ChatGPT Chatbot for Sales: Qualify Website Leads Without Losing Them

Put a chat widget on your site that asks four things (need, size, timing, contact details), scores the answers against rules you write, and hands the promising ones to a person with the conversation attached. Use a ready-made bot such as Tidio's Lyro or Intercom's Fin if you want it live this week; build your own only for control.

A custom GPT is not an option here: OpenAI has scheduled custom GPTs to retire on 11 December 2026, moving them to ChatGPT plugins, and both live inside ChatGPT rather than on your own website. "ChatGPT-style" now means a bot that uses a language model behind a chat box, with your rules around it. The model is the easy part. The qualifying rules and the handover are where leads are won or lost.

Follow me on Instagram@sagnikteaches

Decide what a good lead is before you touch a tool

A bot that asks questions without a purpose is a form with a personality. Write down, in a sentence each, what separates a lead worth a person's time from one that is not. For a service business, four questions usually carry it.

Connect on LinkedInSagnik Bhattacharya

1. Need. What are they trying to get done, in their own words?
2. Size. Quantity, number of people, budget range or a rough scope.
3. Timing. When do they want it started or delivered?
4. Contact. Name, email or phone, and the best time to reach them.

Then define three outcomes. Hot: matches what you sell, has a size above your minimum and a start date within your window. A person is told straight away. Warm: a fit, but vague on size or timing. They go to the inbox for a same-day reply. Not a fit: outside what you do, a student asking for free help, a supplier selling to you. The bot answers politely and points to the right place. Without those three outcomes written down, any bot you buy will qualify everyone or no one.

Subscribe on YouTube@sagnikteaches

Ready-made, DIY or a mix: what each costs and where it bends

There are three routes. The prices below are the vendors' own list prices as I read them in November 2026; plans change, so check the page before you commit.

For most small businesses with fewer than a few hundred chats a month, the ready-made route wins on time, because the widget, human takeover and inbox already exist. The DIY route earns its place when you need a very specific flow, a particular CRM, or want conversations to stay on your own systems. The comparison of Shopify Inbox and Tidio goes further for stores, and the Fin pre-sale tutorial covers that tool's sales use in detail.

The conversation script for a service firm, with sample exchanges

Take an illustration: a small commercial cleaning firm whose website gets about 400 visits a month. The owner wants enquiries qualified out of hours. The bot's job is to greet, ask four things and hand over. A sample exchange (illustrative):

Bot: Hi, I can help you get a quote. What kind of premises do you need cleaned?
Visitor: an office, 2 floors
Bot: Thanks. Roughly how big is it, and how often would you like it cleaned?
Visitor: about 300 square metres, 3 evenings a week
Bot: Got it. When would you like to start?
Visitor: ideally next month
Bot: Great. What's the best name and email for our team to send a quote?
Visitor: [first name], [email]
Bot: Thanks [first name]. A person on our team will reply by 10am tomorrow with next steps. Anything else we should know?

That lead is hot: it fits the service, has size and a start date inside the window, and gave a contact. It goes to the owner with the transcript, not a cold summary. A second visitor who types "do you do carpet cleaning for my house?" is not a fit for a business that serves offices; the bot says so kindly, and offers a link to a directory or a referral if the owner wants one.

Two points about wording. First, keep the bot's replies under about 30 words; long answers read as a lecture in a chat window. Second, do not hide that it is a bot. A clear first line such as "I'm the assistant for [business]; a person takes over for quotes" sets expectations and avoids complaints later.

The instructions that keep the bot on rails

On a ready-made tool you enter these as the bot's instructions and training content; on a DIY build they are the system prompt. Either way, put the rules in plain words.

ROLE: You are the website assistant for [business], which provides [services].
You qualify enquiries and pass good ones to our team. You do not close sales.

ASK, in this order, one question at a time:
1. What do they need? 2. How big is the job (area, quantity, people)?
3. When do they want to start? 4. Name and email (or phone).
Stop asking once you have all four. Never ask for payment details.

FACTS YOU MAY STATE: [opening hours], [service types], [what is included],
[minimum job size]. For anything else, including prices, say:
"A person on the team will confirm that." Do not guess.

