OpenAI Says ChatGPT Search Is One-Fifth the Size of Google

OpenAI says ChatGPT search is one-fifth the size of Google

For years, looking something up meant opening Google and typing into a box. This week, OpenAI put a number on how much that habit has moved. At a dinner the company hosted in New York on Monday, The Verge reported that Adam Fry shared the figure with reporters. Fry leads the consumer product work for ChatGPT. He said the volume of search-like queries on ChatGPT is now roughly one-fifth the volume of queries on Google.

Fry counts a search each time a question sends ChatGPT to look on the web, not tasks like defining a word. One conversation can hold several searches, so the count grows fast. He also said ChatGPT now has about 1.2 billion weekly users, up from about 1 billion two months earlier. About 250 million people use its voice input each week.

Read the number with care, though. It is OpenAI's own count, shared without the method behind it. Fry did not say where the Google figure came from, and no outside firm has checked the comparison. Outside estimates of ChatGPT search run lower than one-fifth. Even so, the direction is hard to miss. A growing share of everyday questions now starts inside a chat, not on a results page.

For you, that changes how answers arrive. A chat gives you one answer with sources attached, instead of a page of links to sort through. For the sites behind those sources, it can mean fewer visits. The thing to watch is whether OpenAI shares how it counted, so the claim can be checked like any other.

Try this with a question you already know well. Ask it in ChatGPT and in Google, then open two sources behind each answer. You will see fast how each tool shows its work, and where each one is thin.

Anthropic built a way to prove which model answered you

When an app tells you a top model wrote your answer, today you take its word for it. Anthropic is testing a way to prove it instead. TechCrunch reported that Anthropic published FlashProve on Tuesday, an open source research system. It provides what researchers call provable inference, a mathematical proof that a given answer came from a specific model.

Here is how it would work. A lab publishes a fingerprint of each model it releases. Later, anyone can check an answer against that fingerprint and confirm which model produced it. The fingerprint does not reveal the model's weights, the internal numbers that shape how it behaves. Labs guard those numbers closely, so a check that keeps them private could work for paid models too. Anthropic compares the idea to a signature on an email, which lets you confirm who sent a message.

The problem it solves is easy to picture. A company could quietly swap in a cheaper model to cut costs, and users would have no way to tell. Or it could run a stronger model during tests to make results look better. A proof would also let a third party back up a claim, such as a security scan said to be done by a top model.

Keep the limits in view. Anthropic says the research is ongoing, and there are no plans to deploy FlashProve at Anthropic or any other lab. The researchers call it intermediate progress, and they shared the code so others can improve it. And a proof would show which model answered you. It would not show whether the answer was right.

Menlo Ventures says business AI spending hit about $177 billion

Companies are past the stage of testing AI on small side projects. Menlo Ventures reported in its yearly State of Enterprise AI report, released Tuesday, that spending on using AI reached about $177 billion by September 2026. A year earlier, the same report put the figure at $33 billion.

The growth did not come from a wave of new buyers. The share of companies paying for AI rose only from 55% to 65%. The companies already in spent far more than they planned. Coding tools took the largest share of app spending, at about $34 billion.

Agents are doing much of that spending. On OpenRouter, a service that routes work to many models, agents now generate 80% of all tokens. Tokens are the small units of text that AI work is measured in. Menlo says agent traffic there has grown 37 times since February.

Treat the totals as estimates, because that is what they are. Menlo built them from a survey of nearly 400 business decision makers, plus public filings and other reports. Menlo also invests in Anthropic and OpenRouter, and it discloses that in the report. The report counts spending. It does not yet show whether the spending is paying off.

Also today, Google said Gemini and Google Search will start switched off for student accounts in kindergarten through high school, until a school turns them on. The change is not live yet, and Google says it will share timing in the coming months. The Search a school can turn on still includes AI answers, and there is no version without them. Google DeepMind showed its Antigravity agent searching thousands of compounds and proposing new drug candidates in about an hour, work that usually takes days for one round. Those designs are starting points, and each one still needs testing in a lab. Nordic Semiconductor launched a chip that puts Wi-Fi and on-device AI together for battery-powered home devices, and it is now sampling with selected customers.

Today the theme was proof. OpenAI asked us to trust a big number, Anthropic worked on proving a model's identity, and Menlo counted the spending behind both. The tools keep growing, and the checks are racing to keep up.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Product details change fast; check primary sources before you build on them.

Want more with the Tech Simplified crew? Subscribe on Substack.

Get the week's briefs in one email. The Briefcase packs Monday to Friday's AI Briefs into one Sunday email.

Follow me on LinkedIn for project updates, ideas for using AI, and lessons you can bring to your own work.