Instead of a cold pitch, a finished website: 200 of them in one day

The idea

Google Maps is full of small businesses with no website: a florist with 25 reviews, a shoe-repair shop with 185, a motorcycle shop with photos of every bike it has rebuilt. Customers did the marketing; there's just nowhere to send new ones.

For a consultant selling automation, that's an opening. A cold pitch is easy to ignore. A finished site with the business's own photos, hours and services takes seconds to judge, before any conversation.

So the client asked for software that builds such sites in bulk and introduces their services. Five rules:

1. No "AI slop": most owners could publish the site as is.
2. Nothing invented. Photos from the listing, not stock. Services from the listing, reviews and the business's own pages; prices only if the business published them. Services seen only in reviews aren't confirmed by the business. Too little data, no site.
3. The right person or nobody: email only an address that saved evidence ties to the business.
4. Cheap at volume: at least 20–30 sites a day on a tight AI budget.
5. No purchase required: the site stays up whether or not the owner answers the survey or buys anything, and comes down only on request.

What the owner gets

1. A personal email with a screenshot of their finished site and two sentences on where it came from. Instead of a pitch, one question with buttons: what takes the most time in your week? A link below skips straight to the site.
2. A four-question survey: the time sink, hours, who does it, what they'd hand off. The link identifies the business, so no typing names or emails. The answers show the consultant where automation could really help.
3. The handover page: files, plain instructions for a free host or their own domain, and an offer of paid help: edits, a domain move, booking reminders, review requests, projects from scratch.

The site opens right from the email; the files sit behind the survey. Useful first, questions later, on purpose.

By the way, email clients strip out forms, so the question is a set of link buttons. A click records the answer and opens the survey with it preselected.

What's on every site

One template, filled with the business's own data:

• First screen: the best photo, an "open now" tag computed from the hours and time zone, the Google rating, call and services buttons.
• About: from what reviews actually mention.
• Services in tabs. Prices only if the business published them, say on an old site or a menu; otherwise none, and "call us for prices" is cut.
• Gallery of listing photos with a full-screen viewer.
• Why customers come back: three or four recurring positive review themes, paraphrased, not quoted.
• Visit: hours with today highlighted, tap-to-call, directions and a map.

Each site is one HTML file plus a photo folder. It opens straight from the download, runs on any free host and can be edited in Notepad. A plain-text guide comes along for someone who has never touched a website.

How it ran

The September campaign on this pipeline introduced the client's brand and services to small businesses in 13 states.

I ran it on a Windows PC, from the command line or a simple menu. One command per stage (staged, qualify, make, deploy, mail); mail without --send only drafts emails for review.

A command guide and project notes cover every stage. Since the campaign, the pipeline has sat idle, ready for new cities and niches.

Scale and cost

• AI: 2.4–4.1¢ per site at list prices, measured on five production batches.
• Hosting: $0, with one Cloudflare Pages project on the free plan and a folder per business.
• Data, search and email are billed by their providers per request or per plan. They aren't included here: I didn't track them separately in this build.

How it works

Photos needed the most care: listings are full of menus, blurry shots, duplicates and decor. In one early run, a site's hero image was a restroom sink😅 Now a vision model screens every photo before it reaches a page.

Facts come from data. Words come from the model.

Claude built the first site end to end in a chat: research, writing, checking. It looked fine. But a model writing the whole page can "improve" a phone number or invent a price.

So every page splits in two:

• Facts from data: name, phone, email, address, hours, rating, photos and published prices. The template inserts them, and the structured contact fields never reach the copy prompts. Review texts and photos still go to Claude for analysis.
• Words from the model: headlines, about text, service descriptions and review themes — each written separately for its section of the page.

Mechanical guards sit on top. Promises like "I'll fix it", "reply to me" or "the same day" fail the build, and "call us for prices" is cut. The prompt bans quoting reviews word for word (they belong to their authors) and keeps criticism out of "why customers come back": "better three honest positive themes than four with a criticism among them."

Bonus: it got cheaper too. Splitting facts from copy meant $0.18 a site on the most capable model. Moving copy to Claude Sonnet, photos and classification to Claude Haiku, and caching every step brought it to 2.4–4.1¢ per site.

An email address only with proof

The hardest part wasn't building sites but finding someone to send them to. Businesses without a site rarely publish an email, and a loose search happily "finds" someone else's.

Version one took an address whenever the business name appeared anywhere in the page URL. It "found" emails for 14 of 27 businesses; about 3 were right. The rest belonged to a gift-card marketplace, a map website, an awards vendor and a different cleaner in another town.

Aggregators put business names in their URLs, so I rebuilt the rule for precision:

• a distinctive word from the name must be in the mailbox, email domain or page host, not just the path; generic words like "nails" or "cleaners" don't count;
• platform domains are blacklisted;
• three search steps: open web → social pages → the owner's name; most genuine addresses were on the businesses' own social pages;
• in production, an address stays only if the same page shows the business's phone or street address with the city, and that snippet is saved as evidence.

The moral: precision over coverage. Better to miss a lead than email the wrong business.

Who really needs a site

An empty "website" field on a listing doesn't prove there's no site. In the first test batch, four of the ten businesses judged to have no usable site had a working one: a booking site, an ordering page, a JavaScript app.

Wix, Squarespace, Square and ordering apps render with JavaScript, so a live site's raw HTML can hold as few as 27 words. Hence conservative rules:

• if a page loads, it's a site;
• a site found off the listing counts only if its host has two distinctive words from the name (one is often just the city);
• social profiles and link-in-bio pages don't count.

Listed sites that failed the automated check form a separate group: of 19,100 checked, about 470 didn't respond or only redirected to a social profile. Those businesses got a different email: "The website on your Google Maps profile is down. I made a working one."

Honestly, "is down" was too confident: the automated check only shows the site didn't open for it, not that it's down for everyone.

Takedowns

If an owner asks, their site is replaced with a short "removed at the owner's request" notice and the business goes onto a suppression list. Discovery, build and sending all check it, so the site is never rebuilt or re-sent.

The catch was caching: Cloudflare's edge first kept pages for seven days, so an edit or a takedown would have taken a week to show. Pages now revalidate every five minutes, while images, which don't change, stay cached.

Development: 9 days

Who this is for

• Web studios and freelancers selling sites to local businesses: show the finished site before the first call.
• Agencies that sell through outreach: a personal landing page or demo for every prospect.
• Software for local businesses (booking, reviews, ordering): a ready-made demo on each prospect's own data.
• Consultants who, like this client, want to introduce their services through something useful.

The same data could feed review summaries, local search audits or ad landing pages; this version is built for targeted outreach, with contact checks, draft review and suppression handling.