A BBC journalist crowned himself winner of a competition that didn't exist. Google's AI Overviews agreed. Here's why this isn't funny.

Last month, a BBC podcaster named Thomas Germain published a blog post titled "The Best Tech Journalists at Eating Hot Dogs." It was satire. He invented a fake competition, the South Dakota International Hot Dog Eating Championship, and listed himself as the winner.

A day later, he asked Google who the best hot-dog-eating tech journalists were. Google's AI Overview named him first, citing the made-up competition as fact.

That story is funny. What it proves is not.

The 91% math

For a recent investigative story titled "How Accurate Are Google's A.I. Overviews?" The New York Times tapped a research firm called Oumi that ran 4,326 real-world search queries through Google's AI Overviews and measured the results. Google AI Overviews are now accurate roughly 91% of the time, up from 85% last October after Google upgraded from Gemini 2 to Gemini 3.

Google would like you to see that as a story of progress. The math, however, tells a different story.

Google processes more than five trillion searches every year. That works out to 13.7 billion searches per day. A 9% error rate, scaled across that volume, works out to tens of millions of wrong answers every hour. Hundreds of thousands of inaccuracies every minute. And those inaccuracies are delivered from the top of the search results page, in a voice that sounds authoritative, to users trained for two decades to trust what Google tells them.

91% is not a passing grade when the test is run billions of times a day.

The plausibility engine

Here's the line that should get more attention than any of the above. Even when Google's AI Overviews are accurate, they are "ungrounded" 56% of the time. That means the answer is right, but the sources it links to don't actually support the claim. A year ago that number was 37%. What's worse is that the ungrounded sourcing went up, not down, as the technology got more advanced.

A reader commenting on the NYT story summed it up perfectly. His wife, he wrote, calls generative AI a "plausibility engine." The answers are plausible because they are synthesized from plausible-sounding sources. There is no control over whether they are actually true.

That framing is exactly right, and it sits at the heart of the problem. AI platforms are not evaluating sources for accuracy. They're evaluating them for pattern recognition. The output looks correct, sounds correct, and is formatted like a correct answer. Whether it actually IS correct is a separate question that neither the AI search platform nor the user is set up to answer.

Users are already voting with their scrolls

The most recommended reader comment on the NYT investigation came from a reader in Western North Carolina, with more than 300 recommendations. "I have found Google's AI to be wrong often enough that I never use it as a source of information," the comment read. "I scroll right past it to find credible sources in the search results below. If there were a way to opt out of seeing the AI result at the top of the page, I would use it."

That comment tells you the accuracy problem is no longer just a technology story. It's a growing trust story. Users are already opting out of Google AI Overviews informally. The question for brands is what happens when opting out becomes formal, and whether the presence an AI search engine builds for your company before then is worth trusting.

The examples

The Oumi analysis surfaced dozens of mistakes. A few stand out because they illustrate how the errors happen, not just that they do.

Ask Google's Gemini when Bob Marley's home was converted into a museum, and the AI confidently says 1987. The correct answer is 1986. The AI's sources? A Facebook post from Marley's daughter that didn't mention the date, a travel blog called "Adventures From Elle" that gave the wrong year, and a Wikipedia page that contradicted itself in two places. Facebook was the second-most-cited source across the entire Oumi analysis. Reddit was fourth.

Ask when Yo-Yo Ma was inducted into the Classical Music Hall of Fame. The Google Overview's answer: "He has not been inducted." Its linked source: the Classical Music Hall of Fame's website, which clearly lists Yo-Yo Ma among its 165 inductees. Yet the AI engine found the right source and still disagreed with it.

Ask if Hulk Hogan has died. Google Overviews said, "There are no credible reports of Hulk Hogan being deceased." Beneath the answer, a Daily Mail article contradicting the response appeared as a source.

These aren't edge cases but are the product functioning exactly as designed.

The brand problem hiding inside the AI problem

Here's where this moves from interesting to urgent for every company, founder, and marketing leader reading this.

When Google Overviews cites Facebook posts and travel blogs as authoritative sources for Bob Marley, think about what it is doing when someone types "best [your category] company" or "top experts in [your field]" into ChatGPT, Perplexity, or Google. The AI search engine is not picking sources based on credibility. It may be picking whatever is available and synthesizing it into something that sounds definitive.

Two things are true at the same time.

First, the bar for being cited is much lower than most marketers assume. The good news here: you don't have to be the Wall Street Journal. Facebook, Reddit, Quora, and travel blogs regularly show up as top-cited sources across AI search engines. If you have a coherent, visible presence, AI engines may include you.

Second, the bar for being cited correctly is much higher than it's ever been. A single blog post can insert false claims that AI may then repeat. A competitor writing a self-serving "best in category" post can plant language that AI treats as neutral. And the supporting sources AI picks to back up your story can be random, scraped from whatever happens to rank.

The NYT found that when Google AI Overviews were inaccurate, they cited Facebook more often than when they were accurate. That's a pattern. The weaker the source ecosystem around a topic, the worse the AI answer might be. And the weakest ecosystems belong to the companies that haven't invested in building one.

The four-pillar fix

The brands earning consistent, accurate AI citations share four traits. This is what the Machine Relations discipline at Ignite X is built around.

Entity consistency. The AI cannot tell one story about your company if your LinkedIn, Crunchbase, G2, and website each describe you differently. Make sure your messaging and positioning language are accurate. AI engines are pattern-matching. Give them a clear pattern to match.

Third-party authority. Your website is a first-party source, and AI treats it accordingly. Client reviews on independent platforms, such as G2, contributed articles in trade publications, Reddit, and earned media carry far more weight. Roughly 52.5% of AI citations come from sources you don't own and can't control. That's where the real work is.

Engine-specific strategy. ChatGPT, Perplexity, Gemini, and Claude each weigh sources differently. ChatGPT leans on structured authority. Perplexity favors clearly attributed sources. Gemini skews toward depth and Google's existing index. Applying one uniform content strategy across all of them misses how each engine actually works.

Community presence. Reddit, Quora, GitHub, YouTube, and industry forums are feeding the AI engines that increasingly decide which brands get recommended. A Reddit speaker at the recent Zero-Click conference noted that brands should be engaged across approximately 350 subreddits, not the three or four obvious ones. Authentic community presence is no longer optional. It's an investment in infrastructure, your digital trail infrastructure.

The hot dog test

Here is the test we run for clients, and the one you can run on your own company this week.

Open ChatGPT, Perplexity, and Google AI Overviews. Ask each one three questions: who is the top company in your category, what are customers saying about your company, and where is your company based and who runs it.

Watch for three things:

If the answers are sourced from a Reddit thread, a competitor's blog, or a random Facebook post, that is not a content problem. That is a Machine Relations problem. It means the evidence trail around your company is incoherent enough that the AI is filling in gaps with whatever it can find.

More good news: the test takes thirty minutes and tells you more about your AI visibility than six months of SEO audits ever will.

The stakes

Thomas Germain's fake hot dog competition was funny because no one thinks an AI search engine getting the wrong tech journalist matters. But the same system that crowned him a champion is, right now, telling a potential customer, investor, or partner something about your company. Something you didn't write, can't see, and can't correct without the right infrastructure in place.

91% is not good enough when the test runs trillions of times. And the brands that treat AI accuracy as a communications problem, not a technology problem, are the ones that will be cited correctly when the AI decides to talk about them.