If you are trying to figure out how to optimize for AI search in 2026, the honest starting point is that half the advice circulating right now has already been tested and disproven. Google AI Overviews now show up on a large majority of commercial search queries, and ChatGPT alone serves several hundred million people a week, so getting cited inside an AI-generated answer matters as much as ranking in blue links. But the tactics people are rushing to adopt, especially schema markup and llms.txt files, have produced genuinely mixed results in controlled studies. This post looks at what the actual research says works, what does not, and what a normal business website can realistically do about it.
Why answer engine optimization is not just SEO with a new name
Traditional SEO optimizes for a ranking: you want to be link number one, two, or three on a results page. Answer engine optimization, often shortened to AEO, or generative engine optimization, shortened to GEO, optimizes for something different: being the source an AI model chooses to pull from, paraphrase, or quote when it writes an answer. There is no guaranteed top spot. A model might cite you, cite three competitors instead, or blend five sources into one sentence with no link at all.
The stakes are also different once you do get cited. Data from AI visibility firms tracking referral traffic shows that visitors who arrive from an AI answer convert at a noticeably higher rate than typical organic search visitors, in some tracking as high as 15 to 16 percent from ChatGPT referrals versus under 2 percent for average organic search traffic. Fewer people click through from an AI answer, but the ones who do have usually already had their questions answered and are further along before they land on your site.
How to optimize for AI search, according to a real academic study
The most cited research in this space is a 2024 paper from Princeton, Georgia Tech, and IIT Delhi researchers (Aggarwal et al.), later expanded on through 2026, which built a benchmark of 10,000 real queries and tested nine different content strategies against a generative search system. Two tactics stood out. Adding citations to credible external sources improved a page's visibility in AI-generated answers by as much as 115 percent for content that started out ranked low. Adding concrete statistics to a page improved visibility by around 41 percent. Simply repeating keywords, the old SEO reflex, barely moved anything.
The practical takeaway is almost the opposite of classic SEO instinct: instead of optimizing a page to look authoritative through length and keyword density, you make it easier for a model to lift a self-contained, factual, sourced chunk out of your page and use it directly. Write the kind of sentence you would not mind seeing quoted verbatim in someone else's answer.
The schema markup fight nobody has settled
This is where it gets genuinely messy, and where a lot of AEO advice is quietly wrong. In 2026, two well-resourced studies looked at the same question, whether adding schema.org structured data increases AI citations, and reached opposite conclusions.
Ahrefs ran a controlled study tracking 1,885 pages that added JSON-LD schema between August 2025 and March 2026, compared against 4,000 matched pages that never added it. Using difference-in-differences analysis, a method built specifically to isolate cause from coincidence, they found no statistically significant citation increase on Google AI Overviews, Google AI Mode, or ChatGPT. Their conclusion was blunt: if a page is already being picked up by AI systems, adding schema does not push it higher.
Around the same time, a Princeton and Moz study found the opposite using a correlational approach: sites with properly configured FAQ, Article, and HowTo schema were 3.2 times more likely to be cited in Google AI Overviews than sites without it, with the strongest results coming from FAQ blocks containing 3 to 7 questions answered in 100 to 200 words each. To add to the confusion, Google announced in May 2026 that it was removing FAQ rich results from search entirely, killing off the visible payoff of FAQ schema in traditional search results even as its value for AI systems remains disputed.
The reasonable reading of all this: structured data is not a magic switch, and it will not rescue a page nobody already trusts. But it costs little to implement correctly and may help models parse question-and-answer content faster, so there is limited harm in doing it well, as long as you are not counting on it as your main strategy.
llms.txt: the file almost nobody's crawler actually reads
A second widely recommended tactic, publishing an llms.txt file that summarizes your site for AI models, has run into a similar reality check. An SE Ranking scan of 300,000 domains found roughly 10 percent adoption after more than a year of industry conversation, and only about 7 percent of Fortune 500 companies had shipped one by the end of March 2026. The bigger problem: an analysis of over 500 million AI bot visits across a 90-day window found that major crawlers, including GPTBot, ClaudeBot, and PerplexityBot, almost never request the file directly, favoring ordinary HTML crawling instead. As of early 2026, no major AI lab had publicly committed to reading or acting on llms.txt in production. The one place it does appear to help is developer documentation, where coding assistants are the one AI consumer that reliably reads these files today.
If you run a documentation site for a developer tool, an llms.txt file is worth the twenty minutes it takes to write. If you run a small business website, it is not currently doing the work its advocates claim.
Where AI answers actually pull their sources from
A large-scale analysis covering roughly 680 million citations across ChatGPT, Google AI Overviews, and Perplexity found that listicles account for close to 22 percent of citations, more than any other content format, followed by standard articles and product pages. It also found that only about 11 percent of domains get cited by both ChatGPT and Perplexity, meaning the two systems largely draw from different pools of sources rather than one shared "best of the web" list.
Third-party platforms matter more than most businesses expect. Research from SE Ranking found that domains with heavy brand presence on Reddit averaged roughly seven ChatGPT citations compared with 1.8 for domains with minimal Reddit presence, a near four-times gap. But this channel is volatile, not a settled tactic: ChatGPT sharply reduced how often it cited Reddit threads in several categories in August 2026, with some trackers reporting drops of 70 percent or more in a matter of weeks. Betting your entire visibility strategy on one third-party platform is risky when that platform's weighting can shift with a single model update.
A practical checklist for how to optimize for AI search
None of this requires an agency contract or new software to begin. A reasonable, evidence-backed starting list looks like this:
- Write the first sentence after every heading as a complete, self-contained answer, not a teaser that requires reading the rest of the paragraph.
- Add real numbers with a named, checkable source rather than vague claims. This is the single tactic with the strongest research support.
- Link out to credible external sources when you reference a claim. Counterintuitively, citing others appears to help you get cited, not just the site you link to.
- Keep FAQ-style sections short and specific, three to seven questions, answered plainly, rather than long exhaustive lists.
- Maintain a presence on the platforms AI systems already pull from for your category, whether that is Reddit, review sites, or industry forums, but do not treat any single one as permanent.
- Check how your own brand or product is actually being described by asking ChatGPT and Perplexity directly, on a regular basis, rather than assuming your SEO rankings tell the whole story.
The takeaway
The uncomfortable truth about answer engine optimization right now is that it is still being figured out in public, including by the companies selling tools to help you do it. Some of what gets marketed as an AEO best practice, like schema markup and llms.txt files, has already been tested by serious researchers and found to do far less than advertised. What has actually held up under scrutiny is much simpler and less technical: write clear, factual, well-sourced sentences that could stand alone if quoted, and keep checking how AI systems describe you instead of assuming last year's SEO playbook still applies. If you want a sense of how much AI usage is scaling in the background of all this, Panda Tech Bytes' free AI Token Tracker is a quick way to see the volume of activity flowing through the same models now shaping how people find businesses like yours.
