Generative Engine Optimization (GEO): Explanation & Algorithm Breakdown

Generative Engine Optimization, or GEO, is the practice of optimizing a company’s online presence to cause its products or services to be recommended by generative AI chatbots. As a new marketing channel, GEO has not been the subject of enough research to develop a shared understanding of its best practices among marketing professionals. Our team expects the literature about GEO to grow the way its closest antecedent, search engine optimization (SEO), did when it was first studied.

To that end, our research team conducted a study of the recommendation algorithm used by the 4 most popular generative AI chatbots in the U.S. The study took place from March to June 2024, and sought to identify and weigh the factors used by generative AI chatbots to make product and service recommendations. In total, we produced 11,128 commercial queries between the 4 chatbots, asking each for buying recommendations across a range of service and product categories. (The list of industries in which we conducted queries is listed in Appendix A below.)

In the table below, we break down the data we gathered from the study, listing the factors that influenced each chatbot’s recommendations in weight order. Afterwards, we define each factor and detail how each chatbot utilizes them to make recommendations

Generative Engine Ranking Factors

Below, we break down our research on the factors that influence commercial recommendations on each of the engines. While all generative AI engines share this group of factors, the weight each engine places on each factor varies quite a bit, as detailed above.

NOTE: The most advanced version of all 4 of the top generative AI chatbots personalize their recommendations as you tell them more about yourself, which may alter the weight of the below factors.

Authoritative List Mentions

Generative AI Engines, are, by definition, predictors. When generating content, their job is to predict the words, sentences, and paragraphs most likely to come next in a way that imitates the language of experts. They make their predictions by “studying” what multiple authoritative sources have to say on the subject, then blending the knowledge from those sources into a natural, human-like communication. In the same way, generative AI engines’ product and service recommendations come from their study of multiple authoritative sources, such as highly-ranked compendiums of the top products, services, and companies in a particular industry. Google has already invested enormously in vetting the authority of websites and ranking them, and gen AI engines use the top-ranked Google search listings to inform their output.?

One exception to the chatbots’ reliance on Google-vetted lists is Claude, which searches the open Internet sparingly, and mainly relies on traditional compendiums such as encyclopedias and directories.

Awards, accreditations, and affiliations

If an award, accreditation, or affiliation given to a company or product is mentioned on a web page, and the web page is deemed trustworthy as per the LLM’s training sources, it signals that company or product’s authority, making it more likely to be recommended.?

Online Reviews

ChatGPT, Gemini, and Perplexity place a substantial minority of weight on online reviews from trustworthy platforms such as Amazon, Better Business Bureau, Glassdoor, TrustPilot, Capterra, and CNet.

Social Sentiment

Social sentiment is a measure of how positively or negatively a company is talked about in news articles, public social media accounts, and discussion forums. While it is currently a relatively minor factor, used only by ChatGPT, we expect its weight to increase in the future due to its importance in real-world recommendations.

Customer Examples & Usage Data

When recognized brands publicly associate with products or companies, as in an endorsement, announced partnership, or case study, AI chatbots can make inferences about the credibility of the product or company. Similarly, third-party data about product usage or customer base size is an indicator of authority. Currently, two AI chatbots – ChatGPT and Claude – use this factor to inform their recommendations.

Google Website Authority

Google assigns an authority score to domains and website pages, originally known as PageRank. It is primarily based on consistent publication of helpful content and backlinks from other domains. Gemini places substantial weight on this factor.

Local Business Reviews

Gemini and Perplexity use online reviews from popular platforms such as Google Business Profiles (GBP), Yelp, TripAdvisor, and Angie’s List to make recommendations for local queries.

Traditional Databases & Directories

All generative AI chatbots train their LLMs using a base of widely-trusted texts such as Wikipedia and Encyclopedia Britannica; the New York Times and the Wall Street Journal; and the literary canon. They also use business databases and directories such as Hoovers, Bloomberg, nd IBISWorld. Claude draws from these sources directly for the purpose of making business recommendations.?

ChatGPT’s Recommendation Algorithm

ChatGPT’s recommendation algorithm mainly relies on searching Bing (whose ranking algorithm is largely based on Google’s) for lists, reviews, and directories that rank highly. It then provides its own amalgamated recommendation based on those sources.?

Sometimes, it relies heavily on the #1 ranked Bing search result. For example, in one query we asked ChatGPT “Who are the top demand generation agencies?” and it returned a verbatim copy of a list published on First Page Sage’s website, which currently ranks #1 on Bing for “top demand generation agencies 2024” – the exact keyword into which ChatGPT translated our query.

