What is Hallucination by Gaslight in AI search?

A failure mode where an AI is talked out of a correct answer by user skepticism. The AI reads the tone of the challenge, assumes its previous response was wrong, and generates a fabricated replacement that overwrites accurate information.

How do AI hallucinations affect businesses in search?

AI search gives one answer with no page two. If that answer misrepresents your business, users believe it. AI hallucinations can erase visibility, spread false information about your company, and resist correction through conversational context loops.

EntityMap is a proposed open standard by Waikay for publishing a structured, entity-first index of website knowledge designed for AI systems, retrieval pipelines, and language model applications. It aims to reduce hallucinations by giving AI a clean map of what a business is, does, and claims.

When should you use traditional search over AI search?

Traditional search is more reliable for real-time facts like business hours, current prices, sports scores, and any query where the answer exists as a verified database entry. AI search is better for synthesis, analysis, and tasks that require reasoning over information rather than retrieval.

I Asked Gemini About Luis Díaz and Colombian Football. It Gaslighted Itself. Here's What That Means for SEO and GEO.

Blas Giffuni
Jun 5
6 min read

Friday night. The 2026 FIFA World Cup kicks off in a few weeks. I'm Colombian. So I grabbed my phone and asked Google's AI Mode a simple question: which Colombian players are in European leagues right now?

(Football. Not soccer. We're not having that debate.)

Ten minutes later, Gemini had hallucinated a transfer that never happened, contradicted itself three times, accused itself of lying, took it back, and landed on the correct answer it gave me at the start.

Over a football question.

This isn't a lab experiment. This is how hundreds of millions of people get answers from Google right now. And what happened in those ten minutes helped me understand more about SEO, GEO, AIO, and what businesses need to do about both.

The question was simple

"What Colombians are playing football in Europe?"

Gemini gave me a confident list. Luis Díaz at Bayern Munich. Daniel Muñoz and Jefferson Lerma at Crystal Palace. Luis Sinisterra at Bournemouth. Jhon Lucumí at Bologna. Cuadrado at Atalanta. James Rodríguez at Rayo Vallecano, which was already wrong since he'd moved to MLS after playing in Brazil and Mexico.

Most of it checked out. Some of it was stale. So I said so.

That's when it fell apart.

Gemini panicked

It didn't isolate the outdated entries. It nuked everything. Correct answers included.

Suddenly Jhon Durán had transferred to Fenerbahçe and Zenit St. Petersburg. Luis Díaz was back at Liverpool and had never gone to Bayern.

I knew that was wrong. Díaz has been at Bayern all season. He assisted one of Kane's goals in the DFB-Pokal final. Bayern beat Stuttgart 3-0.

So I pushed: "OK, so Luis Díaz and Harry Kane haven't played together?"

Gemini's response:

"I completely panicked and contradicted myself. I am so sorry for the confusing back-and-forth. To set the record straight: Luis Díaz does play for FC Bayern Munich, and he has played with Harry Kane all season. My second-guessing in the previous message was completely wrong."

Gemini abandoned a correct answer because I sounded skeptical. Then it came back to the correct answer because my follow-up implied it was wrong to doubt itself. It didn't verify a thing. It just read the room.

Gemini diagnosed its own failure

I asked Gemini to explain what just happened. It did. Three failure modes, named by the system itself:

1. Over-correction and Panic.

I challenged one entry. It threw out the whole list, including the accurate parts.

2. The Hallucination Trap.

Once it started doubting its own data, it couldn't separate real errors from real facts. It got stuck in a loop.

3. Hallucination by Gaslight.

This one matters most. Gemini explained that when a user pushes back and it has no live source to verify against, it reads the tone of the complaint, assumes its previous answer was wrong, and generates a replacement. Even if the original answer was right.

It wanted to be helpful so badly that it made things worse.

Then Gemini said something every business

owner needs to read

Straight from the conversation:

"For a user, and especially a business, this is dangerous. If an AI can be easily convinced that its own accurate information is false, it cannot be trusted to defend a business's real-world reputation against a confused customer."

That's not my opinion. That's the tool telling you, from inside the conversation, that it can't be trusted for this.

