Rewriting product copy barely moved AI rankings in a new audit. Being a known brand did

For two years the pitch for generative engine optimization has gone roughly like this: rewrite your pages the way the models like, and the models will like you back. A paper posted on November 5, 2026 by three researchers at Politecnico di Milano is the cleanest test of that pitch we have read, and it lands hard on one side. On identical product pages, the brand name printed at the top moved the AI ranking. The rewrite did not.

We think every challenger brand should read it, and we think most of them will draw the wrong lesson. The wrong lesson is "only famous brands win, so why bother." The right one is narrower and more useful: stop spending money on making copy sound different, and spend it on being named in more places and saying more specific things.

What the researchers did

The paper, The Popular Get More Popular, by Giorgio Monaco, Jacopo Lenti and Francesco Pierri, has two parts.

The first is an audit. They took commercial keywords from four sectors in the Italian market (sport, fashion and beauty, insurance and travel), turned each into a recommendation question that names no brand, and asked it of three engines with and without web search: OpenAI's gpt-5.5, Google's gemini-3.6-flash and Perplexity's sonar. Each question was asked plain and with three buyer personas, ten times over. That came to 19,200 answers, collected on October 14, 2026, with the user location set to Rome. They then counted which brands were mentioned and which brands' own websites were cited, and compared that with how often each brand is searched on Google.

The second part is the experiment, and it is the part that matters. They took 24 of the questions, pulled five real product pages for each, and asked four models to rank them. One page, the target, was then changed in two ways. Its signature (the brand name, product name, model codes and web address) was swapped among the real brand, a better-known real brand, a less-known real brand and an invented brand with no registered domain. And its text was rewritten three ways: restructured with the key facts first and headings and lists, rewritten for fluency, and rewritten in an authoritative tone, the last two following the methods in the original GEO paper. Every rewrite had to keep all the facts and numbers and stay within 15% of the original length, and a separate model checked that nothing was added or removed. In total: 38,400 rankings, collected between October 28 and November 3.

The result in one chart

Show the data

Change made to the page	Positions gained
Real brand and product vs invented brand	+0.44
Less-known real brand vs invented brand	+0.19
Real product name vs invented one, same brand	+0.14
Fact-preserving GEO rewrite vs original text	+0.02

Source: Monaco, Lenti and Pierri, arXiv 2611.07614, Table 2. Means over 24 questions and four models (GPT, Gemini, Perplexity, DeepSeek), search off. The rewrite figure is the mean of three rewrites and was not significant.

With its real brand and product name, the target page averaged a rank of 3.21. Signed with an invented brand, word for word the same, it averaged 3.65. That gap, 0.44 positions, was positive for each of the four models and in each of the four sectors. Even a little-known real brand ranked the page 0.19 positions above an invented one.

The rewrites moved the page by about 0.02 positions. The authors report the 95% interval as running from slightly below zero to 0.049, so a gain of more than 0.05 positions is ruled out. The rewrites were written by gpt-5.5 and did no better on GPT than on the other models.

There is one more detail we keep coming back to. Swapping the target onto a more popular real brand did not help it compared with a less popular one. Among real brands that did not own the page, popularity made no measurable difference. What mattered was whether the name was real and whether it matched the page.

The audit says the same thing from the outside

The first study looked at the world as it is, not a lab setup, and it points the same way. Brands in the most popular fifth by search volume were mentioned and cited 1.4 to 2.8 times as often as the average brand of their engine and sector. A brand searched ten times more often on Google was mentioned in about 0.87 more answers out of 100 with search off and 0.89 more with search on. Turning search on, in other words, did not loosen the hold of the brands the model already knew.

The authors found one reason why. On GPT, which reports its searches in order, the most popular brands were often written into the very first search query, before any web result had come back. In travel, 45% of the mentioned brands in the most popular fifth were already in that first query, against 7% in the least popular fifth.

They also checked whether the brands an engine mentioned actually appeared in the sources it returned. Mostly they did. But the more interesting number is where.

Show the data

Engine	Share of mentions
Gemini	42.9%
Perplexity	42.6%
GPT	29.1%
All three engines	36.2%

Source: Monaco, Lenti and Pierri, arXiv 2611.07614, Appendix C.2. 30,006 mentions in calls with observed search queries. Pages were read after the calls; a lexical match.

