Yes. Two research teams showed in 2024 that content can be engineered to appear more often in AI-generated answers. Optimizing for AI bots is the practice of shaping content so that large language models such as ChatGPT and Perplexity select it, cite it, and recommend it when they answer users. The interesting question is no longer whether it works. It is which techniques work, because they are not the ones SEO taught us.
What did the Harvard study find?
In their 2024 paper "Manipulating Large Language Models to Increase Product Visibility", Aounon Kumar and Himabindu Lakkaraju of Harvard University tested whether carefully crafted text could steer an LLM's product recommendations.
Their method: insert a Strategic Text Sequence (STS) into a product's information page. The STS is generated with Greedy Coordinate Gradient (GCG) optimization, an algorithm built for adversarial attacks and repurposed here to boost visibility.
The authors describe it plainly:
"We develop a framework to game an LLM's recommendations in favor of a target product by inserting a strategic text sequence (STS) into the product's information."
The experiment used fictitious coffee machines. The results:
- ColdBrew Master, a high-priced product the model rarely recommended, reached the top recommendation slot within 100 optimization iterations, even against affordability-focused queries.
- QuickBrew Express, already reasonably placed, improved further. STS helped both invisible and moderately visible products.
The researchers flag the consequences themselves:
"This ability to manipulate LLM-generated search responses provides vendors with a considerable competitive advantage and has the potential to disrupt fair market competition."
What did the GEO study find?
Where the Harvard team manipulated one product at a time, Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, and colleagues proposed a systematic framework in their 2024 paper "GEO: Generative Engine Optimization".
Their premise, in their own words: "Traditional SEO methods are not directly applicable to Generative Engines." Keyword matching alone does not cut it; subtler techniques are needed.
They tested concrete techniques across many queries and measured visibility inside generative answers:
- Keyword stuffing: little to no improvement.
- Statistics addition, embedding relevant, verifiable data: strong gains.
- Quotation addition, including credible quotes: strong gains.
- Cite sources, linking authoritative references: strong gains.
The best techniques produced visibility gains of up to 40%.
One finding matters especially for smaller sites. A fifth-ranked website using the cite-sources technique increased its AI visibility by 115%, while the top-ranked site lost 30%. The authors note that factors such as backlink building should not disadvantage small creators. Generative engines flatten part of the old hierarchy.
What do the two studies say together?
They answer two halves of the same question.
The Harvard work proves the mechanism: LLM outputs respond to targeted text changes. The GEO work proves the practice: ordinary editorial improvements, meaning evidence, quotes, and cited sources, shift visibility at scale and without adversarial tricks.
One is a warning about manipulation. The other is a playbook for optimization. Both confirm that AI answers are a surface you can compete on.
How do you apply this to your own content?
- Add verifiable statistics. Concrete, dated numbers with the source nearby make a passage easier for an engine to select.
- Quote credible voices. An attributed quotation gives the model a ready-made block to reuse.
- Cite and link your sources. This was the technique that helped lower-ranked sites the most in the GEO study.
- Skip keyword stuffing. The research measured almost no effect, and it makes the text worse for humans.
- Monitor the answers. Ask the engines your customers' questions and check over time whether you appear, and how.
It helps to know why these techniques work. Generative engines retrieve pages and compose answers from the most useful passages, a process we explain in our guides to how large language models work and retrieval-augmented generation (RAG).
Where are the limits?
STS-style manipulation is fragile ground. Adversarial strings can be detected and neutralized by model providers, and an advantage gained that way can vanish with the next update. The GEO techniques, by contrast, improve content for human readers as well. That is why they hold up over time.
Access is the other boundary. Engines cannot cite pages their crawlers never reach. A file like llms.txt gives them a map of your site, and we explain the format in our guide to llms.txt.
RankWit monitors how AI engines answer questions about your brand and shows which sources they cite.
FAQ
Can you really influence what ChatGPT recommends?
Yes. Kumar and Lakkaraju showed that a strategic text sequence added to a product page pushed a poorly placed product to the top recommendation. That was a controlled adversarial experiment, though. For a real brand, the durable path is the editorial one: verifiable evidence, attributed quotes, and cited sources.
Which techniques improve AI visibility the most?
In the GEO study, adding statistics, including quotations, and citing sources produced the largest gains, up to 40% more visibility in generative answers. These are changes that also make content better for human readers, which is why we recommend them as the default starting point for any AI visibility effort.
Does keyword stuffing work on AI engines?
No. Aggarwal and colleagues measured little to no improvement from keyword stuffing. Generative engines select passages that carry evidence and clarity, not passages that repeat a target phrase. Time spent stuffing keywords returns more when invested in data, quotations, and source links instead.
Do small websites stand a chance in AI answers?
Yes, and sometimes a better one than incumbents. In the GEO experiments, a fifth-ranked site that cited sources gained 115% visibility while the top-ranked site lost 30%. Generative engines weigh the quality of a specific passage more heavily than the historical authority of a domain.