When Did You Last Look at What the Machines See?

Here's something I bet you've done. You open your own website, or you walk into your own shop, and everything looks right, because you're seeing it from the inside with the lights on and the coffee going.

But have you ever stood out on the sidewalk on a rainy evening and looked in through the window the way a stranger would? It can be a funny little shock. The sign you love is hiding behind a plant, and the price card fell over sometime last week, and nobody told you.

That gap between how things look to you and how they look to whoever's actually showing up is what this whole episode is about. And lately, a lot of whoever's showing up isn't a person at all.

Hi! I'm Bob Michaels, and this is Evolving the Web. Every two weeks I talk through the essays I just published, plainly, as one story. This time there are four essays, and the thread through all of them is that you should go look at the thing itself instead of trusting that it's fine.

So, back out on that sidewalk. The crawlers behind ChatGPT and Claude read the raw page your server sends, and many of them don't run the JavaScript that fills in the rest. Picture a restaurant where the menu only appears once a waiter brings it to your table. These visitors read the chalkboard by the door, and then they leave. If your price or your service text only shows up after JavaScript runs, they may never see it.

And this isn't some rare visitor wandering in once a year. A single network measured an enormous pile of requests from these crawlers in just one month, so they're out there on your sidewalk all the time.

Some systems do render the full page, like Gemini, AppleBot and Bing Copilot, but the published tests disagree on which ones actually do. I find that a little comforting, honestly, because it means you don't need to win an argument about who renders what. You just need the important stuff sitting on the chalkboard.

Here's the check I'd love for you to try. Fetch one of your pages with JavaScript switched off, and look for your price, your service description and your policy. It takes about a minute, which is less time than this coffee's been cooling.

If something's missing, the fix is to send that important text in the page itself, and then confirm it in your server log, so you've seen it with your own eyes and you're not just hoping.

Now, I tell people to check things for a living, so I have to own up to something. In September I found defects in work I'd already called finished, and my own review had passed every one of them. That stung a bit.

What I learned is that a check you've never seen fail on the problem it guards tells you nothing when it passes. Think about the smoke alarm in your hallway. If you've never pressed the test button and heard it shriek, that quiet little green light is just a light.

So before I trust a new check, I press the test button. I plant the exact defect the check exists to catch, I confirm the check names it, and then I put the original back. If the check stays quiet with a planted problem sitting right in front of it, I've learned something important before it matters.

I also stopped grading my own homework. My finished work goes to a different AI engine than the one that built it, with orders to reproduce defects, and if that engine couldn't run, the report says so plainly. That way a silent report can't pretend to be a clean one.

And once you start pressing test buttons, you notice other places where the machines are reading something you rarely look at. Your product feed is a big one.

AI shopping agents read structured product data, so your feed is the storefront they see first. It's a bit like the catalog a store used to mail out. Plenty of shoppers never set foot in the building, so if the catalog's wrong, the store might as well be wrong too.

OpenAI asks merchants for a feed that's refreshed regularly. You send it first as a sample so they can validate it, and after that you send daily updates. Google has a shared standard for agent shopping, and Shopify's catalog is the part agents use to find products. Agent checkout through it is open only to select merchants in certain countries for now.

So how do I audit a feed? The same way I audit a page. I look for every fact as plain data, a refresh that keeps up, and a match with the live site. Is the price in the feed the same price a person sees? Does the feed still list the thing you stopped selling last spring?

A fact that isn't there as plain data is a fact the agent won't state about you. It can't recommend the free returns it never read about.

All of this checking got me thinking about timing, because there's a point where looking costs almost nothing and saves the most.

Imagine you hire someone to redo your kitchen. The cheapest moment to say "wait, not that wall" is when they show you the sketch on a napkin, well before anyone picks up a hammer.

That's why I think human review pays off at the plan, before any work runs. Going back over finished work can recover a result, but it never recovers a decision.

So every plan I write has the same fixed fields. They run from a single sentence about the purpose all the way to an approval boundary, which simply says none when nothing leaves the project. Because the shape never changes, you know exactly where to look, and a person can reject a plan in a minute.

For the riskier stuff, like client copy and compliance records, a second AI model reviews the plan first, and its findings go right into the plan.

There are also gates that refuse a plan when it has unanswered lessons from earlier work. But if there's no route through one, people just switch it off, the way you'd unplug a smoke alarm that goes off every time you make toast. So every gate needs a way to pass it honestly.

And a long run only counts as done when another AI engine has reviewed it and found nothing.

So that's the whole story. Look at the page the crawler actually gets, the check you've watched fail, the feed the agent actually reads, and the plan before anyone starts work. Each time, the trick is the same small walk out to the sidewalk to see what's really there.

If you'd like the longer versions, all four essays are up on bobmichaels.ai, and I'd love to hear what you find when you switch JavaScript off.

I'm Bob Michaels. Thanks for hanging out. See you in two weeks.