SCORING: Hot = fits our services AND job at least [minimum] AND start within
[X] weeks. Warm = fits but vague. Not a fit = outside services.

HANDOVER: If Hot, or if the visitor asks for a person, a call, or is upset,
say a person will reply by [time] and stop. Output the lead record in the
format below.

NEVER: promise a price, a discount, a date or a result. Give legal, medical
or safety advice. Claim to be a person.

Note the line "stop asking once you have all four". Models like to keep being helpful, and a bot that asks a fifth and sixth question loses people at the finish line. Test it by deliberately answering several questions in one message and by refusing to give an email.

The handover: what the person receives and how fast

Qualifying is wasted if the lead lands in an unread inbox. Design the handover as carefully as the questions. Each hot lead should trigger an alert in the place your team already looks, whether an email, a team chat message or a CRM task, and it should contain the same fields every time.

LEAD RECORD
Score: Hot
Name / contact: [first name], [email / phone]
Need: office cleaning, 2 floors, about 300 sq m, 3 evenings a week
Timing: start next month
Budget stated: none
Questions the visitor asked: whether we bring our own equipment
Conversation: [link or full transcript]
Captured: [date, time], page: [URL of the page they were on]

Agree a response time and write it into the bot's last line. A hot lead contacted within an hour is a different proposition from one contacted two days later, so promise only what the team can deliver, including evenings and weekends. If nobody is available out of hours, the bot says "we reply by 10am the next working day" and the lead record arrives in the morning queue. Tools like Tidio can pass the chat to a live agent and also pipe leads to a spreadsheet or CRM; the Tidio and Zapier tutorial shows the plumbing: Zapier's Tidio trigger fires when the bot sends a contact from a step you choose, which is the moment a lead becomes hot.

Logging so you can improve the bot

Every conversation should leave a row in a sheet or CRM, even the ones that went nowhere. Capture: date, page, source of the visit if your tool records it, the four answers, the score, who took over, what happened next (replied, quoted, won, lost) and the visitor's first message. After a month you can ask an AI to read the log and tell you where visitors dropped out, which questions stalled, and which fits the bot wrongly rejected.

Check privacy while you are at it. Visitors are giving names and contact details to a chat box, so say on the widget what you will do with them, keep transcripts only as long as you need and make sure your privacy notice covers it. If you have visitors in places with data-protection rules, ask an adviser what consent wording applies. The chatbot prompt template tutorial has a test set you can reuse to find leaks before launch.

Cost of a conversation on a DIY build, in numbers

A small build calls a model through its API, and the bill depends on how many tokens (the small chunks of text a model reads and writes) go in and out. Each turn re-sends the instructions and the conversation so far, so the input grows with the length of the chat. Take an illustration: instructions of 1,500 tokens, a visitor who sends eight messages of about 25 tokens, and eight bot replies of about 120 tokens each.

Input across the eight turns is 8 x 1,500 = 12,000 tokens of instructions plus the growing history: turn 1 sends one visitor message, turn 8 sends all eight visitor messages and seven bot replies, about 4,300 tokens over the whole chat. That makes around 16,300 input tokens. Output is 8 x 120 = 960. At Anthropic's list prices for Claude Haiku 4.5, $1 per million input tokens and $5 per million output (as of November 2026), that is about $0.016 plus $0.005, near $0.021 a conversation, or about $21 for a thousand chats. The newest Haiku on the same price page is listed lower still for short prompts, and the larger models are listed at two to ten times Haiku 4.5's rates. The model cost is small next to everything else, so choose on answer quality and test with real questions. Prices change, so check the current price page, and add the cost of hosting the widget and your own time to keep it working.

Compare that with a ready-made plan. Tidio's entry Lyro tier, $32.50 a month for 50 conversations, works out at about $0.65 a conversation; larger blocks are priced on its page. Fin charges per outcome, and qualified prospects are the dear ones: say 150 chats a month end in 30 qualifications at $9.99 and 120 disqualifications or self-serve routings at $0.99. That is $299.70 + $118.80 = $418.50 a month, and four times that at 600 chats with the same mix (chats passed to your team without an outcome are not charged). The DIY route costs more to build and less to run at volume, so the crossover depends on your chat count, your lead mix and how much you value your own hours. Run the sum with your own traffic before deciding.