More than a third of the time, the brand reached the answer through somebody else's page. That matches what Babel42 found in US data last week, where only 15% of AI brand recommendations cited the brand's own site; we covered it in our post on own-site citations.

Why this does not contradict the GEO paper

On November 9 we read the original GEO paper for product pages and found that its best methods helped lower-ranked sources a lot. That looks like the opposite of this result. It is not.

The methods that worked in the original paper added substance: quotations, statistics, citations to outside sources. The Milan team deliberately excluded anything that added facts. Their rewrites could reorder, restyle and reword, but not add. Read together, the two papers say something consistent: making the same facts sound better does almost nothing, while adding real, specific facts can help. The first is what a lot of GEO retainers sell. The second is what product and regulatory teams already own.

Our view: stop paying for tone

We think this paper should end one line item in most challenger brands' plans: paying anyone to rewrite existing product copy so it sounds more authoritative, more fluent or more AI friendly. The best evidence available says the effect is close to zero. If a page is hard to read, fix it for people. Do not expect the model to notice.

What should replace it? Two things, and neither is fast.

Be named where the models read. If more than a third of mentions travel through third-party pages, the job is to be on those pages, correctly. That means complete retailer listings, reviews that describe what the product is for, coverage in the trade and consumer press, and category roundups that include you. Tip 11 covers how to get into the sources AI answers cite, and tip 18 covers making sure the model sees your brand as one consistent entity. None of this is new advice. The paper just makes it the main advice.

Make the product facts more specific, not more polished. The experiment held facts constant on purpose. In real life you can change them: add the numbers, the certifications, the who-it-is-for and the who-it-is-not-for. The persona results in the same paper hint at why that matters. When a buyer persona with a budget or a life stage was placed before the question, the popularity slope fell from 0.98 to 0.80 for the budget persona, and the engines' searches added price words far more often. A constrained question gives the model a reason to look past the default names. A product page that answers constrained questions is the one it can find. Tip 20 covers writing for specific needs.

What it means if you are the big brand

If you run a category leader, the paper is good news with a catch. The model's familiarity with your name is an asset you did not have to optimize for. The catch is in the experiment: when the product name did not match the brand, the page lost ground. An invented product name under the real brand cost 0.14 positions. For a house of brands with dozens of sub-lines, renames and regional names, inconsistency between what the model remembers and what your pages say is a real cost. Keep product names stable and identical across your site, retailers and press.

It is also worth noting how thin the lead can be in packaged goods. In EMARKETER's food and beverage index for the third quarter, the top 10 brands overall had mention rates of just 3% to 5% across 16 categories, according to its September 8 summary. Familiarity helps, but no one owns the shelf.

The limits

This is one paper, posted as a preprint, and it has clear boundaries. The prompts were in Italian, run from Rome, in four sectors, only two of which (sport and fashion and beauty) are near packaged goods. The ranking experiment ran with web search off, so it measures what a model does with pages it has been handed, not what a live shopping answer does after fetching results. The models were called through APIs, which may behave differently from the consumer apps. And the authors measured mentions and citations, not sales.

None of that makes the central result weak. It was positive for every model and every sector, and it lines up with an earlier audit by Will Jack and colleagues, which found that category leaders kept about 80% of their recommendations across buyer personas while mid-market brands swapped up to 75% of the set, per its abstract. The middle of the market is where the movement is.

What we would do this month

If we were running AI visibility for a challenger CPG brand, we would do three things before touching any copy for style.

First, list the 20 third-party pages that come up most when you ask assistants about your category, and check whether you are on each one and described correctly. Second, open your top five product pages and count the specific, checkable facts on each: sizes, counts, certifications, ingredients, who it is for. If the count is low, add facts, not adjectives. Third, test with personas, not bare prompts, so you see the questions where you can win. Teams that track this across assistants sometimes use a tool like SearchPanel from XENA Intelligence; a shared sheet and a fixed list of questions is a fine start.

The paper's title is a fair summary of how assistants behave today. It is not a sentence on challengers. It just tells you that the way in is through being known and being specific, not through sounding right.