By Paragon AI MAG — Series 4, Episode 5: "How the AI Tools You Use Actually Work"
In May 2024, Google placed an AI answer box at the top of its search engine. Within days, it advised users to mix about an eighth of a cup of non-toxic glue into pizza sauce to keep the cheese from sliding off, and suggested eating at least one small rock a day for digestive health. The system was not being malicious. It simply could not tell an 11-year-old satirical Reddit comment or an article from The Onion from verified reality.
Two years later, Google's AI Overviews reach over 2.5 billion people a month. And search itself has flipped upside down: between January and April 2026, 68.01% of Google searches in the United States ended without a single click to the open web. More than two in three searches, and nobody clicked anything.
The classic bargain of the internet — you type a query, the search engine shows ten blue links, you click a publisher's site and read the page — has collapsed into an answer box that reads the web for you and summarizes the result on the search page.
Our verdict: 🟡 WATCHING
AI search is fast, it works, and it reads live web pages before it answers — a powerful first pass. But the two biggest independent tests found about half of its answers have a significant problem, delivered in the exact same confident tone as the truth. Use it to start. Never use it to decide.
The scale: a trillion-query machine
Google handles more than 5 trillion searches a year — about 13.7 billion searches every single day, roughly 158,000 every single second. Google's AI Overviews now reach over 2.5 billion people a month, automatically, whether they asked for it or not.
The competitors run at massive scale too. ChatGPT takes 2.5 billion prompts a day, with over 900 million weekly users as of early 2026. Perplexity answered 780 million queries in May 2025 alone — roughly 30 million a day in June 2025 — up from roughly 3,000 queries on its launch day in 2022.
The five-step machine
Classic search matched keywords and ranked pages by links. Modern AI search does not guess answers from its training memory. It runs a five-step machine:
1. Crawl and index. A page can only be cited if the search engine has already catalogued it. There is no secret backdoor.
2. Query fan-out. The engine splits your one question into many simultaneous searches. "How to fix a lawn full of weeds" fans out into best herbicides, removing weeds without chemicals, and preventing weeds — all at the same time.
3. Vector embeddings. Words and phrases are plotted as numbers in a high-dimensional space, where meanings sit near each other — "dog" near "puppy", far from "car". The system understands context, not just keywords.
4. Retrieval-augmented generation (RAG). The retrieval engine pulls fresh, ranked pages from the index and feeds them into the language model's context window. Without RAG, an answer stays frozen at the model's training cutoff date.
5. Synthesis, ranking and citation. The model writes one answer from the retrieved pages, with citations. In one large test, ChatGPT Search averaged 10.42 links per answer, Google AI Overviews 9.26, Perplexity 5.01, and Microsoft Copilot 3.13. Fewer sources means a narrower pool — and more blind spots.
The big players
All the major engines use the same five steps, but they run them differently:
Google AI Overviews appear automatically at the top of standard results — which is why they reach 2.5 billion people. Google AI Mode is a separate conversational tab, powered by Gemini 2.5, built for long, multi-part questions.
ChatGPT Search, rolled out in October 2024, decides for itself when a question needs a live web search and when its own memory is enough.
Perplexity searches the open web on every single query and shows its sources prominently — one study found it cites about 20 sources per answer.
Bing Copilot is built directly into Windows, the taskbar and the Edge browser. YouTube Ask lets viewers ask questions and jump straight to the most relevant moment in a video.
When AI search fails
The central flaw is that the machine cannot express doubt. Two large independent audits measured it:
The Columbia Tow Center study (March 2025) tested eight generative engines across 1,600 news queries. More than 60% of answers were incorrect. Perplexity was the best, wrong 37% of the time. Grok 3 was the worst, wrong 94% of the time. ChatGPT misidentified 134 out of 200 articles, signalled doubt only 15 times, and never declined to answer. Over half of the citations in Gemini and Grok 3 led to fabricated or broken URLs. And premium paid models gave more confidently incorrect answers than free ones — because they refused to say "I don't know".
The EBU study (October 2025) had 22 public service broadcasters across 18 countries evaluate more than 3,000 AI answers. 45% of answers had significant issues, 31% had serious sourcing problems, and 20% contained major factual errors. Gemini had significant issues in 76% of its responses.
The failures fall into three buckets: the confidence problem (wrong in the same tone as right), broken citations, and fresh-news failures — the AI confidently named a deceased Pope Francis as the sitting pontiff, and falsely claimed surrogacy is illegal in the Czech Republic. Fiction, delivered with absolute certainty.
Who loses when nobody clicks
Pew Research Center found that searchers who saw an AI summary clicked an organic web link in only 8% of visits, compared to 15% on pages without one. Users clicked a link inside the AI summary itself in just 1% of visits.
Seer Interactive tracked a 61% drop in organic click-through rates on queries where an AI Overview appeared (from 1.76% down to 0.61%). Queries without an AI Overview also fell 41%, so AI is not the only factor — but the direction is clear. There is one twist: brands explicitly cited inside an AI Overview saw their click-through rates rise from 0.74% to 1.02%. Being cited inside the box is the new currency. If you are not in the summary, you practically do not exist.
Your practical toolkit
Ask full questions. Queries of 10 or more words trigger an AI response 53% of the time, compared to 8% for short keyword searches — and longer questions get better answers. Add the words "with sources" at the end of your prompt to force a live web search.
The 60-second citation check. Always click at least one citation. Check that the source page actually says what the AI claims it says. If a claim has no citation, treat it as a pure guess.
Divide by query type. Use AI search for exploration, synthesis and first overviews. Use classic link search for buying decisions, medical advice, fresh breaking news and local business lookups.
Turn the AI box off. Google offers no permanent account toggle. Per search, tap the "Web" tab below the search box for clean, unsummarized links. To make it permanent, set up a custom search engine in your browser pointing to https://www.google.com/search?q=%s&udm=14 — the &udm=14 parameter forces classic web results with no AI summary.
The bottom line
AI search is genuinely useful. It reads live pages before answering, it is far faster than scanning ten blue links, and 2.5 billion people use it every month for good reason. But because the two biggest independent tests found about half of its answers carry a significant problem — and because the machine is confidently wrong and never admits ignorance — you have to change how you use it.
Let the AI do the first pass. Make the final decision yourself. And always, always check at least one citation.
When did an AI answer last get something wrong for you — and did you even check? Tell us in the comments.
Paragon AI MAG is a daily AI magazine: one video and one podcast episode every day, every fact checked against sources several times before release. Watch the video version on the Paragon AI MAG YouTube channel and listen to the full podcast episode.
For the engineers: this breakdown dissects the AI search pipeline — crawl gating, concurrent query fan-out, vector embeddings, RAG injection into the context window, and citation ranking. Plus the benchmark numbers: ChatGPT Search averaged 10.42 links per answer, Google AI Overviews 9.26, Perplexity 5.01, Copilot 3.13 — and why fewer sources means more blind spots. Follow Paragon AI MAG on Dev.to for daily source-checked explainers. Full video: https://youtu.be/QyStmo6Apk0