The same bot for a wholesale distributor and a translation agency

The four questions stay; the wording and the score change with the business. Seeing two more cases shows which parts are yours to adapt.

A wholesale distributor of catering supplies. Visitors are mostly trade buyers or members of the public who found the site. The need question becomes "What products are you looking for, and are you buying for a business?" Size is a monthly spend or order quantity, and timing is "one-off or regular". The score rule is simple: a business buyer with a regular order above the account minimum is hot; a one-off small order goes to the web shop link; a consumer is not a fit and is pointed to retail stockists. The handover includes the products named, so the sales rep can open the call with a relevant price list instead of "what are you after?"

A translation agency. Need is language pair and document type; size is word count or pages; timing is the deadline; contact adds the company name. A 12,000-word contract due in four days is hot; a visitor asking for a free sample translation of a paragraph is warm at best. The bot must never promise a turnaround, because that depends on translator availability, so its instruction says "a project manager will confirm timing". It should also ask whether the document is confidential and tell the visitor not to paste it into the chat, a rule that belongs in the greeting. Sensitive files go through a secure upload your team chooses, not the widget.

Before and after: a contact form against a qualifying chat

A contact form is the usual starting point, and the comparison is worth making honestly. The form asks for name, email and "message". Visitors type "need a quote" and leave, and the owner replies with four questions. The first exchange is spent finding out what the chat could have collected immediately.

With the bot, the same visitor answers the four questions in under two minutes, in a format that is quick on a phone, and the owner replies once with a quote or a call time. In an illustration, a firm that received 20 form enquiries a month and replied to each with a follow-up of questions saw about half answer; the other half dropped out at the second step. Collecting the answers in the first interaction removes that step. The numbers are invented, and your own may differ, so measure the share of leads that reach a quote before and after, over at least a month, rather than assuming.

What to put in the bot's approved facts

The bot is only as accurate as the facts it may use, so write a short fact sheet of one page, in plain sentences, and update it whenever things change. Include: what you sell and what you do not; areas or customers you serve; minimum job or order sizes; opening and reply times; how quotes work (free, with a visit, with a deposit); how to reach a person; and two or three lines about what makes you different. Leave out prices unless they are fixed and public, and leave out any promise you would not repeat to a customer's face.

Add a short "do not answer" list: legal or medical questions, anything about a competitor, refunds on existing orders (those go to a person), and any request that involves sending payment details. Visitors can ask all of these in chat, and a bot with no rule will make something up. Tidio lets you teach Lyro from your own Q&A, uploaded FAQs and website pages, and you can find the details in the Lyro training tutorial; the principle is the same everywhere: a small, correct fact base beats a large, stale one.

Out of hours and when the visitor wants a person now

The riskiest moment is a visitor who wants a human and cannot have one. Decide in advance what happens in three cases. During staffed hours, the bot offers "connect me to a person" at any point and hands the chat over live. After hours, it says plainly when a person will reply and collects contact details so the reply can happen. When the visitor is upset, it stops qualifying at once and escalates, with the transcript attached, flagged so the first reply is a person's apology, not a sales pitch.

Honesty about availability is part of qualifying. A visitor told "we reply by 10am tomorrow" and then replied to at 9.30 is a better lead than one told "we'll be right with you" at ten at night.

Test it with ten awkward visitors before it goes live

Write ten test conversations and run them yourself. Include these:

1. A visitor who asks for a price straight away.
2. A visitor who gives all four answers in one sentence.
3. A visitor who refuses to give contact details.
4. A visitor who is angry about an earlier order.
5. A visitor outside your service area or line of work.
6. A visitor who asks the bot if it is a person.
7. A visitor who types in another language.
8. A visitor who asks for a discount.
9. A visitor who asks for something you do not offer but could.
10. A visitor who tries to make the bot ignore its instructions.

