Last month I watched Gemini mention my brand for the keyword "pain management." Good news, right? Then I looked at the source it cited. It wasn't our website. It was a press release we published on a third-party site three years ago — and had completely forgotten about.
AI engines cite third-party websites about your brand because they select the most extractable, authoritative passage available — not the official one. Your website competes against every review site, press release, and forum thread that mentions you, and it often loses. Here's why that happens, and what I learned from watching it happen to my own brand.
Three years ago, in my healthcare SaaS days, we did classic link building: a press release on a third-party site, targeting the keyword "pain management," with that exact anchor text pointing to a specific product page. Standard SEO playbook. The campaign ended, we moved on, and I forgot the press release existed.
Three years later, Gemini answers pain management software questions by naming our brand — and cites that press release as its source. Not our homepage. Not the product page we pointed the anchor at. The press release itself.
When I stopped being annoyed about the source, I realized what I was actually looking at: the anchor text association we built in 2022 had become an AI citation in 2025. We taught the machine that our brand belongs to "pain management" — and the machine remembered, long after we'd forgotten the campaign.
Because AI engines don't give your website any special status. When Gemini composes an answer, it retrieves candidate passages from everywhere — your site, review platforms, press releases, forums — and cites whichever passage answers the question most directly. Your homepage competes against every page that has ever mentioned you.
And here's the uncomfortable part: press releases are built to win that competition. Ours had exactly what AI extraction rewards — a dated, factual, quotable statement: company name, what the product does, for whom, in plain declarative sentences. Our own product page, meanwhile, said things like "streamline your workflow." If a machine had to quote one of those two sources to answer a question, I'd pick the press release too.
Your website doesn't automatically represent your brand to AI engines. The most quotable sentence about you on the internet does — wherever it lives.
The lesson I keep coming back to: that citation wasn't random. We chose "pain management" as the anchor text deliberately. The association stuck — not just in Google's link graph, but in how AI systems understand what our brand is for.
Most people still think of anchor text as a PageRank tactic — something you do for a few months to move a ranking. What I saw suggests something bigger: every anchor you place in third-party content is a label you're attaching to your brand in the AI's memory. And those labels outlive the campaign by years. It's also the clearest real-world proof I've seen of what Koray Tuğberk GÜBÜR calls historical data in topical authority — old signals don't expire, they compound.
Three practical rules I've drawn from this:
My press release happened to age well. But think about what else lives in a 3-year-old announcement: old pricing, discontinued features, positioning you've moved away from, executives who left. AI engines repeat all of it with total confidence, because to them a confident 2022 press release and your current homepage are just two competing sources — and the older one might be more extractable.
You can't fix this by editing your website, because your website isn't the source. You first have to know which sources AI engines are using for your brand. That's a different question from "am I mentioned?" — and honestly, it's the question that surprised me more when I started tracking it. (I explain the difference between mentions, citations, and share of voice in the AI brand mentions glossary entry.)
If AI engines are citing third-party content about your brand, treat it as a diagnosis, not an insult. It tells you three things: your brand has entity associations strong enough to be mentioned (good), your own content is losing the extraction competition (fixable), and content you don't control is defining you (worth monitoring).
I found a forgotten 3-year-old press release speaking for my brand. What's speaking for yours? Citevis shows every mention and its source across ChatGPT, Perplexity, Gemini, and Claude.
Check Your Brand's Sources →