The one study everyone in GEO SEO quotes for its "up to 40%" headline also found that keyword stuffing pushed a page below the version nobody had touched.
That second finding gets left out of most sales decks, and it is the more useful of the two.
A note on the term first, because it means two different things. In this article "GEO SEO" means generative engine optimisation: getting a page retrieved, quoted and credited when ChatGPT, Perplexity, Copilot or Google's AI Overviews answer a question. Some glossaries, Arimetrics among them, use "GeoSEO" to mean geographic or local search. If you are after the second meaning, our local SEO page is the right place, and nothing below is about Google Maps.
What the 2023 GEO paper tested
The paper is GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari and colleagues at IIT Delhi and Princeton, first published in November 2023 and presented at KDD 2024. It is the founding study of the field and, as far as we can find, still the only one that tested keyword stuffing head to head against other rewrites.
The set-up was simple. They built GEO-bench, 10,000 queries drawn from real search logs and question sites. For each query they took the top five Google results, handed all five to GPT-3.5, and asked it to write a cited answer. Then they rewrote one of the five sources using one of nine methods. The measure was how much more, or less, of the answer that source now got.
The nine methods were: authoritative tone, statistics addition, keyword stuffing, cite sources, quotation addition, easy-to-understand language, fluency optimisation, unique words and technical terms. The main metric, position-adjusted word count, counts how many words in the answer are attributed to your source, weighted towards the top of the answer.
Here are the overall results from Table 1 of the paper, on a baseline of 19.3 for the unedited source.
xychart-beta
title "Position-adjusted word count by GEO method (Aggarwal et al., GEO-bench, GPT-3.5)"
x-axis ["Keyword stuffing", "No change", "Unique words", "Authoritative", "Easy to read", "Technical terms", "Cite sources", "Fluency", "Statistics", "Quotations"]
y-axis "Position-adjusted word count" 0 --> 30
bar [17.7, 19.3, 20.5, 21.3, 22.0, 22.7, 24.6, 24.7, 25.2, 27.2]
Keyword stuffing came last at 17.7, about 8% under the untouched page. In the paper's smaller test on the live Perplexity engine it did worse still: the authors write that it "performs 10% worse than the baseline". At the other end, the paper reports that adding citations, quotations and statistics "achieved a relative improvement of 30-40%" on the main metric, with quotation addition topping out at 27.2, a 41% lift.
So the figure you have heard is real. It just comes with a loser attached, and the loser is the tactic most UK SEO briefs still ask for by default.
Why repeating the phrase makes things worse
A generative engine does not look for your string. It reads a handful of pages, pulls out claims it can use, writes an answer, then attributes those claims back to the pages they came from. The paper puts it plainly: the model in a generative engine "is not limited to keyword matching", so techniques built for the old setting do not carry across.
Think about what that means for a stuffed paragraph. Ten repetitions of "emergency plumber Northampton" contain exactly one usable fact, that you are a plumber in Northampton. Everything after the first mention is dead weight. When the model is choosing which sentences to lift, a paragraph with nothing extractable in it loses to a competitor's paragraph that says what the call-out charge is and how fast someone arrives.
There is a second effect. The stuffed version has pushed out words that would have carried information. The rewrite is not neutral. It replaces substance with repetition, which is why the score falls below the untouched page instead of simply failing to rise.
Both big engines now say this in their own rules. Google's spam policies, updated August 2026, define keyword stuffing as filling a page with keywords "in an attempt to manipulate rankings" and give the example of repeating a phrase "so often that it sounds unnatural". Bing went further in February 2026. According to Search Engine Journal's report, its webmaster guidelines renamed the section "Keyword Stuffing and Artificially Engineered Language" and extended it to text written to trigger AI citations. Grounding and citation are now listed as outcomes a page can be excluded from.
The same report says Bing's advice for being cited is to state facts plainly so an AI system can verify them, use consistent entity names, keep to one topic per URL, and put the important information near the top. A good page looked like that before any of this.
"Optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Google Search Central, AI features guide, July 2026
The caveat before you rewrite every page
The 30 to 40% figures describe what happens once a page is already in front of the model. In the Aggarwal experiment all five sources were placed in the context window by the researchers. The study measured share of the answer, and said nothing about whether the page would have been found in the first place.
