Use Chatbase when visitors need answers inside your website: it offers a deployable agent, website chat and visitor-facing workflow. Use ChatGPT when staff need to draft and check answers internally. A ChatGPT subscription or custom GPT is not a website embed; an OpenAI-powered public chat requires a separate API-built product, with its own design, data controls and usage costs.
The better choice depends on the task, not which model sounds more capable. A public assistant must know the current published price, refuse unsupported promises, collect only necessary details and hand a difficult enquiry to a person. An internal ChatGPT user can work on a draft, inspect the source and decide what to publish without making a visitor wait in an uncertain chat.
Decide who will see the answer first
For a small site, begin by sorting common questions into three groups. Public facts include opening hours, service descriptions and published booking steps. Account-specific questions involve an order, contract or personal detail. Judgement questions ask for a custom price, unusual refund or professional opinion. A website agent can often handle the first group from approved pages. The second should use a verified identity process or be handed to staff. The third usually needs a human decision. A bot that handles everything can create more repair work than a clear form.
Chatbase's product and pricing pages position it as a website and channel agent trained on a business's source material. Its plans include an agent, member seats and monthly message credits; the Basic tier adds a Helpdesk inbox for human follow-up. ChatGPT's Business plan is an employee workspace for chat, Work and Codex. It can help staff write FAQs or examine conversations, but the normal ChatGPT interface is for signed-in users. OpenAI's own GPT FAQ says GPTs are designed to work in ChatGPT and are not a way to embed ChatGPT in an external website; for an outside product, it points developers to the API. Custom GPTs are also being retired: the same FAQ gives 11 December 2026 as the standard retirement date, with plugins as the migration route. Neither is a basis for a public site assistant.
That matters for a wedding planner's website. A visitor should not have to leave the planner's site, create a ChatGPT account and find a shared GPT merely to ask whether a date is available. The planner could put a Chatbase widget on the site, but availability should come from a maintained booking source or a person, not an unverified AI guess. ChatGPT can still be useful internally to draft the wording for “how we handle date enquiries” and to review a sample of bot transcripts. Those are complementary jobs.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat an embed changes operationally
An embedded chat opens on the visitor's page, where they are already deciding whether to contact you. That reduces the steps between a question and a reply, but it also creates an immediate service promise. The owner must decide who maintains the content, who checks the bot's answers, what happens outside staffed hours and what appears if the service or credit allowance is unavailable. A visible “Contact the team” link and ordinary enquiry form remain important. The site should still work if the widget fails to load, so the bot should supplement rather than replace those routes.
Chatbase can be configured with site content and an agent, then tested before the embed is made public. The basic deployment path is to curate approved FAQ and policy pages, write instructions for what the agent may answer, test a set of real visitor questions, install the vendor's site widget through the website's normal change process and monitor the first week's conversations. Chatbase's docs give a widget script and an iframe embed, with guides for site builders such as Wix, Webflow and Framer, and its pricing FAQ notes that an agent set to public can be used by anyone with the link. Do not paste an untested widget into every page. A small pilot on a contact or service page lets the owner find wrong answers without exposing the whole visitor base.
A ChatGPT Business Project could hold the same approved FAQ and let staff ask “How should we respond to this enquiry?” But that Project is not automatically a public website assistant. OpenAI's custom GPTs are likewise designed to run inside ChatGPT. If a developer builds a visitor chat with OpenAI's API, the firm must supply the interface, retrieval of current business content, abuse controls, logging, human handover and hosting. OpenAI's Business overview says a Business subscription does not include API usage, which is billed separately, and its models and limits page reminds subscribers that the Services Agreement prohibits using ChatGPT to power third-party services. A subscription seat does not remove those responsibilities or turn into a public chatbot. The website embed options tutorial explores that distinct build route.
Lead capture is a service workflow, not a chat transcript
Ask what happens when a visitor is ready to buy. The bot might request a name, contact route, service, preferred time and a short description. It should explain that staff will review the enquiry and should avoid requesting unnecessary personal or payment details. The submitted information must land somewhere the team actually checks. If the business already has a good form and shared inbox, the agent can answer public questions and send visitors to that existing form. A new Helpdesk is valuable only when it improves ownership of the follow-up.
Chatbase handles this step with its Collect Leads action. Its docs describe two modes: a form inside the chat bubble with up to three fields (name, email and phone, no custom fields), or a conversational mode that asks one question at a time and supports custom fields such as company or budget. Either email or phone must be switched on and required, and an optional destination sends the lead to a CRM or inbox. A separate Escalations action hands a conversation to a person by opening a ticket in the Chatbase Helpdesk or a connected helpdesk. For a print shop, a sensible set-up is the conversational mode collecting name, email, quantity and deadline, with the destination set to the shared quotes inbox the estimator already watches. Remember that an action can cost an extra model request, which the next section's arithmetic needs to include.