Score each: did it stay in role, avoid inventing, collect what it needed, and hand over when it should? The usual failures are quoting a made-up price, promising a specific date, and failing to hand over an upset visitor. Fix by tightening the instructions, then re-run all ten, not only the one that failed.

How a good bot still lost leads, and the fix

In one illustrative case, a bot qualified leads accurately for a month, but the owner noticed that hot leads were not turning into quotes. The conversation log showed the cause: the handover emails went to a shared inbox that two people each assumed the other was watching. Median time to first reply was 26 hours, against a promise of "by 10am tomorrow".

The fix was not a better bot. The team named one owner for the alert, set a rule that a hot lead unclaimed after two hours pings a second person, and changed the bot's closing line to match what they could do. In the next month, the median first reply fell below four hours. Handover is a business process, and the bot is only as good as the process behind it.

Keeping the bot honest week by week

For the first month, read ten random conversations a week and tick three things: no invented facts, correct score, clean handover. After that, a monthly read is enough. Add every new question you see to the bot's approved facts so it stops saying "a person will confirm" for questions you can answer. When prices or opening hours change, change them in one place and re-run your ten tests.

Track four numbers: share of visitors who start a chat, share who finish the four questions, share of finished chats scored hot, and share of hot leads that become quotes. A drop at any step tells you where to look. Low starts mean the greeting is weak or hidden; low finishes mean a question is too hard; low hot share means your rules or your traffic need review; low quotes mean the handover is the problem. For a wider view of leads from several sources, the sales chatbot tutorial on qualifying and booking calls covers the pattern without a particular tool. If you want help choosing between a ready-made bot and a build for your own site, that is what my AI implementation consultation covers.

Questions before you put a sales bot on your site

Can I embed a custom GPT on my website?

Not as a website bot. Custom GPTs live inside ChatGPT, and OpenAI has scheduled them to retire on 11 December 2026, moving them to ChatGPT plugins. A website bot is a separate product: a ready-made chat tool with its own AI, or your own small build that calls an AI model's API from your site.

How many questions should the bot ask?

Four or five at most before offering a person: what they need, how big the job is, when they want it and how to reach them. Every extra question loses visitors. If the bot has to ask more, ask them after a person has taken over.

Will the bot make up prices or promises?

It can, if you leave it free to answer. Give it only the facts you approve, tell it to say "a person will confirm" for anything else, and test it with ten awkward questions before launch. Review real conversations weekly for the first month.

What does a ready-made bot cost compared with building my own?

Ready-made tools charge monthly plans or per conversation or outcome, so costs track traffic. A small build calls an AI model's API, where a short conversation usually costs cents, but you also pay in setup time and upkeep. Compare the totals at your actual lead volume, not the headline price.

What if the bot cannot reach anyone after hours?

Tell the visitor honestly when a person will reply, capture the contact details and the need, and send the lead to your inbox or CRM at once. A bot that promises a reply it cannot deliver loses more leads than having no bot.

Further reads

• Ecommerce AI Chatbot Design: Order Status, Returns and Sizing Flows — Design order-status, returns and sizing flows for a store.
• AI for Self Storage: A Unit-Size Chatbot for "What Fits in a 10x10?" — How to build a self-storage unit-size chatbot: size guide, item-to-unit logic, sample conversation, lead capture and a test set.
• Garage Conversion Leads: Qualify by Door Size, Floor and Planning With AI — A seven-question intake, a six-photo checklist and green, amber, red sorting that tells a garage conversion firm which enquiries deserve a survey visit.
• Solar AI Chatbot for Installers: Qualify Leads by Bill, Roof and Shade — Question order, a lead score, a rules-first prompt, a hand-over note and ten test enquiries for a solar chatbot that never promises savings.
• Avoca AI for ServiceTitan Shops: What It Does, Who It Fits — Assess Avoca's ServiceTitan booking, AI CSR and lead follow-up against a small shop's actual missed-call problem.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Tidio pricing page, Intercom Fin pricing page, Zapier's Tidio integration page and OpenAI's custom GPT retirement FAQ (read November 2026); Anthropic API pricing page; worked figures are illustrative. Check every price on the vendor's page before buying.