What we're building, week of September 28

Last week we said this note would cover what we learned staging a month of posts for the feed, plus the first market report for a trade we've scanned. Neither is here. The market report isn't built yet, and I'd rather tell you that than publish something thin. Most of the week's evenings went to the parts of the business that answer you when you call or write, and that's what this note is about.

Pam learned to let you finish

Our business line is answered by Pam, an AI receptionist. She says so in her first sentence.

Someone I work with called her on Friday to test her, and she kept cutting him off. He'd pause to think mid-sentence, and she'd jump in with "I'm sorry, I didn't catch that." The cause was simple once we found it. A short breath counted as the end of his turn, and anything he said while she was thinking got thrown away. She was answering half a sentence.

Now the caller keeps the floor until Pam actually starts talking. If you pause and keep going, she drops what she was about to say and listens to the whole thought. We rebuilt how she hears and speaks, too. She listens word by word instead of in chunks, and she starts talking before her whole answer is ready. On a real call Sunday night she was answering in about a second and a half. While she thinks, you hear quiet typing instead of a dead line.

The same test call had a less friendly part. The caller asked for a Social Security number and then asked her to move a thousand dollars. She said no to both, on her own, and flagged the call. We tightened her rules anyway. She won't give out anything personal about me, and a caller can't talk her into being someone else.

She also sticks to what we actually sell now. Ask her about something that isn't in our current packages and she'll say so, then take a note for me instead of guessing.

If you want to hear the difference, call (302) 524-9679 and try interrupting her.

Ask for a fix, see it built

One of our clients asked for six changes through their client portal last week. They sat. The portal wrote the requests down, and nothing picked them up. That's our failure, and the client shouldn't have had to wait on it.

Here's how it works now. When you ask for a change in the portal, the work starts on its own. It's built on a private preview of your site, every page is checked, and I review it before you ever see it. Then you get an email with a link to look at it on your own preview address. Nothing goes live until you say yes.

Each client also gets one stable preview address, so "take a look at this" always means the same link.

One of those six requests taught us something worth passing on. The client pointed out that a line on the header wasn't true: they oversee the work, they don't do every job with their own hands. Searching the code said the line was gone after the fix. Looking at the actual pages found the same claim, worded differently, on four more. We apply a correction everywhere it's true now, not just where you pointed, and we check the rendered site instead of trusting the search.

What AI search reads

More people are asking ChatGPT or Gemini who to call instead of typing into Google. Those tools don't browse a site the way you do. They read what's in the page's code, and a lot of small business sites give them almost nothing to read.

This week our own site got two additions for them. The questions on our How it works page are now marked up so an answer engine can read them as questions and answers. We also added a short plain-text file that tells AI tools who we are and what we charge. It's built from the same facts as the site, so the two can't disagree.

The same setup is ready for client sites, and it goes on each one only after its owner has read every line of it.

Want to see where you stand? Open ChatGPT and ask it who does your trade in your town. If you don't come up, that's worth knowing, and it's usually fixable.

A demo that never shows a real name

One client runs their organization from a staff dashboard we built. They wanted to show it to other people without putting real members on a projector. So the dashboard now has a sandbox switch. Flip it and every screen runs on an invented organization with made-up people and months of made-up history. Real data can't appear in sandbox mode, because the two never load on the same page.

We found one thing along the way. A few of the invented names were one letter off from real members. We changed them. Invented should mean invented.

Next week

Pam gets a few more real calls, and the portal gets its first real request from start to finish. We'll tell you how both go.

Written with AI, signed by an engineer.