Harvest a client's whole web presence with AI

Before you send a client a list of questions, find out what the internet already knows about them.

Half of what you'd ask is already online: services, prices, opening hours, photos, what customers say. Collect it first, so the questions you send are the ones only the client can answer. It also turns up things the client forgot they ever published.

Start with a map of where to look. The site, every page of it. Map listings and booking platforms, which often carry different hours from the site. Social accounts, old and current. Directories, local press, and any old domain that's still up.

Then hand the job to your coding agent. This is the bakery's research day, the fifth of October. Give it the site's address and every profile you know: the social page, the reviews listing, an archive of old versions, and the owner's emails to check against. Ask for one file per source, word for word, stamped with where it came from and the day it was captured. A summary loses the exact wording, and that's what you're after. And tell it to mark anything it can't confirm, never fill it in.

It works through the list and writes one file per source. Every file carries its source and its capture date, the fifth of October. Open the hours. They're copied exactly as the site prints them, and marked: the owner's email from the third gives different hours. Which is current is a question for the owner, not a guess for the agent. The About text is word for word, and the founding year isn't on any page, so the file says so and leaves it blank. Keep that gap open. It's exactly where a first draft will be tempted to invent one. The reviews keep the customers' own words, like the sourdough sells out by ten. You'll want that wording later, and no summary keeps it. And the old copy from the archive keeps the archive's own date, because an old page tells you what used to be true.

An agent that follows links is blind to anything nobody links to, and that's where the surprises usually are: a booking profile with old prices, or an old domain still in search results. Both can tell customers things that aren't true.

So check by hand. Search the business name with the town. Search the phone number and the street address, since those survive a rebrand. Run a site: search on every domain you find, though Google's own documentation says those results aren't complete. Look each domain up in an archive of old versions. Everything new goes into the same folder, with its address and the date you found it. And at the end, ask the client one question: is there anywhere else you're listed?

Harvesting also tells you what to ask. Anything public, you find yourself. What's left goes in a short questions file: which hours are current, who took the photos, what they'd like gone from the web. That's the file the client actually sees.

Last, capture the site as it is today, at the same fixed widths every time: 1440, 768 and 390 pixels. Your agent can take them if it has a browser tool connected. They're for fair before-and-after comparisons, not proof the site is responsive, and once the old site is gone you'll be glad you have them.

The crawl finds what's linked. Finding what isn't linked is your job.