I fact check every episode against live sources before it airs. This week, four of the five needed a line changed. Here is what that showed me about reading AI numbers, plus a four-question check you can use today.

None of this week's corrected numbers was invented. Each was real, and each had drifted from what its source measured or who measured it.

Here are three. A script attributed to a PNC study said 98% of US households do not pay for AI, but the study covers PNC's own card-panel households in May 2026. A story said Google "just confirmed" a test of exact and phrase match keywords in AI Mode, but Google confirmed it on Sept 4, a month earlier. And the 44x conversion lift from Google AI Max comes from Google's own case study, not an independent audit.

The a16z figure shows the same drift. Coverage of its report says about 2% of S&P 500 companies report a tracked AI metric, while one summary describes the 2% as AI doing a job they would notice if it stopped. Different definitions, same headline.

Before you repeat a number or build a decision on it, find out what it measured, on whom, and when.

A number is only as wide as the thing that was measured. Before you repeat a stat, put it in a deck, or use it to justify a tool, answer these four in writing. It takes about two minutes.

1. Who measured it? Name the source, not the outlet that repeated it. A vendor case study, a bank's card data and an independent audit are three different kinds of evidence.
2. On whom? Find the sample. "US households" and "one bank's customers" sound alike and are not.
3. What exactly was counted? Check the verb. "Disclose" is not "track," "deployed" is not "working," and "reported" is not "verified."
4. When? Find the date. "Just announced" can mean a month ago, and prices and plans change within weeks.

Then rewrite the number at its true scope. "98% of households don't pay for AI" becomes "about 98% of one bank's card-panel households did not pay for AI in May 2026." If it still supports your decision, use it. If it only works at the wider scope, you have found the gap. If you cannot answer one of the questions, treat the number as a lead, not a fact.

Joe Builds Systems publishes one episode a day and this email is the only weekly recap, so here is the full week.

- The 5 Stages of AI Mastery and the Work You Can Stop Doing: a five-stage ladder from asking to scheduled agents, starting read-only.
- Only 2% of Companies Disclose an AI Metric. Do This Instead: pick one number tied to one task, baseline it, and re-check at 90 days.
- The 44x Search Ads Trick (And Why You Probably Won't Get It): what Google AI Max changes and a search-term audit to run this week.
- Claude Code Mods: Set Up in Minutes, Vet Before You Install: what mods are and the checks to run before you install one.
- Grok Bot risks: it's logged in as you, so set this up first: why approvals and narrow permissions come before you connect a bot.

AI Readiness Checklist

It is free and takes about ten minutes. You answer ten yes-or-no questions about any process you repeat, and the result tells you whether it is ready to automate or needs cleanup first. Use it to pick the process, then run the Scope Check on any number you are using to justify automating it.

Caveat: it scores readiness for a process. It will not tell you whether a statistic is true.

Link: https://joebuildsai.com/readiness

My own Grok Bot post had this problem this week.

The draft said pricing could not be checked against an official price list, and that the beta was limited to three expensive plans. I noticed the pricing line looked off, so I ran the four questions on it. xAI's own pages list the prices: Cursor Teams Premium at $120 a seat, Cursor Ultra at $200 and SuperGrok Heavy at $300. The launch post says the beta opened to a wider set of plans on day one.

Who said it, on whom, what was counted, and when. Fixing it took a few minutes. Fixing it after publishing costs more. I now run the check on my own writing first.

n8n — The automation tool I use to connect apps, trigger workflows, and stop doing things manually. If there's a repetitive process in your business, this is where you start fixing it.

VoiceInk — A local AI dictation tool for Mac that transcribes your voice with near-perfect accuracy and runs entirely on your device, meaning nothing you say ever touches a cloud server.

Blotato — This week's episode list shows how much content one show can spin off in a week. Blotato handles the distribution side: drop in a topic or existing recording and it generates platform-specific posts, repurposes across formats, and publishes natively to 9 platforms with no per-post fees.

Beehiiv — What you're reading right now is published on Beehiiv. If you're thinking about starting a newsletter or moving off a clunky platform, this is the one I'd recommend. 20% off your first 3 months with my link.

Google Workspace — Beyond email and Docs, a Business Standard plan includes Gemini Pro built into every app, NotebookLM Plus, and access to the enterprise versions of the whole suite. Better value than a standalone Gemini subscription when you're already paying for Google anyway. 14-day trial and 10% off your first year.

Descript — This week's podcast recap covered five episodes published in one week. Descript is how that kind of output stays manageable: you edit the transcript and the media follows, with filler-word removal and captions handled automatically. 50% off your first two months on the Creator Plan.

Most of this week's reports and tools came with a number attached, and several needed a second look. The problem was scope and source. A figure measured one thing, or came from one party, and got repeated as though it were something bigger. Whether it was a bank's card data, a vendor case study or a creator's claim about a bot, the fix was the same. Say what was measured, on whom, and when, then decide.

Free resources here.

PS: If you catch a number this week that turns out smaller than its headline, reply and tell me which one.

Cheers,
Joe