The GEO paper everyone quotes, read for CPG product pages: what it tested and what it didn't

If you have sat through a pitch for "GEO" this year, you have probably heard that the right content changes can raise your visibility in AI answers by up to 40%. The number comes from a real paper, GEO: Generative Engine Optimization, by researchers at Princeton and IIT Delhi, published at the KDD conference in 2024. Search Engine Land cited it again on November 6 in a piece about big brands losing ground in AI search.

The paper is worth reading, and fewer people quote it accurately than quote it. This guide covers what it actually tested, the results that matter for a packaged goods product page, and the limits that should stop anyone from promising you 40%.

What the researchers did

The authors built a simplified answer engine of the kind that was common in 2023: take a question, fetch the top five Google results, and have a language model (gpt-3.5-turbo) write an answer that cites those sources. They then picked one of the five sources at random, rewrote it using one of nine methods, and measured how much more or less of the answer drew on that source.

They measured visibility two ways. The main one, position-adjusted word count, counts how many words of the answer cite a source, giving more weight to sentences near the top. The second, subjective impression, asks a model to rate things like how relevant and influential a citation is.

The questions came from a benchmark they built, GEO-bench: 10,000 queries drawn from nine sources, including real search queries, Reddit's Explain Like I'm Five, debate questions and queries from Perplexity's discover page. The mix was 80% informational, 10% transactional and 10% navigational. They also ran a smaller check on Perplexity itself.

The nine methods, ranked

Show the data

Method	Visibility score
Quotation addition	27.2
Statistics addition	25.2
Fluency optimization	24.7
Cite sources	24.6
Technical terms	22.7
Easy to understand	22
Authoritative tone	21.3
Unique words	20.5
No optimization (baseline)	19.3
Keyword stuffing	17.7

Source: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024, Table 1. Overall position-adjusted word count. Baseline is the unmodified source.

The three methods that did best all add substance rather than style. Quotation addition adds relevant quotes from credible sources. Statistics addition replaces qualitative statements with numbers. Cite sources adds references to credible sources. The paper reports relative improvements of 30% to 40% on the main measure for these, which is where "up to 40%" comes from.

Two results get less attention than they deserve. Keyword stuffing, the oldest trick in search, scored 17.7 against a baseline of 19.3: worse than leaving the page alone. And making the text sound more authoritative or persuasive did little on the main measure, which the authors read as engines being "somewhat robust" to tone.

The finding that matters most: who gains

The result we would put first for a challenger brand is in the paper's Table 2. The researchers broke the gains down by where the rewritten source had ranked in the original Google results.

Rank 1 sourceRank 5 source

Show the data

Method	Rank 1 source	Rank 5 source
Cite sources	−30.3%	+115.1%
Quotation addition	−22.9%	+99.7%
Statistics addition	−20.6%	+97.9%

Source: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024, Table 2. Relative improvement in visibility for the rewritten source, by its rank among the five sources fetched.

When the fifth-ranked source added citations, its visibility rose 115.1%. When the top-ranked source did the same, its visibility fell 30.3%. The pattern held for quotations and statistics. The authors conclude that these methods help lower-ranked sites far more than top-ranked ones, and they frame it as a way for smaller creators to compete with large companies.

For a small or mid-sized CPG brand, that is the encouraging part. It suggests the answer engine's choice of what to quote is driven by what a page contains, not only by where it ranked. It fits Ahrefs' brand data, which we covered in today's data post: authority gets you considered, but it does not guarantee you are the source the answer leans on.

What happened on a real engine

To check the results outside their own setup, the authors ran a smaller test on Perplexity, using 200 queries and supplying the sources as uploaded files.

Method	Main measure (position-adjusted word count)	Second measure (subjective impression)
No optimization	24.1	24.7
Keyword stuffing	21.9	28.1
Quotation addition	29.1	32.1
Statistics addition	26.2	33.9

Source: Aggarwal et al., Table 5. The paper describes quotation addition as a 22% improvement on the main measure and statistics addition as up to 37% on the second, with keyword stuffing 10% worse than baseline on the main measure.

