Research
Where AI Engines Pull Their Citations From
Most AI citations come from the first third of a page. Visuals get picked far more than text alone. And a win on one engine rarely carries to the next.
Plenty of research already tells you which pages get cited by AI engines. Less of it tells you where on the page the citation actually comes from, and that turns out to be just as actionable.
Three findings from 2026 studies change how you should structure a page if you want it quoted. None of them is about authority. All of them are about placement, format, and platform.
Here is what the data says and what to do with it.
Citations come from the top of the page
Zyppy analyzed thousands of ChatGPT citations and found that 44.2% of them came from the first 30% of the page. The introduction, the opening section, and any summary block near the top account for nearly half of all citations. The middle third of a page contributes another 31%. The bottom third barely registers.
The takeaway is blunt: if your answer lives in the fifth section, an AI engine may never reach it.
This rewards a structure that classical SEO often discouraged. The old instinct was to build toward the answer, setting context first and paying off the reader at the end. AI extraction inverts that. Put the direct answer in the first paragraph or a short summary block, then expand below it for the reader who wants depth.
A practical test: read only the first two paragraphs of your page. If they do not contain a quotable, self-contained answer to the query, an engine has little to lift from the place it looks hardest. The guide to structuring pages for AI extraction goes deeper on this.
Multimodal content gets selected far more
A separate analysis of nearly 16,000 AI Overview results across 63 industries found that content combining text with images or video had a 156% higher selection rate than text-only content.
That is a large gap, and most informational pages are still text plus a single stock image. A page with an original chart, a labeled diagram, or a short explanatory video gives an engine more to work with and signals a more complete treatment of the topic.
The word “original” matters. A generic stock photo adds nothing an engine can use. A diagram that encodes real information, a chart built from your own data, or a screenshot that shows the thing you are describing all add extractable substance. The format is doing work, not decoration.
For text-heavy sites, this is the most underused lever available. Adding one genuinely informative visual to a strong page is often a faster win than another 500 words.
Each engine cites a different set of sources
The third finding reshapes how you measure success. Across ChatGPT, Perplexity, and Google AI Overviews, only 2% of cited URLs appeared in all three. Ninety-one percent of citations showed up on a single engine.
Getting cited by Perplexity tells you almost nothing about whether ChatGPT or an AI Overview will cite the same page for the same query. The engines retrieve differently, weight sources differently, and draw from different indices.
Two consequences follow. First, you cannot optimize for “AI search” as one target. You optimize for each engine and track each separately, because a win on one does not carry to the others. Second, a single citation-tracking snapshot from one platform badly understates or overstates your real visibility. You need coverage across engines to see the picture. The post on how ChatGPT, Perplexity, and Google cite differently breaks down the per-engine behavior.
What to do with this
The three findings stack into a short set of changes.
Front-load the answer. Lead every informational page with a direct, self-contained response to the query it targets, before any setup. Depth goes below, not above.
Add one real visual. Give your strongest pages an original chart, diagram, or screenshot that carries information a reader and an engine can both use. Skip the decorative stock image.
Track per engine. Measure citation presence on ChatGPT, Perplexity, and Google AI Overviews separately. Do not assume a win on one transfers.
One caveat keeps this honest. Placement and format help only on a page that is eligible to be cited in the first place. If the page lacks the authority and trust signals an engine needs, front-loading the answer changes nothing, because the engine never considers it. Structure is a multiplier on a page that already earns consideration, not a substitute for earning it. The work behind that eligibility is covered in how to get cited by AI search systems and what AI visibility is.
FAQs
Where on a page do AI engines pull citations from?
Most come from the top. Zyppy found 44.2% of ChatGPT citations come from the first 30% of a page, with another 31% from the middle third. The bottom of the page is rarely cited, so the direct answer should appear near the top rather than at the end.
Does adding images or video really help with AI citations?
Yes. An analysis of nearly 16,000 AI Overview results found multimodal content (text plus images or video) had a 156% higher selection rate than text-only content. The visual needs to carry real information, such as an original chart or diagram, not a decorative stock photo.
If I get cited by one AI engine, will the others cite me too?
Usually not. Only 2% of cited URLs appear across ChatGPT, Perplexity, and Google AI Overviews together, and 91% appear on just one engine. You have to optimize for and track each engine separately.
Is page structure more important than authority for AI citations?
No, it is a multiplier on top of authority. Placement and format help only on a page an engine already considers trustworthy enough to cite. Without the authority and trust signals, structuring the page changes nothing.