Technical SEO for AI Citations: Which Claims Hold Up (2026 Evidence)

1 Does Adding Schema Markup Increase AI Citations? The 1,885-Page Test

This is the most important study in the whole topic because it is the only one built to separate cause from coincidence. Each treated page was matched with control pages from other domains that had similar citation levels and never added schema. If schema worked, the treated group should have pulled ahead. It did not.

The AI Overviews dip deserves context rather than alarm. Ahrefs puts the loss at around 12 daily citations per page, in a sample where most pages were getting hundreds, and both groups were already falling before anything changed. The honest reading is that schema had no clear effect in either direction.

There is one real limit. Every page in the study was already inside AI’s consideration set. The study cannot say whether schema helps a page nobody cites yet get picked up, and it only looked at schema in the HTML, not schema injected by JavaScript.

If you want certainty for your own site, run a small version of the same test. Pick a handful of pages that already earn some AI citations, add schema to half of them, leave the other half alone, and compare both groups after a month. If they move together, the platform moved, not your markup.

2 Why Do 53% of AI-Cited Pages Have Schema If It Doesn’t Cause Citations?

Semrush’s study is useful, and I use Semrush daily, but its own authors are careful: the patterns in 5 million cited URLs point to strong technical foundations, not to any single lever. Well-maintained sites add schema, write tidy URLs, load quickly, earn links and publish better content. AI systems cite those sites. Remove schema from the picture and the rest of those signals still carry the page.

The slug-length finding shows why correlations need care. URLs with 17-40 characters slugs collected the most citations, but short slugs are often homepages and very long ones are often deeply nested or stuffed. Rewriting your URLs to hit a character count would cost you redirects and gain you nothing a study has shown.

The takeaway for learners is simple. When a statistic describes what cited pages have, read it as a description of good sites, not a recipe. Keep schema for rich results and entity clarity. Just stop selling it to clients as an AI citation fix.

3 Can AI Crawlers Read JavaScript? What 569 Million GPTBot Requests Showed

This is the one technical factor where the evidence and the logic line up. If your main content only appears after client-side rendering, GPTBot, ClaudeBot and PerplexityBot receive an empty shell. No amount of content quality fixes a page the crawler never sees. Server-side rendering, static generation or pre-rendering for key pages is the fix.

Two cautions on dates and wording. This study was published in December 2024, so crawler behaviour may have shifted since. And the finding has drifted as it spread. The original says Vercel saw 569 million GPTBot requests and that no major AI crawler renders JavaScript. Many articles now say a study of five hundred million fetches found zero JavaScript execution, which is close but not what was written.

You can check your own site without trusting anyone. Your server logs show which AI bots fetch which URLs and which responses they get. My step-by-step log file analysis process works for AI crawlers too; just filter by their user agents and compare what they fetched against what a browser renders.

4 Does llms.txt Get You Cited? 97% of Files Received Zero Requests

I deployed llms.txt on this site myself, so I have no reason to talk it down. But two very different methods now agree. SE Ranking correlated adoption with citations and found nothing, and removing the file from its model actually improved the model’s accuracy. Ahrefs read server logs and found the bots barely ask for the file.

Adoption figures vary wildly, from 10.13% in SE Ranking’s sample to 28% among Ahrefs’ more technical customers. Neither describes the whole web, so treat any single adoption number with care.

What the logs do show is worth acting on. AI bots fetch your normal HTML pages constantly, so the effort that would go into polishing an llms.txt file is better spent making sure those pages load cleanly, return the right status codes and say something specific enough to quote.

My position: publishing llms.txt is cheap housekeeping and harmless. Just do not put it in a proposal as a citation lever, and do not let it replace the work that the rendering and content sections of this page point to.

5 Does Content Freshness Affect AI Citations? 17 Million Citations Compared

Freshness is where good practice and hype collide. The data supports keeping important pages genuinely current, and ChatGPT shows the strongest pull towards newer pages. Google AI Overviews behaved differently, citing pages slightly older than the organic results.

What the data does not support is changing a date without changing the page. The study measured publication and update dates, not whether an edit caused a citation. Update when the facts change, add what is new, and record what you changed. That is the same discipline I recommend for anyone trying to write content that ranks on Google and gets cited by AI.

Keep the averages in proportion too. AI-cited pages were still close to three years old on average, so long-lived reference content keeps its value.

6 Which Technical AI SEO Claims Have No Primary Source?

None of these may be wrong in spirit. JavaScript rendering really is a problem, and stale pages really can fade. The issue is precision. A specific percentage suggests someone measured it, and when a learner repeats it in an audit or a client pitch, they are vouching for a study that may not exist.

Page speed is the clearest blind spot. It is widely listed as an AI citation factor, and speed genuinely helps users and Google rankings, which feed AI Overviews indirectly. But I could not find a public controlled test isolating speed or Core Web Vitals as a cause of AI citations. Until one exists, call it a supporting factor, not a proven lever.

Here is a simple test before you quote any AI SEO number: can you name the organisation, the sample size and the year, and open the page where it lives? If not, leave it out or label it. That habit is the backbone of the approach in my Hybrid Engine Optimisation guide, and it will make your recommendations harder to argue with.