Editorial correction: . Examples are writing guidance, not citation requirements.
Two suppliers pitch you. The first says they are reliable, experienced and committed to excellence. The second says they handled 4,200 containers last year, that their average clearance took 1.8 days, and quotes their largest customer by name saying so.
The second pitch provides claims you can check. Specificity alone does not make them true: verify the records, reporting period and permission to quote the customer. Do not assume a language model applies a human buyer’s judgment.
The experiment
The 2024 paper that named this field — GEO: Generative Engine Optimization, presented at KDD 2024 — is the closest thing to a controlled trial we have. The authors built GEO-BENCH, a benchmark of 10,000 queries drawn from nine sources including MS MARCO, Natural Questions, ORCAS and real Perplexity queries, then rewrote source content nine different ways and measured which versions got pulled into generated answers more often. TIER A
| Rewriting method | Visibility change | What it means |
|---|---|---|
| Add quotations | +40.5% | Quote a named, credible person |
| Add statistics | +32.3% | Replace qualitative claims with numbers |
| Improve fluency | +28.7% | Make the writing flow properly |
| Cite sources | +27.6% | Reference credible external sources |
| Authoritative tone | varies | Confident, declarative writing — best in debate and history topics |
| Easy-to-understand | modest gain | Simplify language, keep the information |
| Technical terms | modest / mixed | Add domain-specific vocabulary |
| Unique words | modest / mixed | Rarer vocabulary |
| Keyword stuffing | −9.1% | The classic SEO move. Negative result in this experiment. |
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization", KDD 2024. Figures shown are on the Position-Adjusted Word Count metric; the paper also reports a Subjective Impression metric where Statistics (+29.8%) and Quotations (+27.8%) again lead. Headline finding: up to 40% visibility improvement.
Three things in that table deserve to be read twice.
Keyword stuffing produced a negative result. Repeating the target phrase reduced the study’s measured visibility in this experiment. If you take one thing from this chapter into a conversation with an agency, take that.
Fluency scored as highly as citation. Simply making the writing good was worth roughly as much as adding sources. This is not usually how optimisation works, and it is a reassuring signal about where the systems are pointed.
The effect was strongest for the underdog. Sites ranked lower in results gained far more than sites at the top. Adding source citations lifted a fifth-ranked site's visibility by 115.1% in the paper's measurement. If you are not the biggest name in your category, this is your chapter.
It was published in 2024 and tested against the generative engines of that moment; the systems have changed. And its visibility metrics are proxies, constructed by the authors, not revenue. Treat both the ordering and the percentages as results of that experiment, not a durable ranking of tactics for every current engine or industry. Domain, query, model and measurement method can change the outcome.
What this looks like on a real page
"We have extensive experience handling hazardous goods and a strong track record with the authorities."
"In 2025 we filed 3,928 import declarations covering 10,400 containers at Port Klang, for 56 customer groups. Roughly two-thirds of those containers were 20ft."
One sentence became four verifiable statistics. None of them is a claim about how good the company is — which is precisely why they carry weight. Every figure here comes from DNE's operational records; see the method note in Chapter 10.
Four editorial practices to consider
- Quote someone real and name them. Your operations manager, a customer, a regulator's published statement, an industry body. A name and a role attached to a sentence changes its status.
- Turn every adjective into a number. "High volume" is nothing. "10,400 containers across 3,928 declarations in 2025" is a fact. Go through your site and hunt adjectives.
- Cite the source of every external claim. Link to the regulation, the study, the official figure. It costs a sentence.
- Then edit for fluency. Read it aloud. Fix what makes you stumble. This scored as highly as the clever techniques.
Take any page and count two things: the number of specific figures, and the number of external citations. For each consequential claim, check whether the source supports it and whether the context is clear. Add relevant evidence where it helps; do not insert statistics or citations merely to meet a quota.
Go through your five most important pages and replace every unsupported superlative with a number you can defend. If you cannot defend a number, delete the claim rather than softening it.