Researchers tested multiple ways to optimize websites for AI search and discovered exactly what to do to increase visibility. They were able to increase the visibility of small, low-ranking websites by 115%, allowing them to outperform the large corporate sites that typically dominated the top of search results.
Researchers from Princeton University, Georgia Institute of Technology, Allen AI Institute, and Delhi Technological University observed that a generative engine optimization technique called GEO can generally improve visibility by up to 40%.
Nine optimization techniques were tested across multiple knowledge domains (law, history, science, etc.) to find out which ones work, which don't, and which approaches actually hurt rankings.
Of particular interest is that some techniques worked particularly well in specific knowledge areas, and three of them worked particularly well for all types of sites.
The researchers highlight GEO's ability to democratize the top search results, writing:
“This discovery highlights the potential of GEO as a tool to democratize the digital space.
Importantly, many of these low-ranking websites are created by small content creators and independent businesses that traditionally struggle to compete with the large companies that dominate search engine results. That means there are many things. ”
Tested with Perplexity.AI
Researchers tested the Perplexity.ai search engine and an AI search engine modeled after Bing Chat and found that the results were similar to those of the engine modeled after Bing Chat. .
In section 6 of their research paper, they state:
“Similar to our generation engine, Quotation Addition performed best on alignment word count, with a relative improvement of 22% compared to the baseline. Additionally, Cite Sources, Statistics Addition Methods that have performed well on generation engines such as , show high improvements of up to 9% and 37% on the two metrics.”
Tested with AI search modeled after Bing Chat
The researchers tested the method on a generative search engine modeled after the Bing Chat workflow, and also on Perplexity.AI, an AI search engine.
Those people write:
“We describe a generation engine that includes several backend generation models and a search engine for source searches.
The generation engine (GE) takes a user query qu as input and returns a natural language response r. Here, PU represents personalized user information such as preferences and history.
The generation engine consists of two important components.
a.) Set of generative models G = {G1, G2…Gn}. Each serves a specific purpose, such as query reformulation or summarization.
b.) Given a query q, a search engine SE returns a set of sources S = {s1, s2…sm}.
Here we introduce a typical workflow. At the time of writing, this workflow looks very similar to BingChat's design. This workflow breaks down the input query into a set of simple queries that are easy for search engines to consume. ”
Related: Bing explains SEO for AI search
Search query used for testing
Researchers created a benchmark containing 10,000 search queries from nine different sources, across multiple knowledge domains and varying levels of complexity. For example, some queries required inference to solve the answer.
The research paper explains:
“…we have handpicked GEO-BENCH, a benchmark consisting of 10,000 queries from multiple sources reused for the generation engine, along with synthetically generated queries. The benchmark includes queries from nine different sources, each further categorized based on target domain, difficulty, query intent, and other aspects.”
Below is a list of nine search query sources.
1.MS macro,
2. Orcas-1
3. Natural questions
4. AllSouls: This dataset contains essay questions from 'Oxford University All Souls College'.
5. LIMA: Contains difficult questions that require the generation engine to not only aggregate information but also perform appropriate inferences to answer the question.
6. Da Vinci's Debt
7. Perplexity.ai Discover: These queries come from the Discover section of Perplexity.ai, an up-to-date list of trending queries.
8. ELI-5: This dataset contains questions from the ELI5 subreddit
9. GPT-4 Generated Queries: Generate queries based on query intent (e.g. navigation, transactions) and difficulty from different domains (e.g. science, history) to compensate for variation in query distribution. instructs GPT-4 to: Range of responses generated (e.g. open-ended, fact-based)
9 ranking strategies tested
Researchers tested nine different ways to optimize websites for different types of searches, including law and government, business, science, people and society, health, history, and other topics. We've tracked how different approaches work.
They found that each type of niche topic responds well to different optimization strategies.
The nine strategies tested were:
Authority: Change your writing style to make your authoritative claims more persuasive.
Keyword optimization: Add keywords from search queries
Add statistics: Modify existing content to include statistics instead of interpreted information.
- Cite sources (cite reliable sources)
- Add citations: Add citations and citations from high-quality sources.
- Ease of understanding: Make content easier to understand
- Fluency optimization is about making content clearer
- Unique words: Add rare, unique words that aren't widely used without changing the meaning of your content.
- Technical terminology: With this strategy, add both your own and technical terms as needed without changing the meaning of the content.
- cite source
- Add a quote
- Add statistics
Which method was most effective?
The top three optimization strategies are:
- cite source
- Add a quote
- Add statistics
These three strategies achieved relative improvements of 30-40% compared to baseline.
The researchers write about the success of these strategies:
“These methods include adding statistics related to the website content (Add Statistics), including authoritative citations (Add Citations), and including citations from authoritative sources (Citation Sources) However, changes to the actual content itself are minimal.
Still, it significantly increases the visibility of your website in the generation engine's response, increasing both the authenticity and richness of your content. ”
Fluency Optimization and Easy-to-Understand techniques also helped improve visibility by 15-30%.
These results were interpreted by the researchers as an indication of how the AI search engine evaluates both the content and the presentation of the content.
Related: AI for SEO: Can we work faster and smarter?
What optimization strategies didn't work?
Researchers were surprised to find that using a persuasive and authoritative tone in your content generally did not improve rankings on AI search engines, as did other approaches.
Similarly, adding keywords to content from a search query didn't work either. In fact, keyword optimization performed 10% worse than the baseline.
Optimization works differently in different knowledge areas
An interesting finding of this report is that which type of optimization is most effective depends on the knowledge domain (law, government, science, history, etc.).
They found that content related to the historical domain ranks better when an “authority” optimization is applied, where more persuasive language is used.
Citation optimization, which improves content with citations from trusted sources, worked very well for fact-based search queries.
Adding statistics worked well for legal and government-related questions. Statistics also worked well for “opinion” questions, where the searcher asks her AI for its opinion on something.
The researchers observed that:
“This suggests that incorporating data-driven evidence can increase the visibility of websites in certain contexts, especially these situations.”
Adding quotations worked well in the knowledge areas of People and Society, Description, and History. The researchers interpreted these results to mean that the AI search engine probably prefers “credibility” and “depth” for this type of question.
The researchers concluded that performing domain-specific optimization was the best approach.
read: Impact of AI search on SEO
Low-ranking websites improve their rankings with GEO
The good news from this research is that even websites that typically rank poorly can benefit from these strategies for optimizing for AI search engines.
They concluded:
“Interestingly, websites that rank lower in the SERPs, which typically struggle to gain visibility, benefit significantly more from GEO than websites that rank higher.
For example, using the Cite Sources method, the visibility of the website ranked #5 in the SERP increased significantly by 115.1%, while the visibility of the top-ranked website decreased by 30.3% on average. did.
…The application of GEO techniques provides an opportunity for these small content creators to significantly improve their visibility in the response of the generation engine.
Powering your content with GEO allows you to reach a wider audience, leveling the playing field and allowing you to compete more effectively with larger companies in the digital space. ”
read: List of 16 GPT SEO AI tools currently available
SEO game changer
This research study shows a new path for SEO regarding AI-based search engines. Those who said AI search would trump SEO were premature. This study seems to indicate that SEO will eventually evolve into GEO to compete with the next generation of AI search engines.
Read the research study here:
GEO: Generation engine optimization
Featured image by Shutterstock/ProStockStudio
