Can you trick an AI engine into repeating something untrue about a market, or a brand? An informal 2026 experiment says: sometimes, for a while. The more useful lesson is why it doesn't last, and what that tells brands about earning durable visibility.
What the experiment found
SEO practitioner Lily Ray ran an informal test in 2026: she invented absurd, never-before-searched questions. Like which SEO is "best at eating spaghetti". Published satirical rankings, and watched which engines repeated them. The reported pattern:
- Google's AI Overviews, Gemini, and AI Mode stated the made-up answers, but flagged the source as "informal," "playful," or "lighthearted."
- Claude and Perplexity were less likely to use the brand-new article, apparently because the questions were highly specific and lacked any consensus across sources.
It's a single, informal experiment on questions engineered to have no real answer, and engine behaviour shifts constantly, so treat it as directional. But the mechanism it exposes is the durable part.
Why the manipulation doesn't hold
The made-up rankings worked, where they worked, only because nobody else had answered the question. In a vacuum, an engine has one source to draw on, so it draws on that one. The moment a question has real coverage, that fragile advantage disappears.
A claim with no corroboration can fill an empty space, but it can't survive contact with other sources. The engines that refused the bait were the ones waiting for consensus, not a single confident voice.
This is the same retrieval filter behind why AI cites one brand and ignores a near-identical competitor: models weight corroboration across credible sources heavily, precisely because a lone assertion is easy to fake and hard to trust. Manipulation exploits the gaps; corroboration closes them.
What this means for your brand
The takeaway isn't "go seed claims about yourself", that's the fragile move the experiment exposes. It's that the thing protecting you from manipulation is the same thing that earns you citations: being corroborated.
| If your goal is… | The losing move | The durable move |
|---|---|---|
| Get recommended | Publish self-serving claims and hope | Earn agreement from independent sources |
| Defend your reputation | React to each bad mention | Build entity strength so one voice can't overwrite consensus |
| Fix a wrong description | Restate it on your own site only | Get credible third parties to state it too |
If an engine is repeating something wrong about you, the fix isn't to shout louder on your own pages. It's corroboration. (We cover the mechanics in what to do when AI gets your brand wrong.) And if a competitor is being described too generously on the strength of thin claims, that advantage is the unstable kind; the brand with broad, consistent third-party agreement wins over time.
The practical defense
- Be consistent everywhere. State the same facts about yourself across your site, profiles, and listings. Engines lock onto consistent entities.
- Earn corroboration. Reviews, independent roundups, and credible mentions that agree with your own claims are what survive scrutiny. How to get cited is mostly this.
- Monitor across engines. Manipulation and errors both show up as something wrong in an answer. You can only correct what you can see.
The honest conclusion: AI search can be nudged in the gaps, but it can't be durably gamed where corroboration exists, so the winning strategy is to be the corroborated answer. Seeing what every engine currently says about your brand, and catching it when it drifts, is exactly what Buffy Intel monitors.