AI in the wild: Confident, wrong, and weirdly expensive

Imagine working in SEO for years, researching a problem you know inside and out, only to have an LLM confidently explain why your experience is wrong.

That happened to me. Actually, it happened three different times last week with Gemini.

The scary part wasn’t that the answers were obviously bad. It was that they sounded good. The responses were polished, believable, and directionally accurate enough that most people would never question them.

And if you aren’t deeply familiar with the topic, you probably wouldn’t know how to challenge the answer.

Funny enough, I caught it twice. The third time cost me money. All within the same week.

Example 1: Gemini educates me on technical SEO

Let me give you some context.

I’m currently helping migrate a client’s FAQ hub from a provider-hosted subdomain to a self-hosted implementation.

The FAQ lives under a /faq/ folder, but individual Q&A pages are parameter-based URLs. Normally not an issue… except Shopify forces canonicals back to the root /faq/ page, effectively preventing those pages from being indexed.

While researching Shopify-specific solutions and duplication considerations, Gemini gave me this response:

Excuse me?!

You will absolutely not be penalized for conflicting SEO signals.

At best, Google indexes what it wants. At worst, it ignores directives it doesn’t trust.

The bigger issue is the wording.

“Penalty” is basically the magic word in SEO. The moment leadership hears it, priorities shift, momentum dies, and recommendations become harder to implement.

Then Gemini doubled down.

I asked whether removing the canonicals and allowing the parameter pages to exist independently was an option.

Gemini responded:

• “Google generally ignores query parameters.”

Wow… just wow.

Below is an implementation I worked on with the Saatva team, where we intentionally indexed parameter URLs within the shopping experience.

Search Console and URL inspection confirmed those pages indexed just fine.

Parameter pages can absolutely rank, index, and generate value.

The problem isn’t that Gemini was wrong (as in it’s a more difficult implementation).

It’s that the answer sounded believable enough that someone without SEO experience would probably accept it and move on.

Example 2: Gemini says solve the issue with a $3,000 part

This one hit differently because I’m not a mechanic.

I’ve been troubleshooting an issue on my Jeep SRT and spent hours outside in the sun collecting data, testing fuses, reviewing OBD2 logs, and trying to narrow down the root cause. After reviewing everything, Gemini confidently concluded the issue was likely a rear differential failure and suggested a full replacement.

Estimated damage? Roughly $3,000 in OEM parts alone.

The response sounded fantastic. It was detailed, logical, and even complimented my troubleshooting process.

The problem? It was wrong. Really wrong.

After pushing back and sharing additional OBD2 data I had been tracking during diagnosis, Gemini completely reversed course and admitted it jumped to a worst-case scenario without enough evidence.

In other words, I almost spent thousands replacing parts I didn’t need.

That’s what makes these examples interesting. In SEO, I immediately knew Gemini was wrong because I had years of experience to challenge the recommendation. Here, I didn’t have that advantage. I had to rely on skepticism, continue testing, and avoid treating the answer as fact.

Same AI. Same confidence. Completely different outcome.

Example 3: Gemini cost me $20 million dollars!

OK, technically, this was video game money.

I was playing Madden, working through team logistics, and trying to optimize salary cap spending to re-sign key players and keep the roster together.

At some point, I got lazy and tagged in my partner, Gemini. I shared a screenshot of the team finances and asked for a roadmap to restructure contracts and optimize the cap situation.

Gemini gave me a detailed plan. I followed it.

And it put me $20 million over the salary cap.

The funny part is this perfectly mirrors the earlier examples.

The answer looked good. It was organized, confident, and gave me a clear player-by-player action plan.

So I trusted it. Then I called Gemini out.

You’ll notice in the second screenshot that Gemini essentially points out that I blindly trusted the recommendation without validating the math.

Honestly… fair.

The difference is that this mistake only cost me fake money and a Madden franchise. The Jeep example almost cost me real money, and the SEO example could have cost implementation momentum and trust.

The value of expertise was never memorizing answers. It was knowing when something feels off, asking better questions, and recognizing when the answer smells like bullshit.

AI didn’t remove that skill. If anything, it made it more valuable.

AI isn’t replacing experts. It’s replacing people who have stopped thinking.