OpenAI Built a Weather Station and Nobody's Doing the Forecast

Recipe requests to Claude spike 2.3x at 6 p.m.

Personal conversations jump from about 35% of Claude’s traffic on weekdays to just under 50% on weekends. Tax questions run eight times their normal volume around filing season. Users who’ve been around for six-plus months have a 10% higher success rate than newbies, who are mostly asking for haikus, sports scores, and what to make for dinner.

I know all of this because Anthropic published it. In a report. For economists.

And that, in one sentence, is both the problem and the opportunity with the most valuable audience data ever assembled.

I. They Actually Built The Weather Station

A few months ago I wrote that the whole trick to a brand newsroom is becoming the weather station in an industry that’s been squinting at clouds for fifty years. At Freightos, the weather was ocean freight prices. It took us years to publish them consistently, mostly because I kept losing internal arguments about whether giving away data was “strategic.” (It was. I was wrong for about three years. I have the Slack threads.)

So let me give credit where it’s due, because the foundation labs have done the hard part.

OpenAI Signals launched in early 2026 as, in its own words, “a public data initiative that provides regularly updated, aggregated insights into how ChatGPT is being used around the world.” It has an interactive dashboard, a downloadable de-identified dataset under a CC BY 4.0 license, quarterly research briefs, a B2B extension, and a global report with a foreword that reads like it was written by a very earnest central bank. In August it dropped its first country-by-country release across 147 countries. Singapore sends the most ChatGPT messages per capita. Multimedia is 7.8% of messages globally but over 10% in Brazil and Colombia. At work, people are more than twice as likely to ask ChatGPT to do something than to explain something.

Anthropic’s Economic Index has been at it since early 2025 and is, if anything, nerdier. Three reports so far this year. January introduced “economic primitives,” five dimensions scored on every conversation: complexity, skills, use case, autonomy, success. March, titled “Learning curves,” tracked how users change as they gain tenure. June, “Cadences,” mapped the hour-by-hour and day-by-day rhythm of millions of conversations, then bolted on a survey of roughly 9,700 people. The whole thing runs on a privacy-preserving system called Clio, where the only entity reading your transcript is another instance of Claude. Patterns, not people.

If a Series B SaaS company published a fraction of this, I’d write a teardown and call them geniuses. (I have. Ramp. Expense reports.)

So why does it feel like a missed layup?

II. A Weather Station Is Not A Forecast

Because a weather station measures. A forecast tells you whether to bring an umbrella.

Both programs are explicitly built for the same audience. OpenAI’s stated purpose is to help “researchers and policymakers.” Anthropic’s is to make sure “researchers and policymakers have adequate time to prepare.” The bylines are economists. The cadence is quarterly. The distribution is a PDF, a dashboard, and a Creative Commons license. The prose is written the way you’d write if you knew the Fed might read it. Which, to be fair, it might.

Nobody is writing for the person who asked about dinner at 6 p.m.

That’s the gap. The instruments are running beautifully. Nobody’s standing in front of the map at six o’clock pointing at the cold front.

Here’s what makes this weirder than a normal “company sits on data” story. Both labs already have the data, already cleared privacy, already built the classification pipeline, already got legal to say yes. The marginal cost of the newsroom is a handful of journalists and a publishing cadence. Every other B2B company I’ve helped build a data program for had to fight for eighteen months just to reach the spot these two are already standing on.

III. Why This Matters More Than The Next Model

Let me take a step back, because “OpenAI should blog more” is not the thesis.

Models are commoditizing. Not “will.” Are. The gap between frontier models on the stuff normal people actually do narrows every quarter, the open-weight models keep catching up, and the wrapper era (the app on top of the model that actually owns the relationship with the user) is already here. If you’re a lab, your moat used to be the model. Increasingly, it’s the audience that trusts you enough to keep showing up. Distribution, not weights.

And trust, in 2026, is the scarce asset. Everyone I know has a “which AI is lying to me less” heuristic roughly as rigorous as picking wine by the label. Into that vacuum, the company that becomes the credible narrator of how humans are actually using AI wins something no benchmark buys. They become the source. The Reuters citation. The stat in the board deck. The chart your CFO screenshots and sends to the group chat.

Freightos did this with freight rates. Ramp did it with corporate spend. Redfin did it with housing. Every one of them had a worse dataset than a foundation lab and a smaller audience than a mid-sized subreddit.

OpenAI and Anthropic are sitting on hundreds of millions of humans, every day, showing exactly how people are learning to think with a new kind of tool. What they try first. What they give up on. What they come back to at 11 p.m. on a Tuesday. It is the best behavioral dataset in the history of marketing, and it is currently being used to (a) inform workforce policy and (b) sell API credits.

Both worthy. Neither is a brand.

IV. Six Stories Already Sitting In Their Own Data

Here’s what makes this not-hypothetical: most of these stories are half-written. The economists already ran the query. Someone just needs to turn the table into a headline.

