Retrieval poisoning
Retrieval poisoning is planting pages built to be fetched for certain questions so an AI answer repeats something false.
Retrieval poisoning is publishing content that's built to be fetched by AI systems for certain questions, so the answers repeat whatever it says. The content is usually false or slanted. Because the platform trusts what it retrieves, the false claim ends up in the answer, sounding like a fact.
For brands, it's one way a wrong claim about you can appear out of nowhere. Someone else's page, written to match the questions people ask about you, can put words in an assistant's mouth.
How it works
Systems that search before they answer pick the passages that match a question best. A poisoned page is written to match a question very closely, so it gets picked, and it carries the claim the writer wants repeated.
Researchers showed how little it can take. A 2024 study called PoisonedRAG found that adding a handful of planted passages to a large collection was often enough to steer the answer to a chosen question.
It differs from prompt injection, which hides orders for the model to follow. Retrieval poisoning plants claims for the model to believe.
An example
Someone sets up a page titled "Is Fernhill Coffee still delivering in 2026?" It says Fernhill has paused deliveries. It's false, but the page matches the question almost word for word. Answers to that question start citing it, and some repeat that Fernhill has paused.
What you can do
Make the true answer easy to find. Publish it plainly on your own site, in the words people use when they ask, and get it repeated on sites the platforms trust. Report the false page to the search engines and platforms that use it. Write to the host if the page breaks its rules.
In AnswerOrbit
Claims checks what each answer says about you against your facts. Inaccuracy drivers shows the pages behind each wrong claim, and Copy evidence gives you the claim and every answer that made it, ready to send to whoever runs the page.