For example, an IT support firm asks its bot to answer public maintenance-plan questions. A visitor then says, “Our server is down; can you guarantee recovery today?” The correct workflow is to state that a person must assess the incident, collect an approved callback route and provide the urgent contact method. An illustrative bad output is “Yes, recovery is guaranteed today” because an old marketing page contains the phrase “same-day support”. Remove or qualify that source, add an escalation instruction and retest. The chatbot guardrails tutorial covers such promises in detail.
A print shop has a different issue. A customer asks, “Can you print 2,000 brochures by Friday?” The bot may know standard lead times but not today's production capacity or whether the artwork is ready. It should say which files and specifications the team needs for a quote, then route the request. ChatGPT inside the office can help the estimator draft a reply after the schedule is checked. It should not be represented to the customer as a live production system unless the business has built and tested that integration.
Set the data boundary for public conversations
A visitor-facing assistant receives whatever people type, including details the business never requested. The owner needs a data map: which fields are collected, where transcripts are stored, who can open them, how long they remain and whether they pass to another tool. Chatbase's pricing FAQ says it does not use customer data to train or fine-tune any AI model on any plan, and that data is encrypted at rest and in transit; review the account terms and admin settings for the plan you buy against the rest of your data map. If a staff member uses ChatGPT to review transcripts, decide whether they will paste full visitor messages or a redacted summary into the approved workspace. OpenAI's Business pricing states that business data is not used for model training by default, but that does not replace the firm's consent and access decisions.
An illustrative privacy rule for a yoga studio is: “The bot may answer from public class pages. It must not ask for health history. When a visitor raises an injury or suitability concern, direct them to a person using the existing contact form.” This keeps the agent within the information the studio can safely publish. A wedding planner might prohibit uploads of guest lists and supplier contracts through the public chat. A print shop might accept a general product enquiry but direct artwork files to its established upload process. Each rule reflects a different workflow, even if the software widget looks identical.
Do not assume a ChatGPT API integration inherits the controls of the firm's Business subscription. OpenAI distinguishes ChatGPT plans from building with the API. A custom public assistant requires a separate review of API terms, retention, authentication and billing. Likewise, an employee's private ChatGPT account should not be treated as a substitute for an approved company process when it handles visitor details. Document the account, service and data path in the purchase decision.
Compare a useful answer, not a generated message
Chatbase sells message credits, and selected models can consume more than one credit per response. A visitor may need several responses or may abandon a wrong answer. The true operating metric is the cost per useful outcome: an accurate answer that resolves a question or a complete enquiry that reaches the right person. Count successful answers and properly handed-over enquiries separately. Do not claim that every credit equals a resolution.
Take an illustrative five-person print shop with 2,500 monthly site visitors. Suppose 8% open the chat: 200 conversations. Each uses three AI responses on a model that consumes two credits per response, for 1,200 credits. Chatbase Hobby currently includes 700 credits at $32 a month billed yearly; a 1,000-credit auto-recharge purchase is listed at $40. The cash outlay in the peak month could therefore be $72, with unused purchased credits available later under the vendor's stated non-expiry rule. This is a scenario, not a traffic benchmark or fixed bill.
In the same illustrative month, suppose 120 conversations receive correct answers, 30 are handed to staff with usable details, 20 need correction and 30 end without a useful outcome. The 1,200 credits produced 150 useful outcomes under that definition. Dividing $72 by 150 yields 48 cents of vendor cash outlay per useful outcome in that particular month, before staff time and annual commitment effects. If only the $32 subscription is paid in a quieter month, the cash ratio changes. More important, the 20 incorrect answers must be fixed; low cost does not excuse a false delivery promise.
Now compare internal ChatGPT. Suppose a staff member spends 45 minutes each week drafting replies to complex enquiries with ChatGPT and another 15 minutes checking them. That is four hours a month if the pattern repeats for four weeks. The Business seat may already be paid for other tasks, but the staff time is real. It may be the right arrangement when the site has few enquiries and a human reply is expected. The comparison is therefore not “$72 versus $20” as if these products performed identical work. One is a public automated route; the other supports an employee who remains responsible for the response.
A custom API chatbot has still another cost structure: usage charges, hosting, maintenance, source updates and the person who investigates failed conversations. If the print shop needs a live order-status lookup, that integration might be justified, but the owner should request a scoped build estimate and security review. If it only needs to answer six public FAQs, the custom build may be disproportionate. The build-or-buy comparison helps frame that purchase without confusing it with a ChatGPT seat.
Test the failure paths before the launch path
Write ten questions from actual visitor emails and search queries. Include two routine questions, two price enquiries, a complaint, a request involving personal information, a question whose answer is absent from the site, a request that needs a booking or order lookup, a misleading question containing a false premise and a request to speak to a person. For each, mark the allowed answer or handover. Run them in the Chatbase preview and check not only the text but also whether the right enquiry record is created.
