I built SiteScout to show my PM skills end to end: not just understanding the business problem, but rapid prototyping, fixing bugs in production and deploying it myself. The question it answers: can one person run a complete outreach product, from discovery and qualification to follow-up, on its own for $0 a month?
SiteScout is an outreach product I designed, built and run. It finds local businesses that are losing bookings to their own websites. Then it shows each owner a better version of their site before asking for anything.
Plenty of barbershops and salons rank on Google Maps but run websites that quietly cost them bookings. The site is slow on a phone, there's no way to book online, or the phone number is buried.
Fixing that is straightforward. Finding these businesses is the slow, manual part. You search a city, open every site, judge it, hunt for an email address and write a message. The message that finally goes out sounds like every other agency pitch the owner has already ignored.
That gave me two problems to solve. The first was finding the right businesses at a volume one person can't reach by hand. The second was earning a reply from an owner who has never heard of the sender.
Show, don't pitch. An owner who sees their own shop on a better site will reply.
This bet shaped the whole product. The first email names one specific problem the owner can check for themselves. After that, a personalised demo site, which I call a ghost site, does the persuading. It uses their real name, services and photos, and it's live for seven days.
Hours: Tue to Sat, 9 to 6. Walk-ins welcome.
To book an appointment please call us during business hours. Click here for directions.
Sharp cuts on Douglas Avenue.
Open today until 6pmIllustrative example. The business is made up.
I type a niche and a city. Everything after that runs on a schedule, and I only step in when an owner replies.
The only thing I type.
It splits the city into 3 km search circles on Google Places, because a single broad search stops at about 60 results.
It looks for online booking and a call button, runs Google's PageSpeed test, and scores the opportunity from 0 to 100.
An AI-written first email names what's wrong. Up to four follow-ups go out, timed for each owner's local morning.
It builds a demo from the business's real content and publishes it with a seven-day expiry.
It tracks clicks on the demo, reads replies with AI to separate interest from "no thanks", and catches emails that bounced.
An admin dashboard shows the queue, settings and run history, and can run any stage on demand. Built with Python, Notion as the CRM, Google Places and PageSpeed, Gemini, Cloudflare Pages and Workers, and a free Google Cloud server.
My first version qualified leads on seven technical PageSpeed signals. It found sites that were slow but not necessarily losing business. I rebuilt the score around conversion, so missing online booking and a missing call button carry most of the weight. Mobile performance counts, but it can't push a site to the top on its own.
"No online booking, and the site loads poorly on phones. Someone who can't book at 10pm books with the shop down the road."
When the crawler can't tell whether a feature exists, it scores a quarter of the weight. Not finding something isn't proof that it's missing, so "couldn't tell" should never score as high as a confirmed gap. It shouldn't score as if the site were fine, either.
I also detected each business's website platform, on the theory that an old platform meant an old site. It didn't predict pain, so I removed it. Online booking now decides qualification, and other signals only rank the leads that pass.
Trade-offFewer leads qualify, but each has a sharper reason to reply.
The system was already sending when I rebuilt the score. I shipped it in three steps: standalone with tests, then alongside the old rules without affecting anything, then as the live gate.
Trade-offAbout a day slower, with no risk of emailing the wrong businesses.
Google Places allows 5,000 free searches a month, then charges $32 per thousand. Discovery now stops at 4,999 calls however many leads it has found, and resumes on the 1st.
Trade-offLead volume has a monthly limit. Costs can never surprise me.
When real emails started landing in spam, I added one switch that stops all outreach, follow-ups included, while the mailbox warms up. New sends are capped at 10 a day.
Trade-offA full pause, which beat burning the domain's reputation.
The system used to ask for a new city only once the current one was used up. Now, when fewer than a day's worth of leads are waiting, it asks every cycle and suggests a mid-sized US city.
Trade-offMore prompts in my inbox, and no more idle days.
Every change to qualification traces back to the user story it served. That made it easy to cut scope later without losing track of why each piece existed.
ResultScope cuts stayed deliberate, not accidental.
Every one of these failed silently. My recurring fix was to make failures loud.
After the move to the cloud, the scheduler couldn't find the email tool, and two stages failed on every run.
Fixed the environment, documented it, and added failure alerts.A failed send still showed the lead as contacted in the CRM.
Added bounce detection, plus a free check that a domain can receive mail before a lead is accepted.Without the AI key, email copy couldn't be written, and qualified leads were dropped with no error.
The system now refuses to start without the key.A slow city search ran into the next cycle and created duplicate records.
Added a lock that every run, scheduled or manual, has to respect.The AI model's free tier allowed 20 requests a day, and anything over the limit was discarded.
Added retries with backoff and moved to a model with a workable free quota.The real bottleneck turned out to be deliverability, not lead quality.
Paused all outreach for mailbox warmup.I spent the early sprints on finding and scoring leads. The real constraint turned out to be whether emails reached the inbox at all.
Most of the incidents above caused no visible damage and still cost real leads. Every stage now fails loudly.
Free-tier limits, a monthly API ceiling and a daily send cap forced clearer choices than an open budget would have.
Running new logic alongside the old before switching let me change the core of a live system without harming a single lead.
SiteScout is a personal project I built on my own time to show the full range of PM work, from business understanding to prototyping, bug fixing and deployment. Business names and contact details are left out on purpose.