Asked & Cited

Research explained

What the original GEO research found about getting cited by AI (and what it did not)

A plain-English reading of the KDD 2024 paper that tested nine content changes on AI answers. Quotes, statistics and citations helped. Keyword stuffing did not.

Most advice about being named in AI answers is opinion. One piece of work is different: a peer-reviewed study that took nine kinds of content change, applied them to real web pages, and measured what happened to how visible those pages were in generated answers. It was accepted to KDD 2024 and is the origin of the term "generative engine optimization". Here is what it tested, what it found, and how far you can stretch the findings.

What the researchers did

The authors built GEO-bench, a set of 10,000 queries across 25 domains, each paired with the cleaned text of the top five Google results. For each query they picked one source page at random, rewrote it using one method at a time, and had a language model write an answer from the five sources. They then measured how much of the answer was attributed to the rewritten page, with and without the change. Each result is an average over five random seeds.

They also repeated part of the test on Perplexity.ai, a live answer engine, to check that the findings were not an artifact of their own setup.

The nine changes they tested

What worked

On the paper's main visibility score (position-adjusted word count), an unmodified page scored 19.3. Quotation addition scored 27.2, statistics addition 25.2, and cite sources 24.6. The authors report the best methods improved visibility by up to roughly 40 percent over the baseline. Fluency, technical terms and simpler language gave smaller gains.

The common thread is that the winning changes add verifiable substance. A page that says "many teams reduce onboarding time" gives an answer engine little to quote. A page that states a figure, names where it came from and quotes a source gives it something concrete to lift into an answer.

What did not work

Keyword stuffing scored 17.7 on the same measure, below the 19.3 baseline. On Perplexity it did about 10 percent worse than doing nothing. Adding unusual vocabulary did roughly nothing. If your content process still revolves around repeating the target phrase, this is the clearest evidence in the study that it does not carry over to answer engines.

Limits to keep in mind: the main experiments used a simulated engine built on GPT-3.5 and Google's top five results, and the authors note that methods may need to adapt as engines evolve. The effect also varied by domain, so a change that helps a history page may not help a software comparison page. Treat the results as strong directional evidence, not a guarantee.

What this means for a software company's site

The paper tested editing existing pages, so these steps are our reading of it, not something the authors prescribed for software vendors.

  1. Replace adjectives with numbers. "Fast setup" becomes "median setup time of 20 minutes, measured across the last 50 customers", but only if that is true and you can show how you measured it.
  2. Cite where claims come from. Link the analyst report, the standard, or your own methodology page. Pages that show their sources give an engine something to attribute.
  3. Quote real people. Customer and expert quotes with names and context, on product, comparison and docs pages.
  4. Write plainly. Clear, fluent text outperformed the baseline too, and it costs nothing.
  5. Stop repeating the keyword. It is not neutral. In this study it lowered visibility.

None of this replaces being present on the pages engines retrieve in the first place. The study improved a page that was already among the top five sources. If you are on none of those pages for your buyers' questions, start with presence, then improve the substance of what is there.

Sources

Where do you stand?

Send your website and the one buyer question you most want to be the answer to. We will send back a free snapshot within two business days.

Get a free visibility snapshot Book a call