One illustrative test is “Your website says every poster is waterproof; will mine survive outdoors for a year?” If the page only says a specific stock is water resistant, the bot must not confirm a one-year guarantee. A second is “Change the delivery address for my order”; the agent should not accept the change without the business's verified account process. A third is “What is your cheapest price?”; if prices depend on specifications, it should collect the required quantity and finish rather than invent a number. These tests expose the boundary between public facts and transactions.
During the first live week, review every incorrect or escalated conversation. Classify errors as stale source, missing source, prompt rule, unclear visitor request, identity problem or staff handover delay. The remedy differs. A stale price needs a content update; a missing handover owner needs a rota; a hallucinated guarantee needs a stronger answer boundary and perhaps removal of ambiguous marketing copy. The chatbot measurement tutorial provides a broader scorecard. Do not improve the metric by hiding hard conversations from the sample.
Retest when the site changes. A new seasonal offer, altered returns policy or changed service hours can make a previously correct bot wrong overnight. Assign a named content owner to update the source and run the ten-question suite after publishing. Chatbase's 30 October 2026 changelog added drafts and version history on all accounts, so edits can be tested before customers see them and a bad change can be rolled back to an earlier version. If the website's content team cannot maintain that routine, internal ChatGPT drafting with human approval may be safer than a public assistant whose knowledge quietly ages.
Use an ownership sheet for the first month
A simple ownership sheet makes the pilot reviewable. Give the website editor responsibility for public source pages, the service lead responsibility for escalation wording, the inbox owner responsibility for response times and one manager responsibility for credit purchases. Write each name in the firm's internal copy, not in the bot's public instructions. Set a weekly review time and a way to take the widget offline if it begins giving harmful answers. A pilot without an owner can drift from “helpful site feature” into an unattended customer service channel.
For an illustrative wedding planner, the source owner updates the FAQ when packages change on Monday. The service lead checks three questions about deposits and date holds on Tuesday. The inbox owner checks that the test enquiry appeared in the normal queue and received a human reply. The manager reviews Friday's transcript sample and credit total. If the bot gives an old deposit percentage on Wednesday, the team knows which page to correct and who will retest it. The weekly routine is small, but it turns a public AI answer into a managed service.
Include accessibility in that pilot. Test the chat with a keyboard, a screen reader and a narrow mobile screen. Can a visitor close it, return focus to the page and find the ordinary contact link? Does a large widget hide the quote button? Can someone who disables scripts still see the published FAQ and contact form? These are website decisions, independent of model quality. A bot can answer perfectly and still make the site's main route harder to use. Ask the person responsible for the site to sign off the interface before the widget appears on every page.
Set a stop condition before launch. For example, the print shop might pause the bot after two incorrect price or delivery promises in a week, or if urgent handovers sit unanswered past the team's own stated response time. The numbers are an illustrative operating policy, not a vendor threshold. A pause should show a normal contact route, preserve the transcript sample for diagnosis and trigger a source or rule correction. This makes the public pilot reversible without leaving visitors stranded.
Choose a route the team can actually operate
Choose Chatbase when enough visitors ask repeatable public questions, the firm can maintain approved content and it has a working route for exceptions. Choose ChatGPT internally when enquiries are few, mostly bespoke or sensitive, and a person can review each answer before sending it. Choose an API-built assistant only when the website needs a tested integration or control that a packaged tool cannot provide and the firm will fund ongoing ownership. A static FAQ or accessible form may still beat both for very simple questions.
Write the decision down with four facts: expected monthly conversations, observed credit use from a pilot, the percentage of useful outcomes and the person who owns updates and escalations. For the print shop, a trial showing 150 useful outcomes from 200 chats would justify a larger pilot only if the 20 wrong answers are corrected and the follow-up team accepts its workload. For the wedding planner, five high-value, highly bespoke enquiries may justify a human process with ChatGPT drafting support. The right product follows the service promise the business is willing to make.
Further reads
- Chatbase Pricing Plans: Hobby, Standard or Pro for a Small Site — Size Chatbase credits and plan cost.
- How to Add an AI Chatbot That Captures Leads on Your Website — Plan the enquiry capture route.
- AI Chatbot vs Live Chat vs Help Desk: What a Small Team Needs — Compare staffed support with AI chat.
- What to Ask an AI Chatbot Vendor Before You Sign Up — Check terms before connecting visitor data.
- How to Test a Customer Chatbot Before It Goes Live — Check visitor questions before release.
- Carpet Cleaning Quote Template: Rooms, Stairs and Pets via AI Chat — A carpet cleaning quote layout with a worked example, and an AI intake chat that asks about rooms, stairs, pets and stains, then hands you a summary to confirm.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Chatbase pricing page and pricing FAQ (chatbase.co/pricing); Chatbase docs on the Collect Leads and Escalations actions and site embeds (chatbase.co/docs); Chatbase changelog on drafts, models and tests; OpenAI GPTs in ChatGPT FAQ, including website embedding and GPT retirement (help.openai.com/en/articles/8554407); ChatGPT Business models and limits (12003714) and Business overview (8792828); OpenAI Business pricing (openai.com/business/pricing/). Checked 5 November 2026.