Aisvarya Chandrasekar at Columbia's Tow Center gave eight AI search tools quotes from news articles and asked for the headline, publisher, date and link. Together, the tools gave wrong answers to more than 60% of the questions. That was news, not businesses. But you might be getting named without knowing it. Or described wrong. Or both.
So I'd check what AI tools say about you on a schedule, and not trust any one answer. Our measurement page covers what we track and how sure we can be, and our AI visibility tools directory compares the tools that do the checking.

The generative engine optimization strategies I'd still use

This is my read of all of the above. I haven't seen a study that tests the whole list.
  • Back up what you say. Name the source for every number, and quote people who know the subject. The GEO paper tested this, and it's good writing anyway.
  • Get written about by others. Trade press, reviews and local news count for a lot in the Toronto study.
  • Make your own pages plain and complete. Prices, specs, sizes, hours and service areas. On buy-now questions, the Toronto study found AI tools leaned more on a brand's own site.
  • Keep working on regular search. In C-SEO Bench, a higher spot beat a clever rewrite.
  • Skip keyword stuffing, listicles that rank yourself first, and fake mentions.
None of it is a trick. Tricks stop working once everyone uses them. These are the things I'd still be glad I did if the next model changes again. It's the idea behind our generative engine optimization work.
If you'd rather we take a look, the AI Search Game Plan shows where you stand in Google and in AI answers, who comes up instead, and what to do first. Or start on your own: pick one claim on your homepage this week and add where it comes from.

Common questions

What generative engine optimization strategies does the research support?

The 2023 GEO paper found that adding quotes, statistics and cited sources raised a page's share of AI answers in its lab setup. A 2025 test on newer models, C-SEO Bench, found none of those rewrites reliably moved a page up the list of cited sources, while a higher spot in the list of pages the model reads did help. The two studies measured different things. I'd back up your claims, get written about by others, and keep working on regular search.

Does keyword stuffing hurt AI visibility?

In the GEO paper's lab test, it did. Pages stuffed with keywords got 8.3% less visibility than pages left alone, and about 10% less on one of two scores in the Perplexity check, though the other score went up. That test ran on GPT-3.5 Turbo, so I'd treat the exact number loosely. I've never seen a reason to stuff keywords for AI tools.

Do quotes and statistics help you show up in AI answers?

They did in the GEO paper, where quotes raised a page's visibility by 40.9% and statistics by 30.6%, compared with the unchanged page, on GPT-3.5 Turbo reading five Google results. C-SEO Bench later found none of these rewrites reliably moved a page up the list of cited sources on newer models, and adding statistics often moved it down. I'd still add them, because sourced claims help the person reading too.

Do GEO tactics work the same on ChatGPT and Perplexity?

Not exactly. In the University of Toronto's 2025 study, about 9 in 10 of ChatGPT's citations went to earned media like reviews and publishers, against about 2 in 3 for Perplexity. Perplexity cited social posts about 1 in 8 times, and ChatGPT almost never did. When asked for the most popular sodas, both named big brands far more often than niche ones.

What does Google say about GEO?

Google's guide to its AI features says terms like AEO and GEO are common online, but many of the "hacks" sold under them don't work for Google Search. It says you can ignore chunking content and llms.txt files there. In May 2026, Google also made clear its spam policies cover attempts to manipulate its AI answers.

Sources

Share this article
Add as a preferred source on Google

Topics

Curious how your business looks on Google and in AI answers?

Plate Lunch Collective is an AI SEO agency that gets brands named when buyers ask, in search, on social, and in AI. If you'd like to talk it through, we're easy to reach.

Work with us

Hayden Bond

Hayden Bond, in SEO since 2004, founded Plate Lunch Collective, an AI SEO agency that gets brands named when buyers ask, in search, on social, and in AI.

Stay current on AI search

New research and analysis, direct to your inbox.

Related articles

Explore our AI SEO services

Previous

Instagram Rewards Original Content Now. Does AI Search?

Next

Farm Marketing When Customers Ask AI Where to Go