Generative Engine Optimization Examples

What a GEO edit actually looks like

Most writing about generative engine optimisation stays at the level of principle. This is the same advice at the level of a sentence, with the before and after next to each other, because the gap between "add statistics" and knowing what to type is where the work actually lives.

Everything below applies one of the three techniques a controlled study at KDD 2024 found to beat baseline: cite sources, add quotations, include statistics. The same study found keyword stuffing performed worse than the baseline, so nothing here involves repeating a phrase.

Example 1: an adjective becomes a measurement

Before	After
Our platform is extremely fast and highly reliable, delivering industry-leading performance for demanding workloads.	Median response time is 57 milliseconds across a 15-query benchmark, and the endpoint answered 94.9 percent of 8,593 probes over the last quarter. Both figures and the benchmark queries are published at example.com/benchmark.

Why it works. The first version contains nothing a model can quote, because "extremely fast" is not a fact about anything. The second gives a generated answer something specific to say and a reader something to check. It also survives extraction: read it alone and it still means something.

The uncomfortable part. You have to have the number. Most pages read like the first version because measuring is harder than adjectives, not because the writer preferred adjectives.

Example 2: a claim gets a source

Before	After
Studies show that AI search is rapidly replacing traditional search for product research.	Research published at KDD 2024 by a team at Princeton tested nine method families against a benchmark of queries and found three that measurably improved citation rates: citing sources, adding quotations, and including statistics. Keyword stuffing scored below the baseline it was tested against.

Why it works. "Studies show" is the tell of a claim nobody checked. Naming the venue, the year and the finding makes the sentence usable as a citation rather than as filler. A model attaching a source to a claim needs a source to exist.

Example 3: a paragraph is made to survive extraction

Before	After
This is why it matters so much. As we saw above, the implications are significant, and that is before considering the cost side of the equation.	Publishing a capability manifest costs about an hour and puts a domain ahead of nearly all of its competitors: a crawl of 375,998 domains in September 2026 found 207 serving one, and 176 of those were on a superseded path.

Why it works. The first paragraph is entirely made of references to other paragraphs. Cut it out and it says nothing at all. The second names its subject, states one claim, and carries its own evidence.

Example 4: a heading becomes the question

Before	After
<h2>Our Approach</h2> followed by three paragraphs of context before the answer	<h2>How long does it take to publish an ARD manifest?</h2> followed immediately by: About an hour for a single domain, most of which is deciding what to write in the query list rather than producing the JSON.

Why it works. Retrieval matches a question to a passage. A heading that is the question, with the answer in the first sentence below it, is the easiest possible match. "Our Approach" matches nothing anybody has ever asked.

Example 5: a comparison table instead of prose

Prose comparing three options across four dimensions is hard to extract because the facts are spread across sentences. The same content as a table puts each fact in an addressable cell.

Layer	Winning is	What you publish
SEO	A click	Pages and links
GEO	A citation	Statistics and sources
Agentic Optimization	A call	A capability manifest

This is also honest formatting rather than a trick: if the content really is a comparison across dimensions, a table is what it was always meant to be.

Example 6: the quotation you did not write

Adding quotations was one of the three techniques that beat baseline, and it is the one most often faked. A fabricated testimonial is worse than nothing, and invented customer counts are a liability rather than a lever.

The version that works uses quotations from sources you cite rather than praise you invented. Quoting a specification, a study or a public document, marked as a quotation and linked to the section it came from, adds exactly the property the study measured, and every part of it is checkable.

What to do with this on a real site

1. Take your ten highest-traffic pages. Retrieval mostly draws on what search already surfaces, so these are where an edit pays.
2. Find every adjective doing the work of a number. Fast, reliable, leading, seamless, robust. Each is a place where a measurement is missing.
3. Find every "studies show" and either source it or cut it.
4. Read each paragraph alone. Rewrite the ones that only make sense in sequence.
5. Turn section headings into the questions people ask, and put the answer in the sentence directly below.

None of that requires new pages, which is why it is usually the cheapest work available on an existing site.

Two things not to do

Do not repeat your target phrase. Keyword stuffing was measured performing below baseline. It is the one technique with evidence against it, and it keeps reappearing in GEO advice with a new name.

Do not add llms.txt and consider the job done. Analysis of hundreds of millions of bot events has found no evidence that major engines read it. Publishing one is harmless and is not a strategy.