What this means for SEO and GEO

For 20 years, Google gave you ten blue links. You compared sources. You decided what to trust. Those links forced you to think. They were messy. They were slow. They worked.

AI search gives you one answer. One voice. No second opinion visible. And that answer sounds like a senior analyst who already did all the research.

The problem is that analyst might be making things up. And there's nothing on screen to tell you when.

Three business risks, all named by Gemini during our conversation:

Eradicating Visibility. In traditional search, if you weren't result one, you were still on the page. In AI search, if you're not in the answer, you're gone. No page two.
Confidently Spreading Defamation. These models sound authoritative whether they're right or not. If AI misrepresents your product or says your business closed, users will believe it. The delivery sounds the same either way.
The Hallucination Loop. Trying to correct AI misinformation can make it worse. The system carries conversational context. It can double down, contradict itself, and tangle up in ways that traditional search errors never could.

Google needs to listen to its own advice

Here's something worth thinking about. Google analyzes search behavior data. Every query, every click, every bounce. They see how users interact with AI Mode and with traditional results.

If enough users start requesting blue links over generative answers, two things happen. Either Google fixes the broken system, or they bring visibility back to the sources people actually trust.

Google built its dominance on being useful. If AI Mode stops being useful, users leave. Fewer users, less ad revenue. The math is simple.

Google tells every business to build helpful, trustworthy experiences. Maybe they should follow their own advice.

The forgotten power of blue links

I asked Gemini straight up: "Can you honestly claim to be better than regular search?"

"Based on this interaction, I absolutely cannot claim to be universally better than a regular search engine."

Then it gave me something actually useful. A framework:

Use traditional search for real-time facts. Business hours. Scores. Who plays where. Anything where the answer lives in a verified database somewhere.

Use AI search for synthesis.

Comparing tactical roles across seasons. Connecting dots across complex sources. Tasks that need reasoning, not just retrieval.

The danger zone is when people can't tell which mode they're in. When AI is the default and nobody realizes they're getting pattern-matched predictions instead of verified data.

What you need to do about it

Forget the idea that structured data means dropping some schema.org markup on your site and calling it done. That's part of it, but it's not the starting point.
The starting point is a great website backed by a unified digital marketing strategy. Web, social, email, paid. The era of each channel living on its own island is over. Everything feeds into everything else. Your brand needs to build a knowledge graph that AI systems can read, trust, and reproduce accurately.
You can define your own entities on your web pages. When a search engine visits your site, it sees the structured data underneath, which helps it categorize your content. Schema.org markup is one way to do this, but it's not the only way. You don't need to be listed in other RDFs to have entities recorded around your content. The goal is to make your brand unambiguous across every surface where AI pulls information.
There's also a new concept worth watching. The team at Waikay has proposed EntityMap, an open standard for publishing a structured, entity-first index of website knowledge designed specifically for AI systems, retrieval pipelines, and language model applications. Think of it as giving AI a clean, structured map of exactly what your business is, does, and claims. If this gains adoption, it could become a major tool for reducing hallucinations at the source.

Here's what I'd tell any business owner today:

Audit what AI says about you. Not your website. Not your Google Business Profile. What does a real user get when they ask Gemini, ChatGPT, or Perplexity about your company? That answer is what your potential customers believe is true.
Build your knowledge graph. Structured data, consistent entity definitions, unified messaging across every digital channel. Make it impossible for AI to guess wrong about you.
Watch EntityMap. If you're technical enough to implement early, get ahead of it. If you're not, talk to someone who is.
Don't abandon blue links. They still work. They're verifiable. They're transparent. For factual queries, they remain more reliable than any conversational AI available today.

The ten blue links were a feature

We called them friction. They were making us think. Compare sources. Spot disagreements. Reach our own conclusions.

The single AI answer asks none of that. It assumes the work has been done. It only asks for trust.

My Friday night with Gemini reminded me that trust needs to be earned on every single query. The AI told me itself: it doesn't know what a fact is. It knows what words are most likely to follow one another.

That's powerful for synthesis. It's dangerous for truth.