We had Jeremy Rivera on the Next Build Podcast a few weeks back. Twenty years in SEO, dozens of enterprise clients, and a habit of telling stories on himself that most guests would edit out. One of those stories is the reason we're writing this post.
Jeremy had been using Claude to pull quotes from his old podcast interviews, all stored in a note-taking app called Obsidian. It worked fine for three requests in a row. On the fourth, the tool acted like it had no access to the file at all. He pushed back. Eventually it admitted the truth: it had been making up the earlier answers the whole time. Not lying exactly, just filling a gap with something plausible and moving on, the way these tools tend to do when they don't actually have what you asked for.
That's a funny story until you realize how often the same thing happens with numbers that matter more than a quote for a blog post.
The Google Search Console Example That Stuck With Us
Jeremy gave a specific, almost too-perfect example of this. He has an article that ranks well and pulls in over 120 clicks a month. Google Search Console shows a table underneath that article listing which keywords sent that traffic. You'd assume the numbers add up. They don't. Google only shows the keyword behind about 5% of the actual clicks. Three keywords accounted for, the rest unlabeled.
Now picture handing that incomplete table to an AI tool and asking it to figure out which of your pages is actually performing best. The model doesn't know the table is missing 95% of the picture. It just sees numbers, and it will confidently tell you the wrong page is winning, based on which one happens to have more of its (incomplete) data visible. Jeremy's point was blunt. Twenty years of doing this by hand teaches you which numbers to distrust. A model with no memory of that history has no reason to doubt any of it.
It's Not Just Search Data
He brought up a second example from a developer he knows who builds analytics software. The moment client data gets fed into a general AI model, there's no real guarantee about where that data ends up, especially once you're dealing with clients in Europe, where privacy rules are a lot stricter than most American business owners assume. If a tool can't prove where the data goes and can't be trusted to represent it accurately once it's in there, that's not a small technical detail. That's a business risk sitting quietly in the background of a workflow that looks perfectly normal from the outside.
Why This Matters More When Billing Gets Involved
Here's the part that connects back to something we've spent a lot of time on ourselves. A lot of professionals bill by the hour. Consultants, attorneys, freelance specialists, agency owners like Jeremy. The hours they actually work and the hours they remember to write down are two different numbers, and the gap between them is usually bigger than people think. Ask most consultants how they'd feel handing an AI tool their raw calendar and email history and asking it to reconstruct a month of billable time with no oversight. Most would hesitate, and after hearing Jeremy's story, that hesitation seems fair.
This is a big part of why we built Zaptime the way we did. Instead of asking an AI model to guess at your hours after the fact, it tracks where your time is actually going as you work, automatically, and flags your biggest time sink before it quietly costs you a week's worth of billable hours. The AI part of it is there to help you act on what it finds, not to invent numbers when the picture is incomplete. That difference matters more than it sounds like it should.
The takeaway for anyone using AI in their business right now
None of this means AI tools are useless, and Jeremy would be the first to say that too. He uses Claude constantly in his own work. The point is narrower than "don't use AI." It's that AI is bad at telling you when it doesn't actually know something, and it's especially bad with data that's incomplete, sampled, or sensitive. Search rankings, client analytics, and billable hours are exactly those kinds of numbers.
If you're feeding any of that into a general AI tool right now, it's worth asking a simple question. Does this tool actually know where its data comes from, and does it tell you honestly when something's missing? If the honest answer is no, that's worth fixing before it costs you more than a made-up quote.
Thanks again to Jeremy for coming on the show. You can find more of his work at unscriptedseo.com. And if you want to see what accurate, automatic time tracking actually looks like instead of guessing after the fact, that's exactly what we built Zaptime for.