Do GEO tactics improve AI retrieval?

I wanted to test whether GEO-optimized content affected AI retrieval, so I created an experiment using Cloudflare AI Search.

I wrote 20 marketing studies and two versions of each article.

One was the control: Written like a normal editorial article with longer sections, more narrative, and some points relying on the surrounding paragraphs for context.

The variable used the same facts but changed the structure: more specific headings, answers under the heading, one main intent per section, clearer topic/entity references, and sections that make sense by themselves.

All the GEO optimization practices that are floating around.

Same topic. Same facts. Same H1. Same domain. Same design. Same metadata setup. Similar word count. Same crawl and indexing setup.

40 pages in total.

Cloudflare isn’t ChatGPT or Google, but it follows a standard RAG pattern used in AI search: index the content, split it into chunks, embed it, then rank the chunks against a query before an LLM writes the answer. Google publicly describes the same retrieve-then-generate pattern for its own generative Search features.

So the experiment doesn’t tell me how ChatGPT or Google would rank these pages. It does give me a controlled way to test what I wanted though.

I started with the control and GEO pages in separate mirrored indexes.

For 320 searches (multiple queries across the 20 topics):

Across the 20 topics, GEO won 4, control won 7, and 9 tied.

So there was no GEO advantage here.

Some of the failed queries were too vague. They made sense if you already knew which study they referred to, but not when the retriever had 20 studies to choose from.

I added neutral topic context to those queries and ran them again.

Both versions hit 100% Top-3 retrieval.

Changing the query made a bigger difference than changing the content in those first tests.

But now everything was being found, so I changed the test.

For Experiment 3, all 40 pages went into one Cloudflare AI Search index.

Every query now had both versions competing at the same time.

Same query. Same facts. Same retrieval model. Same candidate pool.

For 120 direct comparisons:

The GEO version performed better than the control on 93.3% of the searches and won across all 20 topics.

The control wasn’t disappearing either.

Both versions appeared somewhere in the Top 10 on every search.

The variable (GEO version) appeared in the Top 3 on 98.3% of searches compared to 60% for the control. It also held the #1 position on 111 of 120 searches.

So normal content was still being retrieved. The GEO version was just being pushed much higher.

The 93% result needed another test.

The GEO pages had been structured around likely user intents. The first set of questions may have been too similar to the way the GEO pages were structured.

So I froze the pages. No rewrites, new headings, or changes to the variable.

Then I wrote 60 new questions after the content had already been fixed. The new set included decision, explanation, and comparison questions.

This time:

GEO still won 68.3% of the direct comparisons.

Across the 20 topics, GEO won 13, and the control won 7, although that difference wasn’t statistically significant.

More importantly, the win rate dropped from 93.3% in the first direct test to 68.3% with the new questions.

That suggests some of the original advantage came from the GEO pages being structured around the questions I expected people to ask.

When I changed the questions, GEO still won more often, but the gap was much smaller.

I think optimizing content for GEO does create an advantage for AI retrieval.

The first two experiments showed that both versions could be retrieved. The bigger difference appeared when they competed for the same query. 

GEO won 112 of 120 direct comparisons, and even after I froze the content and tested 60 new questions, it still won 41 of them.

The drop from 93.3% to 68.3% is important though. It shows the effect is strongest when the content is structured around questions people are likely to ask. It still carried over to new questions, just nowhere near as strongly.

There’s further evidence that extends past my test.

A Semrush study found that pages being cited scored higher for clarity and summarization (+32.8%), Q&A formatting (+25.5%), and section structure (+22.9%).

Those are very close to the changes I made between my control and GEO versions: clearer sections, more direct answers, and less reliance on the rest of the article for context.

My experiment gives a different angle: I held the underlying information constant and changed the structure. When the two versions competed, the structured version was retrieved first more often.

One important thing to remember here is that retrieval is only the first stage.

An Ahrefs study found that only around half of the URLs entering ChatGPT’s retrieval pipeline ended up being cited. Their research also found that semantic relevance to ChatGPT’s fan-out queries played a role in which pages moved further through that process.

So even if these GEO tactics improve the chances of a section being retrieved, that doesn’t mean the page will automatically be cited.

It has to get retrieved first. Then it still has to get through whatever citation-selection process that AI system applies afterwards.