How do you optimize content for AI search?

I published a full guide yesterday on how I’d optimize an article for AI search.

There are always bits of GEO advice floating around, some logical and some pure nonsense.

So I wanted to identify which on-page optimizations carry weight when AI crawlers are looking to retrieve content for a user. But more importantly, the data that backs these optimizations up.

Here’s what I found:

The first part is still normal SEO: make it obvious what the page is about. The difference for AI search is tweaking how the title is written.

Ahrefs analyzed 1.4 million ChatGPT prompts and compared pages ChatGPT cited with pages it retrieved but didn’t cite. Cited pages had titles that were more semantically aligned with the original prompt, and even more closely aligned with the fan-out searches ChatGPT generated.

So I wouldn’t use a title like: AI Search Optimization Guide

If the article is really about: How to Optimize Blog Posts for AI Search

The difference with AI search is that the original prompt is only part of the targeting.

LLMs can generate additional searches before producing an answer (known as query fan-out).

Peec analyzed 5 million fan-out queries and found ChatGPT generated 2.1 searches per prompt on average. Profound found 91% of ChatGPT’s generated searches were unique across repeated runs.

So I wouldn’t optimize around one keyword or exact prompt.

I’d make sure the article properly covers the recurring questions, comparisons, and criteria someone would need to understand the topic.

This is one of the main changes I would make - supported by data.

If someone lands on an article asking whether schema improves AI citations, I wouldn’t spend five paragraphs explaining structured data first. That was the typical layout for older-style articles.

I’d start with the answer:

Adding schema hasn’t been shown to increase AI citations.

Then explain the study and limitations behind that conclusion.

Kevin Indig analyzed 18,012 verified ChatGPT citations and found 44.2% came from the first 30% of page content.

That doesn’t prove moving a paragraph higher will make it more likely to get cited. Important definitions and conclusions naturally tend to appear near the top anyway.

But the optimization here is not to lose the useful information near the end of the article.

This is also why I don’t think every article needs a TL;DR. The label itself isn’t the important part. Getting useful information near the top is.

See my last newsletter on TL;DRs.

This is where optimizing for AI retrieval is slightly different from traditional SEO.

An AI crawler can pull one passage from an article without using the paragraphs around it.

Semrush found AI-cited pages showed stronger associations with clarity and summarization (+32.83%), Q&A formatting (+25.45%), and section structure (+22.91%).

So a sentence like: It increased by 2.2%.

Doesn’t do anything. Compare that with:

In Ahrefs’ 2026 matched-control study, ChatGPT citations increased by 2.2% after pages added JSON-LD, but the change was statistically indistinguishable from zero.

The second version carries the subject, source, data, and conclusion with it.

I found something similar in my own retrieval experiment.

Structured content appeared in the top five results 80% of the time compared with 68% for dense prose. But when I broke the same information into very short Q&A fragments, retrieval dropped to 12%.

So I wouldn’t “chunk” content by making everything shorter.

I’d make each section clear on its own while keeping enough context for the information to still mean something.

Retrieval is only one part of the process. Once information is retrieved, it then goes through a ranking process that cuts the content AI doesn’t need.

Ahrefs found that large numbers of pages entering ChatGPT’s retrieval pipeline never became citations.

That means being relevant enough to get retrieved isn’t the same as your content being good enough to be used as a source.

This is where evidence and data can optimize your content.

Instead of: AI tools prefer fresh content.

I’d write: AI tools prefer fresh content, which is supported by Ahrefs who analyzed almost 17 million citations and found URLs cited by AI assistants were 25.7% newer on average than URLs appearing in Google’s organic results.

Now the claim is backed by a source, data, and result.

I’d do the same with original information.

For example, Kevin Indig and Amanda Johnson audited 301 pages carrying 1,075 AI citations. Only eight qualified as primary research, but those eight averaged 11.3 citations per page compared with 3.4 for the remaining pages.

It’s a small subset, so I wouldn’t turn that into ‘original research gets 3x more citations.’

But it points in the same direction: if your article contains the same information as every other result, there is less reason to use it.

A test you ran, proprietary data, screenshots, expert input, a benchmark, or a point of view backed by data gives the article something everyone else doesn’t have.

The remaining things I would focus on are not new:

Although optimizing for AI search tweaks the final output, I wouldn’t write articles as “AI-First” - instead, my workflow is to write the article as you normally would with user and SEO in mind.

Then run through an optimization checklist for AI search and tweak everything where necessary.

I put the full 10-step guide, examples, and research here: