Skip to content
MakerGain
← Back to project

Paste a web address, Grow a Support Agent: SiteGPT’s Weekend Project and First Month at \$10,000

Edited by MakerGain

To learn AI, Bhanu Teja built SiteGPT on weekends and at night: paste a web address or upload content to generate a customer support chatbot that can be embedded on multiple sites, compiles unanswered questions on Mondays, and hands off to a human. It reached $10,000 MRR in its first month; the next month saw about $5,000 in churn offset by about the same amount in new additions, and at the time of the interview monthly revenue was about $13,000–$15,000.

For example, install a customer support bot on a SaaS website: paste the official website address into the input box, wait for it to finish scanning, and a chat window appears in the bottom-right corner of the site. A visitor asks, “Do you support SAML single sign-on?” and it finds an answer from the pricing page and help docs and replies; when the visitor gets impatient, they click once and are handed off to a human.

SiteGPT’s first version had just one feature: input a website link, and a chatbot is generated from the website’s content (Mixergy interview[1]). The website now describes it as “like a ChatGPT built specifically for your product,” trained on website content to answer visitor questions instantly; the input can be a URL, an uploaded file, or directly pasted text, the chat window can be embedded on any number of sites, and real chat history is used to improve answers (SiteGPT[2]).

Image: source SiteGPT actual image source[3], retrieved 2026-11-10, original interface retained.

What It Takes Over Is the “Repeated Answers” Part

This used to be human work: a visitor cannot find an answer on the site, sends an email or opens chat, and a support person copies a paragraph that was written long ago. What SiteGPT takes over is this repetitive labor, but it does not take over the whole process. The website says AI can handle most inquiries while leaving a one-click handoff to a human; every Monday it sends the team last week’s data, popular topics, and questions it could not answer; analytics group conversations by topic and outcome, and it also supports collecting leads (SiteGPT[2]). That list of “couldn’t answer” items is how it hands the conversation back to a person.

Another unsexy thing happens every day: knowledge expires. The site gets redesigned, a new feature ships, prices change, and the bot is still reciting old answers. So the pricing page writes refresh frequency into the self-serve plans: Starter refreshes manually, Growth refreshes automatically every month, and Scale refreshes automatically every week and automatically scans the sitemap (SiteGPT pricing[4]). This product line is plain: what it sells is a solution to the fact that “content expires.”

The data-source side is similar. The integrations page lists Google Drive, Dropbox, Notion, GitHub, YouTube, and others for accessing, syncing, or using training material; email can be answered automatically, or it can only generate drafts for the team to approve; Slack can answer in channels, DMs, and threads and send escalation alerts. Intercom and HubSpot are still marked coming soon and cannot be treated as already available (SiteGPT integrations[5]).

He Once Shut Down a Product Over a Hosting Bill

SiteGPT was not originally a business. Founder Bhanu Teja was working on another product, Feather, and SiteGPT was a side project he wrote in three or four weeks, on weekends and at night—in his own words, to learn AI. He did not first talk to customers about demand, nor did he expect it to generate revenue (Mixergy interview[1]).

This order is worth pausing on. The usual story is to see the pain point first and then start building; here he built it first and then looked at who needed it. And his earlier failure was his first product: a tool for making blogs for Notion, originally called MDX1, with 25 paying customers, but one month the hosting bill rose to $10,000, so he shut it down; only after rebuilding it and renaming it Feather did it grow to $5,000 MRR (Mixergy[1]). Twenty-five paying customers, and the bill could run into five figures.

One Month at $10,000, and the Next Month’s 5,000

Luck was concentrated in those launch days. After he posted on Twitter, the tweet went viral; the next day the product reached the Hacker News front page for several hours; the day after that it became the day’s top product on Product Hunt; in one month it reached $10,000 MRR (Indie Hackers[6]).

Then came the other side of the bill. He said the next month had about 5,000 in churn, and about 5,000 in new additions was needed to balance it; he described this as close to half of the first month—5,000 was the churn amount, not an exact customer count. The inflow and outflow were equally wide. His own explanation was that many of the people who came in the first month just wanted to try it and see what it looked like on their own website; they were not customers with real needs, and without need there was naturally no reason to renew (Mixergy[1]).

The pricing also lays out the limits very concretely. Starter is $59 per month and includes 1 chatbot, 1 user, up to 1,000 knowledge pages and manual refresh, and up to 4,000 GPT-4.1-mini or 400 GPT-4.1 messages per month; there is no free plan, and the 7-day trial requires a credit card and can be canceled before it ends (SiteGPT pricing[4]). The chat window can be embedded on any number of sites, but Starter also tightly caps the number of chatbots, seats, knowledge pages, and message volume. What it limits first is these capacities, not the number of websites.

Income figures need to be pinned to a point in time. At the time of the Mixergy interview, he said SiteGPT’s monthly revenue was about $13,000 to $15,000, and Feather separately had about $5,000; the two products were disclosed separately and cannot be added together as SiteGPT’s revenue today. In the same period, he said expenses were about $4,000 to $5,000, including hosting and marketing (about $4,000 the previous month), of which OpenAI API was about $1,000; when he was not doing promotion, expenses were about $2,000 to $3,000 (Mixergy[1]). Full costs were not disclosed, and profit remains unclear—one can only say the business is running; one cannot calculate its net earnings for it. On traffic, he said most traffic came from Google and direct visits, with Twitter at about 20% to 25%; AI directory sites brought very little, and at the time he was still trying YouTube and help content (same source).

There Are Others on the Same Shelf

Chatbase’s website says it can build customer support agents across chat, email, and voice, and it shows a Helpdesk for AI and human collaboration; Basic is $150 per month and includes 4,000 message credits per month, 3 members, and API access (Chatbase pricing[7], Chatbase[8]). Both are doing the same customer support work, but Starter is $59 per month and limits you to 1 chatbot, 1 user, 1,000 knowledge pages, and manual refresh; human handoff is a product capability on the website, while the weekly unanswered-questions report and team seats require moving up a tier. Chatbase Basic gives 3 members and API from its entry tier. The units differ—model-adjusted messages and message credits cannot be directly converted—so one can only compare configuration terms, not “who is cheaper.”

What It Sells Is Not the Model

There is nothing mysterious about the first version’s tech stack: OpenAI API, Pinecone for storing content, Cloudflare Workers for hosting; the chat window and sales page were made by a freelance designer who had worked on Feather (Mixergy[1]). The model was rented, and the retrieval was assembled.

What was really bought was another small thing: the company already had the answers written on its website—let it say them itself. Around this small thing is a pile of unsexy manual work—refreshing the knowledge base, reading Monday’s “couldn’t answer” report, turning unanswerable questions into answerable ones, and handing off to a human before the visitor gets impatient. Launch momentum pushed the first month to $10,000 MRR; he later said that business customers with real needs only came four or five months later (Mixergy[1]). That sentence explains whom this business sells to better than any monthly number.

Continue reading

AI support
Human handoffs are free: Fini puts its AI-support bill on “actually resolved”

Fini's AI support agent handles chat tickets inside a customer's existing helpdesk. Behavior trees and a deterministic executor read plan and account state node by node, then call customer system interfaces to perform actions such as canceling subscriptions. Human handoffs are not billed; by default, only end-to-end resolutions without human takeover count as billable tickets. The team cleared its first product and rebuilt after four weeks embedded with a fintech support team, choosing not to replace existing helpdesks and requiring human approval before publishing knowledge corrections proposed by background AI.

AI support
Teach the bot to admit uncertainty first: a side project bringing in about \$500 a month

Andy and his two cofounders at London's HeyDeacon built Deacon for founders and small teams still answering support themselves, starting with their own use. When it cannot answer, it admits that and asks visitors for an email address. Only after an address is actually provided is the conversation marked unread and the owner emailed; human replies can enter the knowledge base. The team focuses on the workflow around the model. Knowledge needs manual recrawling or publishing-script triggers to update, while changing models does not require rebuilding the product.

AI search
When AI does not mention your company, he publishes articles on the customer's own site

MyZia helps brands get mentioned in AI answers: it scans visibility, finds questions where competitors are cited but the brand is absent, then writes articles for the customer’s own website. Founder Dhairya Patel also runs a development agency. He did not stop at a dashboard that offers advice. He chose end-to-end automation. Keeping articles on customers’ domains also requires connecting each CMS, making publishing the first part to hit a scaling barrier.

AI support
The database crashed in his first week: Zipchat wants every store's support ready out of the box

Zipchat is an AI support agent for ecommerce. When Ruslan took over in January 2024, the team struggled to account for its MRR and customer sources, and the database started crashing in the first week. He first fixed the books, using Mantle to make MRR visible and collecting feedback on replies. His blog reports a 25% MRR increase about two months later, reaching US$20,000 for the first time. The product claims 30-minute setup, but brand rules, helpdesk connections and answer corrections still require store-by-store configuration.