A July 2026 survey of 45 GEO studies by Olivier Martinez at Sciences Po makes exactly this point. Its abstract says the gains are "conditional on a source already being present in a fixed context" and that no reviewed technique shows a stable, cross-platform effect on being discovered. Retrieval, in other words, is still won by the things that always won it: crawlable pages, links, and a topic the engine already associates with you. Our review of the generative engine optimisation evidence lays that out in full.
Two further details from the paper matter for a smaller business. The lift was not spread evenly. The source ranked fifth of the five gained 115% from adding citations, while the top-ranked source lost share on average. If your page is the one that scrapes in, the rewriting tactics below are worth more to you than to the market leader.
And the tactics did not stack cleanly. The best pair the authors found, fluency plus statistics, beat the best single method by just over 5%. Doing all nine to one page is not the plan.
What replaces keyword stuffing
Four things, each of which gives the model something it can lift and credit. The examples are illustrative rewrites, not client copy.
1. A cited figure
Statistics addition scored 25.2 against the 19.3 baseline, about a 31% lift, and the paper found it did most for "Law & Government" topics and opinion-style questions. The rule is one number, one source, one link, in the sentence where the claim is made.
Before: "Our Northampton SEO agency helps hundreds of local businesses with their SEO, AI SEO and GEO SEO every year."
After: "In the GEO-bench test, a page rewritten with keyword stuffing scored 17.7 against 19.3 for the untouched version, while the same page with statistics added scored 25.2 (Aggarwal et al., 2023)."
The second version can be quoted in an answer about whether keyword stuffing still works. The first cannot be quoted in an answer about anything.
2. A quotable sentence
Quotation addition was the single strongest method, 27.2, and worked best in the paper's "People & Society", "Explanation" and "History" categories. A quotable sentence is short, makes one claim, names its subject in full, and does not depend on the sentence before it.
Before: "This is something that many of our clients find confusing, and it can lead to problems down the line if it isn't handled properly."
After: "Google treats a page that repeats the same phrase until it sounds unnatural as keyword stuffing, which is a spam policy violation."
Write the second kind of sentence at the top of the section it belongs to, so the model finds it in the first pass. Bing's advice about putting the important information near the top is the same instruction from the other side.
3. A named source
Cite sources came in at 24.6, a 27% lift, and the paper found it "particularly beneficial for factual questions". Name the organisation, the document and the date, and link to the page itself, not to the organisation's homepage. "According to government figures" gives the model nothing to check. "According to the ONS Business Demography bulletin for that year", with a link to the bulletin, does.
This also keeps you inside Google's current guidance, which asks for content you created yourself from what you know. The same page warns that producing content primarily to manipulate generative AI responses falls under its scaled content abuse policy.
4. An entity-clear page
This one is not in the Aggarwal paper. It comes from the engines themselves. Bing's updated guidelines ask for consistent entity names with no ambiguous references, and one topic per URL. In practice that means your company name, product names and location are written the same way everywhere on the page. The page is about one thing, and the first paragraph says what that thing is in a sentence a stranger could repeat.
Keyword stuffing was a crude attempt at exactly this. The difference is that an entity-clear page says "Growthcart is a digital marketing agency in Northampton" once, near the top, and then gets on with the subject. A stuffed page says it eleven times and never gets on with anything.
Which replacement to reach for first
Match the method to the page type, because the paper's domain results say the gains are not uniform.
| Page type | Start with | Why |
|---|---|---|
| Fee, price or deadline page | A cited figure | Statistics did most for law, government and opinion queries |
| Explainer or "how does X work" | A quotable sentence | Quotations did most for explanation and society topics |
| Comparison or "is it true that" | A named source | Citations did most for factual questions |
| Service or location page | Entity clarity | Bing's grounding advice; nothing to quote on these pages anyway |
If you only have time for one pass across a whole site, take the pages that currently rank fourth to tenth for their topic and add the cited figure and the quotable sentence. The paper says those pages have most to gain. In our experience they are also the ones an old keyword brief did most damage to.
The decision rule we use is this: if a sentence could be lifted into an AI answer with its source attached and still make sense, keep it. If it exists only so a phrase appears one more time, cut it. Nothing in the evidence since 2023 has changed that test.
If you want a page-by-page read of where your site sits against it, our generative engine optimisation service starts with that audit. It is work we do often for law firms whose fee and service pages read like brochures rather than answers, or you can get in touch with the URLs you are worried about.