Published 28 May 2026
\nGoogle’s upgraded AI Overview is confidently telling users that there are two “P”s in the word “Google.” The error occurs because large language models process text using tokens—chunks of characters—rather than individual letters, making character-level counting an unnatural task for the architecture.
\n\nThe incident is amusing, but it highlights a persistent weakness in modern transformer models: the inability to reliably inspect their own outputs at the character level. As Google pushes AI-generated answers to the top of the search results page, these fundamental machine-learning blind spots are getting high-profile placement.
\n\nThe Actual News: AI Overview's Spelling Errors
\n\nFollowing a recent upgrade to make its search engine more conversational, Google has placed AI Overview front and center in response to user queries. TechCrunch first noticed that the system severely struggles with basic spelling questions. A simple search for “How many Ps are in Google” triggers an inaccurate response stating that there are two 'P's.
\n\nThe AI doesn't stop at simply giving the wrong number. It confidently follows up the error by linking the query to the mathematical term “googol,” completely failing to assess the literal characters in the word.
\n\n \n \n \n\nWhy does Google AI Overview say there are two Ps in Google?
\n\nWhen users search 'How many Ps are in Google,' the upgraded AI Overview sometimes responds that there are two 'P's. This happens because transformer models read words as whole tokens, not sequential letters.
\n\nPC Gamer confirmed the error. Asking the AI how many “R”s are in the word “enigmatic” results in the system confidently answering “one.” Bizarrely, it then correctly spells out the word using individual letters (e, n, i, g, m, a, t, i, c) to prove its point—proving that, in fact, there are zero Rs. The error is easily replicable across browsers like Chrome and Firefox.
\n\n
The Hidden Variable: The Tokenization Problem
\n\nThe consensus around these errors is usually just “AI is dumb.” That is a lazy read. The actual failure mechanism is the tokenizer.
\n\nWords and letters are represented by tokens within transformer-based models. The AI doesn't read the word “Google” the way a human does (G-o-o-g-l-e). Instead, it converts the text into numerical representations. Depending on the tokenizer's vocabulary, the entire word might be a single token. Asking an LLM to count letters within a word is like asking a human to count the individual brushstrokes in a painting by looking at the finished canvas. The underlying granularity simply does not exist in its default field of view.
\n\n \n \n\n
Why This Matters for Search and SEO
\n\nGoogle is positioning AI Overview as the definitive, zero-click answer engine for the modern web. When the system hallucinates facts about history or science, it is a liability. When it fails to count letters in its own name, it is a blow to user trust. For publishers and SEO practitioners, this illustrates a critical reality: the AI deciding your traffic fate is not a reasoning entity. It is a pattern-matching engine wrestling with fundamental architectural constraints.
\n\nThere is a valid argument that users shouldn't be asking search engines to count letters. But if Google insists on serving a conversational AI for all queries, the system must be robust enough to handle edge cases without hallucinating.
\n\n \nThe AI's confident tone is the real issue here. It doesn't say "I think there are two Ps." It states it as a mathematical fact. (This is a fundamental flaw in alignment: models are trained to sound helpful regardless of their actual competence.)
\n\nWhat is an LLM token?
\n\nA token is a numerical representation of a sequence of characters. Common words like 'Google' might be a single token, meaning the AI recognizes it as one mathematical unit rather than six distinct letters. The tokenizer's job is to break text into these chunks, but the system loses character-level awareness in the process.
\n\n \nMake no mistake: Google is clearly prioritizing deployment speed over accuracy.
That calculation may make sense on a spreadsheet, but it risks eroding the core product promise.
What is Still Unknown
\n\nIt is unclear if Google is actively patching these specific character-counting edge cases or if the underlying model architecture is being adjusted. These errors tend to come and go, suggesting Google might be implementing superficial filters rather than solving the root tokenization issue. As PC Gamer noted, the error appeared consistently in Firefox but had already disappeared in Chrome during later testing, implying aggressive, browser-specific A/B testing or rapid patching.
\n\n \nInitially, it seemed this was a regression caused solely by the new AI Overview upgrade. However, looking at the persistence of tokenization errors across the industry, it is more accurate to say the upgrade simply gave an existing, unsolved LLM limitation a much brighter spotlight.
\n\n
What Players Should Watch Next
\n\n- \n
- Patching Behavior: Watch whether Google addresses this by improving the model's reasoning capabilities or simply hardcoding blocklists for spelling questions. (The latter would be a band-aid). \n
- Competitor Responses: See if OpenAI or Microsoft alter their consumer-facing search tools to preemptively mock or solve this specific issue. \n
- User Backlash: Monitor if these frequent, low-stakes errors eventually impact overall trust in Google Search, or if users simply learn to ignore the AI Overview box. \n
FAQ
\n\nWhy does Google AI Overview say there are two Ps in Google?
\nGoogle AI Overview relies on a large language model (LLM) that processes text as tokens rather than individual characters. Because it does not 'read' words sequentially letter-by-letter, it often fails basic character counting and spelling tasks.
\n\nWhat is an LLM token?
\nA token is a numerical representation of a sequence of characters. Common words like 'Google' might be a single token, meaning the AI recognizes it as one mathematical unit rather than six distinct letters.
\n\n
