Most product teams sketch AI citations as tidy little blue numbers sitting neatly beside generated text. They wire them up once on a fast connection, watch the preview pop open in half a second, and move on. That works in the staging environment. It rarely survives contact with the open internet.
Sources rot. Companies delete blog posts. Government portals reorganize their URLs without redirects. Paywalls go up. CDNs hiccup. The document an AI scanned three months ago might have been rewritten yesterday. If your citation component only knows how to look good when everything goes right, it will lie to your users when things go wrong. Worse, it will leave them confused about whether the evidence ever existed at all.
A reliable citation system needs to communicate state. Not just a hyperlink. Not just a tooltip. It needs to tell the reader what is happening with the evidence in real time.
Design for four real states
Think of a citation as a small status machine. Every source your AI surfaces can land in one of four conditions, and your interface should handle each one honestly.
Loading. The user clicks the superscript number, and the system fetches a preview, a snippet, or the original page. During that gap, show them something is happening. A skeleton placeholder, a pulsing excerpt frame, or even plain text reading "Fetching source..." beats silence. Silence makes people wonder if the click worked. In a streaming interface where sources arrive as the model thinks, loading states also buy you time to resolve links without blocking the text generation.
Available. This is the state everyone designs for. The link resolves, the text still matches what the model saw, and the user gets a clear excerpt plus a direct link. But do not stop at dumping a raw URL. Show the title, the domain, and the relevant passage. Give the reader enough context to decide if they want to leave the page. A raw href in parentheses is technically a citation. It is not a good user experience.
Changed. This one stings. The page loads, but the content no longer contains the quote or statistic the AI attributed to it. Maybe the publisher updated the article. Maybe the AI training snapshot caught a headline that later changed. Your component should flag this discrepancy instead of pretending everything lines up. A small note like "This source has been updated since it was indexed" prevents the user from wasting ten minutes hunting for a sentence that no longer exists. It also preserves trust. If you hide the mismatch, the user will eventually assume your product makes things up.
Missing. The 404. The timeout. The paywall. The server that sends back a 200 status code but serves a generic "Content Removed" page. Do not let the link hang there in blue, pretending it works. Replace it with plain language: "Page removed," "Link no longer available," or "Source offline." If you know why it failed, say so. If the Internet Archive or a cached copy exists, offer it. The goal is to close the loop, not strand the reader.
Accessibility is non-negotiable
Citation components sit in high-traffic, high-stakes parts of the UI. They deserve the same rigor as a checkout form or a flight booking flow.
Use text to show status. Colorblind users cannot rely on a green dot or a red border alone. Pair your indicators with words: "Available," "Modified," "Unavailable." That tiny label helps every sighted user too, because status colors rarely mean the same thing across cultures or products.
Write descriptive names for screen readers. Label your links so a person using a screen reader hears something useful. Instead of a generic "Link," read out "Source 3: Audit-log retention policy." This tells the listener exactly what they will get before they decide to open it.
Protect the flow. Clicking a citation should not vaporize the AI-generated answer. Use a panel, a modal, or an inline expansion. The reader opened the source to verify a specific claim. If your interface wipes away the claim they were just reading, you have forced them to hold the text in working memory while they hunt for proof. That is unnecessary cognitive load.
Manage focus deliberately. When a user opens a citation, move keyboard focus into the new panel or popup. When they close it, send focus back to the trigger. Do not hijack focus while the stream is still writing sources to the page. Let the user stay in control.
Handle streaming with care. If your generative interface is adding citations as they arrive, do not bombard screen-reader users with an announcement for every single new superscript number. Batch them. When two or three have accumulated, announce a polite summary like "Two sources added." Constant interruptions make the output unreadable for assistive technology users.
Remember what a citation proves
This point cannot be repeated often enough. A citation shows provenance. It does not show truth.
When an AI cites a source, it is saying, "I found this wording here." It is not saying, "This expert is correct." It is not saying, "This study was well-conducted." It is not even saying, "I understood the context correctly." A citation is a map back to the original evidence. The reader still has to walk that path and decide whether the evidence holds up.
Your design should make that distinction clear. Use excerpts that let the user scan the relevant lines. Show the publication date so the reader knows how fresh the evidence is. If the source is a forum, a press release, or a primary study, surface that context so the reader can weigh it appropriately. Build the interface to support human judgment, not replace it.
Test against reality
A citation component that passes a visual review is only halfway done. You need