ChatGPT scans the top 5 to 10 search results, verifies their authority, then looks for common items that rank highly on the lists, concluding that the best items are the ones mentioned most frequently. For searches where highly-ranked lists conflict about the top items, ChatGPT moves to seeking out awards, accreditations, and affiliations; online reviews; and to a smaller degree, customer examples, usage data, and social sentiment.?

For example, when we asked it “What are the best lawnmowers under $1,000?” it returned 3 models that it identified based primarily on reviews from the New York Times and Consumer Reports.??

The top Bing search results for this query vary widely and include a number of affiliate-influenced lists, and thus weren’t used to generate recommendations. Secondly, there aren’t any awards given out for lawnmowers, so that factor was also bypassed. Thus, the algorithm moved to the next-weightiest factor, trusted reviews. Once a set of 5 lawnmowers was built from the aforementioned trusted reviews in the New York Times and Consumer Reports, the set was ordered and delivered as a recommendation. We believe the order was influenced by the number of times each of the 5 lawnmowers in the set was recommended within major news sites over the last 2-3 years, as the order was closely correlated with the number of mentions reported in our news monitoring tools. This final factor is an example of Social Sentiment influencing the chatbot’s recommendations.

Google Gemini’s Recommendation Algorithm

Google Gemini’s recommendation algorithm is similar to ChatGPT’s but relies more on Google systems and products such as Google website authority, Google Business Profile authority, and Google local reviews. Its first action is to search the first page of Google and return an amalgam of recommendations from those results, citing each website next to the answer.??

For example, we asked it to tell us the top custom software development firms and it replied as follows

Unlike ChatGPT, Gemini does not rely as heavily on the #1 result from its search engine, instead looking for companies that are common to several top-ranked lists or directories. It places higher weight on companies that have been cited as “award-winning,” recommending them even if they don’t appear on multiple lists or directories. Conversely, it does not recommend companies with low online reviews (<3.5 stars) even if that company appears on several top-ranked lists or directories and is cited as award-winning.?

In product searches, Gemini acts similarly. For example, when we asked it what the best facial moisturizers were for dry skin, nearly all of its recommendations were sourced from the #2 search result, a People.com product review article. Notably, it re-interpreted our search for the “best” cream (by which we meant “most effective”) as “most popular.”

In keeping with that reinterpretation, all 3 of the recommended moisturizers were top sellers by volume according to industry data; and were well reviewed (4+ stars).

Gemini uses a different recommendation algorithm for local commercial queries such as “Can you recommend a plumber near Rockville, MD?” While authoritative list mentions still factor into its recommendations, having a high star rating on Google Business Profile correlates most strongly with receiving a recommendation. Gemini also factors in non-Google reviews, such as from Yelp, TripAdvisor, and Angie’s List.

Perplexity’s Recommendation Algorithm

In many ways, Perplexity had the simplest recommendation algorithm of the 4 we studied. Nearly all commercial queries returned recommendations from lists that ranked in the top 5 search results on Google for the equivalent query. Perplexity picks from 2-3 lists, ordering its recommendations based on online reviews and, to a lesser extent, companies that are cited as award-winning, accredited, or affiliated with an authoritative brand (e.g. Harvard or Apple).

Like Gemini, it has a separate algorithm for recommending local businesses, again relying heavily on high-ranking lists but putting substantial weight on reviews from Google, Yelp, TripAdvisor, Time Out, Eater, and other authoritative review sites.?

Claude AI’s Recommendation Algorithm

Claude stands alone amongst the 4 major generative AI chatbots in that it has limited Internet access. Most of its recommendations are sourced from traditional business databases such as Bloomberg and Hoovers.

It orders its recommendations based on information about the awards, accreditations, affiliations, customer lists, usage data, and popularity of the companies it sees as relevant to the query. Because it draws from databases that aren’t as dynamic as those on the open Internet, its recommendations are more likely to be larger, more established companies. For example, we asked Claude who the top travel agents in the US were, and it favored businesses that have been around 50+ years over younger but higher-rated travel agencies.

Unlike ChatGPT, Gemini, and Perplexity, Claude doesn’t even attempt to recommend local businesses.?

Sources

• Research Study, January 2024 — June 2024, First Page Sage Offices, New York, NY
• Gemini: A Family of Highly Capable Multimodal Models — December 2023, Gemini Team, Google
• GPT-4 Technical Report — March 2023, OpenAI
• How Do You Use Personal Data in Model Training? — May 2024, ClaudeAI
• How Does Perplexity Work? — June 2024, Perplexity
• Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking — February 2024, Nikhil Sharma, Q. Vera Liao, Ziang Xiao

Appendix A: Industries Queried for Generative Engine Optimization Study