1. The Age Curve. Signals already reports users over 35 gaining message share every quarter, and users with typically feminine names crossing 50%. That’s a “who’s actually using AI” story every newspaper on earth would cite. Right now it’s a chart in a Q1 update.

2. The Power User Cliff. Anthropic’s March report literally found it. Six-plus-month users have a 10% higher success rate, tackle higher-education and higher-wage tasks, and iterate instead of dictating. New users write haikus and check sports scores. Veterans revise manuscripts and raise funding rounds. That’s not a finding. That’s a series.

3. The 6 P.M. Query. “Cadences” has the hourly curves: recipes at six, news and sleep advice at their own hours, personal use swelling to half of everything on weekends. It’s the shift from work tool to life tool, with receipts. Publish it weekly. Call it the forecast.

4. The Sophistication Stack. What share of users have ever touched memory? Projects? Voice? Custom GPTs? Nobody outside the labs knows. The gap between “I use ChatGPT” and “I use ChatGPT” is enormous and completely unreported.

5. The Prompt Graveyard. Where people give up. The single edit that turns a failed conversation into a successful one. Anthropic already scores task success on every conversation. The autopsy data exists.

6. The Asking-to-Doing Line. OpenAI’s own taxonomy, asking vs. doing vs. expressing, is the story of AI becoming an agent, told country by country. Singapore’s at the front. Who’s next? That’s a monthly leaderboard, not an August PDF.

Each of these is a Reuters citation. Each one is impossible to replicate without the data. Each one is a chart a competitor can’t paste on their own site without flinching.

And every one of them already has the hard part done.

V. The Feature I’d Build Next To It

If I ran the newsroom at either company, the distribution play wouldn’t be a newsletter. (It would also be a newsletter. I’m not an animal.)

It would be a live, personal progression map. Not annual. Not Spotify Wrapped. Real time.

You log in and see: here’s what you’ve done with AI. Here’s where 68% of users like you go next. Here’s the one tiny thing that levels you up. Save a prompt. Try memory. Connect your calendar. Build a project.

Gamified onboarding that never ends. It teaches you what’s possible while telling you exactly where you stand. The “Learning curves” report is, functionally, the product brief for this feature. It already knows what six-month users do that one-month users don’t.

The newsroom data feeds the map. The map generates the next article (“This week two million people crossed the Power User Cliff. Here’s what they did first.”). The article drives people back to the map.

It requires zero new technology and one uncomfortable decision: that owning the audience matters as much as shipping the next model.

VI. The Monday-Morning Playbook

Because I’m a complainer (it’s a gift), here’s what I’d do if someone handed me either program on Monday:

1. Hire two journalists, not another economist. The economists are excellent. Keep them. Put a former trade reporter and a data-viz editor next to them. The economist’s job is to be right. The journalist’s job is to be read.

2. Go weekly. Quarterly is a report. Weekly is a habit. The freight index that ended up on Bloomberg started as a weekly email I hacked together from a Google Sheet. Cadence is the credential.

3. Put a face on it. Right now the byline is an institution. Institutions don’t get DMs. A named human on the “how people use AI” beat, posting on LinkedIn three times a week, will out-distribute the PDF within a month.

4. One story, one chart, one sentence. Every release should have a single number a reporter can lift without reading page 14. “Recipe requests spike 2.3x at 6 p.m.” is a headline. “Temporal heterogeneity in task distribution” is a footnote.

5. Make it embeddable, keep it ungated. OpenAI already did the brave part with CC BY 4.0. Now build the chart embeds and the press kit so that giving it away actually spreads. (The Hard Signal Protocol covers how to pick which numbers deserve the treatment.)

6. Let the consumer story sell the enterprise. “How your industry is actually using AI” is a better enterprise lead magnet than any whitepaper ever written. B2B Signals already exists. Point it at buyers, not just researchers.

Here’s the uncomfortable kicker, the one that would keep me up at night if I worked there. The Signals data is Creative Commons. Which means a wrapper company with one good writer could run OpenAI’s newsroom for them, on OpenAI’s own data, and collect the audience.

Nature abhors an unstaffed weather station.

The Part Where I Admit Things

Two disclosures. First, I’m a paying customer of both companies and I drafted chunks of this inside Claude Code, which means a toaster helped me write an essay about how its manufacturer should market it. We live in interesting times. Second, I spent years at Freightos sitting on data I should have published sooner, so I recognize this failure mode from the inside. It always feels like the data is the asset. It isn’t. The habit of showing up with it is.

The labs have built the most sophisticated weather station in the history of the industry. The barometers are exquisite. The privacy engineering is genuinely admirable. The economists are doing real work.

It’s 6 p.m. Somewhere, a few hundred thousand people are asking what to make for dinner.

Somebody should be doing the forecast.

(It’s chicken. It’s always chicken.)

This post started as a LinkedIn post. If you want more of this in your feed, follow me on LinkedIn.

Frequently Asked Questions