The direction matched the main study. Substance helped; stuffing did not.

The limits, and why they matter for products

Before applying any of this to a product page, read the fine print the pitches leave out.

It was mostly not shopping. Eighty percent of the benchmark was informational queries such as "what does caffeine do to your muscles" or "why does my cat kick its toys." Only 10% were transactional. Product carousels, merchant feeds, prices and reviews were not part of the test at all.

The engine is old. The test engine fetched five sources and used gpt-3.5-turbo. Today's AI Overviews, AI Mode and ChatGPT shopping answers work differently. Google says its AI features may use "query fan-out," issuing many related searches across subtopics, according to its documentation. Search Engine Land notes the experiments ran on the researchers' own setup and Perplexity rather than Google, so the numbers are directional.

The rewrites were made by a model. Each method was applied by prompting a language model to edit the page. The paper's examples show additions like a survey figure in a sentence about chocolate consumption. The paper does not test whether those additions were true. On a product page, an invented statistic is a legal problem before it is a visibility tactic.

Visibility is not recommendation. The measure counts how much of an answer draws on a source. It does not measure whether a product was recommended, or whether the answer described it correctly.

How we would apply it to a CPG product page

Read with those limits, the paper still points somewhere useful: toward specific, verifiable facts, stated plainly, with their source. That is also the direction the November 5 brand-fidelity study points, where gaps in published facts were filled with guesses; see today's think piece.

Replace adjectives with numbers you can stand behind. "Packed with protein" becomes "12 g of protein per 45 g bar." "Long-lasting" becomes "one 64 oz bottle washes up to 64 medium loads," if that is what your testing shows. Every number should already be on the label or in your substantiation file. Tip 6 covers what kind of content gets cited.

Cite the outside source for claims that rely on one. If a certification, a standard or a test backs a claim, name it and link to the certifier's page. "Certified by" plus a link is the product-page version of the paper's cite-sources method. Tip 21 covers how to show certifications.

Quote credible sources, carefully. The paper's best method was adding quotations. On a product page, the credible quote is usually a third party's: a review publication's verdict, a certifier's description of what the mark means. Only quote what you have permission to use, and never invent a testimonial.

Put facts in tables, not images. A nutrition panel in a photo is invisible to most systems. A plain HTML table of the same numbers is not, as tip 22 explains.

Do not stuff keywords. The paper's most consistent result is that repeating query terms does not help and can hurt. Write the shopper's question once, answer it properly, and move on.

Answer the follow-up questions. The paper's examples are all single questions. Real shoppers ask a second and third. Our playbook on answering follow-up questions covers what to add.

Paper's method	Generic version	Product page version
Statistics addition	"Our soap is very gentle"	"pH 5.5; free of the 26 EU-listed fragrance allergens" (only if tested)
Cite sources	"Certified organic"	"USDA Organic, certified by [named certifier], with a link"
Quotation addition	"Customers love it"	A short, attributed verdict from a review publication you are allowed to quote
Keyword stuffing (avoid)	"best natural soap natural soap for sensitive skin natural"	One plain sentence: what it is, who it is for

How to tell whether your edits worked

The paper measured visibility with its own engine. You cannot do that, but you can run a simple before-and-after. Write down ten shopper questions your page should answer. Ask them in Google AI Mode, ChatGPT and Perplexity, and note whether the answer cites your page or repeats your facts. Make the edits to a few pages, leave a few similar pages alone, and ask the same questions again in three or four weeks. Some brands do this with a tracker such as SearchPanel, from XENA Intelligence; the important part is that the questions and the dates stay fixed.

The paper did not prove that any product page will get 40% more visible. It did show, under controlled conditions, that answer engines lean on pages with specific, sourced facts, that keyword tricks backfire, and that smaller sources gain the most from doing the work. For a brand that is not the category leader, that last result is the reason to start.