Google’s AI Can’t Even Spell “Google” – So Why Is It Replacing Search?

Confirmed! Google has zero “p”s.

“The word “Google” has zero “p”s. It is spelled G O O G L E. It is a famously tricky question, and even Google’s own AI has been known to get it wrong!”

This is the corrected version of Google’s AI answer, but before that, Google claimed there were two “p”s in Google.

That mistake was spotted during a week when Google was promoting a future where AI takes a bigger place in Search.

Why Is Google Changing Search?

For 25 years, Google Search worked around a search box where people typed keywords and clicked through to websites but AI has since changed how people search.

Instead of typing phrases, users now ask much longer questions. The example New York Times gave was how people can ask things such as “Who are the top 24 teams in the World Cup and what chance does the United States have of advancing?” instead of just “World Cup”

Google announced this week that it is bringing more AI features into Search. Duncan Osborn, Product Manager at Google Search, said the company is adding Preferred Sources to AI Overviews and AI Mode. Users can select websites they trust, and Google will label content from those sources in AI generated answers.

Google said people are twice as likely to click through to a Preferred Source, and that users have already selected more than 345,000 unique sources.

The company is also introducing article carousels for developing topics, discussion carousels featuring forums and social media content, and “Highly Cited” labels that identify articles referenced by other reporting.

Google said these updates help people find trusted reporting, original articles and firsthand perspectives more easily.

What Went Wrong With The Spelling Test?

This all came from a problem that sounds almost impossible where Google’s AI Overview told users there were two “p”s in the word Google.

According to TechCrunch, the same system also claimed there was exactly one “r” in the word “poop”, misspelled “journalism” as “j o u r n a d i s m”, and correctly identified that there is one “p” in the surname of the US president before spelling it “t r p u m”.

Google acknowledged the problem, saying, “Counting within words has been a known challenge for LLMs, and we’re working to fix this particular issue,” Google told TechCrunch in an emailed statement.

The incident immediately reminded many people of Google’s first AI Overview launch. TechCrunch reported that earlier versions cited satirical articles from The Onion and Reddit posts, and even advised users to eat rocks and put glue on pizza.

Google fixed many of those problems, but the spelling issue keeps haunting us all, especially the company.

How Can AI Get Something This Basic Wrong?

The explanation comes from how LLMs process text: many people assume AI reads words the way humans do. Researchers interviewed by TechCrunch said that is not how these systems work.

“LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding,” Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch. “When it sees the word ‘the,’ it has this one encoding of what ‘the’ means, but it does not know about ‘T,’ ‘H,’ ‘E.’”

Instead of reading letters one by one, AI converts text into tokens and mathematical representations.

That creates an unusual situation where a model can write software or answer difficult problems and produce long essays, but struggle with counting letters in a word.

Sheridan Feucht, a PhD student studying LLM interpretability at Northeastern University, told TechCrunch, “It’s kind of hard to get around the question of what exactly a ‘word’ should be for a language model, and even if we got human experts to agree on a perfect token vocabulary, models would probably still find it useful to ‘chunk’ things even further.”

“My guess would be that there’s no such thing as a perfect tokeniser due to this kind of fuzziness.”

Does AI Search Have A Reliability Problem?

A while ago you’ll remember when we spoke about OpenAI doing a similar thing with the famous “strawberry” test, where people asked AI models how many “r”s appear in the word. The exercise became the “ultimate” test for checking how AI systems handled spelling and counting tasks.

The fact that Google is facing the same issue shows how persistent these errors can be.

Nobody chooses a search engine because it can count letters in words, though. People use search engines because they want useful information. Mistakes involving basic spelling attract attention because anybody can spot them immediately.

Google wants AI to become the centre of Search. The New York Times reported that Google is even redesigning the search box itself so users can ask longer questions and upload photos and videos.

During the same week Google asked people to trust AI generated answers more often, users were sharing screenshots showing that the system could not always spell Google correctly.

TechCrunch wrote that these failures help people remember that AI outputs should be checked and verified before they are accepted as accurate.

Google’s latest Search updates are designed to guide users toward trusted sources and original reporting. The irony of this week’s announcement was difficult to ignore. One of the most discussed examples involved an AI answer getting the name of Google wrong before correcting itself. That does not mean the technology is useless. It does mean that even in 2026, a spelling test can expose limitations that many users assumed had already been solved.