How to Use CustomGPT.ai to Answer Buyers and Tag Each Lead
By Joris van Huët
Enterprise Interim CMO & Marketing Leader · 15+ years · 50+ orgs
Published
2026-10-26
The short answer: on CustomGPT.ai's Standard plan at $99 a month, a website agent with Lead Capture switched on answers 333 visitor messages, not 1,000, because each message it scans costs 3 credits there (1 for the answer plus 2 for the scan), against 2 on Premium (pricing page, actions cost, checked on 26 October 2026). The job: index your sitemap, write the persona and the fallback line, switch on citations, turn on Lead Capture with UTM tags, test with real questions, route each new lead through Zapier to a person who checks it, embed the live chat, and read the analytics every week.
Disclosure: I may earn a commission through some links on this site. It does not change what this page says.
Settle three things before step 1:
- The plan, from your traffic. For example, 400 visitors a month sending 3 messages each is 400 x 3 = 1,200 messages. With Lead Capture on Standard that is 1,200 x 3 = 3,600 credits, against 1,000 included; on Premium it is 1,200 x 2 = 2,400 of 5,000 (limits article). Either Premium, or Standard with Lead Capture off.
- The material you already own. Your sitemap, the pricing and FAQ pages, and twenty questions copied from the sales inbox, word for word. The questions become the test set in step 5.
- The baseline. Write down how many demo requests your form brought in last month, by campaign. Thirty days after launch, set the agent's leads by utm_source beside it.
What the product includes and costs, plan by plan, is in the CustomGPT.ai glossary entry.
Before you start
- A trial account. Every plan opens with 7 free days, but the checkout takes payment details and its Start trial button activates the subscription (subscription guide).
- Edit access to your site's HTML, to paste the chat script.
- A Zapier account, for the hand-off to your CRM or a sheet.
- A sentence in your privacy notice. The agent will collect names and email addresses in chat. Article 13 of the GDPR says that at the moment you collect personal data you tell the person why, on what legal basis and for how long you keep it. Standard deletes conversations after 12 months, and Premium can set a custom period (retention guide), so the period is yours to state.
The build, step by step
1. Index the pages buyers read
On the dashboard, click New Agent, choose Website, paste your sitemap into "Enter a URL or a sitemap" and click Create Agent; CustomGPT.ai finds and indexes the pages (creation guide). Each page counts as one document. Standard allows 5,000 per agent and 10,000 indexed a month across the account, per the limits article, so a blog archive of old posts can crowd out the pages that sell.
2. Give it a persona and a way out
Under Personalize, then Agent Persona, describe the tone and the job in plain instructions, save, and try typical prompts (persona guide). One agent carries one persona, so a support assistant and a pre-sales assistant are two agents against your agent limit.
Then open Personalize, Citation, and rewrite the "I don't know the answer" message (fallback guide). A useful version names the next step: the demo form, or the address of a person.
3. Show the sources
In the same Citation tab, toggle Enable Citations, then choose numbered references inside the answer or a list after it, with a header and a collapsed or expanded start (citations guide). For pricing and contract questions, a visible source is what lets a buyer trust the answer without phoning you.
4. Switch on Lead Capture and the campaign tags
From the three-dot menu beside the agent, open Actions, toggle Lead Capture and confirm with I understand (Lead Capture guide). Under Configure, keep the fields your CRM uses. Name, Email, Phone number and Company are the defaults; you may add up to 10 custom fields, with no more than 8 active at once (field guide). Fewer fields mean fewer questions in the chat.
In the same settings, toggle Collect UTM tags. The agent then reads utm_source, utm_medium, utm_campaign, utm_content, utm_term, gclid, fbclid, msclkid and three referral tags from the visitor's browser URL, never asks the visitor for them, and stores them with the lead (UTM guide). On its own, without Drive Conversions, the agent asks for contact details within the first few messages and then carries on answering (setup note).
5. Test with the twenty questions
Open Deploy, pick Live Chat and press the Try it out button to talk to the agent as a visitor would (preview guide). Ask the twenty inbox questions and mark each answer right, wrong or missing. On the answers about price, terms or security, click the shield icon: Verify Responses extracts each claim and shows the source text behind it, and visitors never see it (Verify Responses). Keep it on demand. Always-on verification costs 9 extra credits per answer on Standard and 4 on Premium (actions cost).
6. Let the agent file leads, and a person accept them
Here the agent does the work: Lead Capture reads every visitor message, decides what is contact data and writes it to the lead record. In Zapier, create a Zap with CustomGPT.ai as the trigger app and New Lead as the event, sign in by OAuth, choose the agent and test the trigger (Zapier guide). The payload carries name, email, company, phone, custom fields, the tags and a session ID.
Send it to a review sheet or a CRM task, not straight into a sequence. The export guide explains why: each new detail is saved as a separate record, so one conversation can arrive as three rows, name, then email, then company. The human check is a person who merges rows by session, rejects the test chats, and only then lets the contact into nurture.
7. Put it on the site
Open Deploy, click Make Public, go to Live Chat, set the icon position, size and window style, then paste the HTML script into your site (live chat guide). Start on the pricing page and the pages your campaigns land on, where the UTM tags are.
8. Read the analytics every week
Agent Analytics shows Latest Missing Content, queries per conversation, thumbs up and down, and user locations (monitoring guide). Missing content is the to-do list for your website, not only for the agent.
What the docs warn about
- Limits stop the account, not one agent. When a monthly limit runs out, the account "may not function normally until the next cycle", per the limits article, which also warns that auto-sync can burn through indexing and word limits fast.
- Old conversations go for good. On Standard, conversations older than 12 months are deleted and cannot be restored (retention guide). Export leads monthly.
- Verification scores are a guide. The vendor calls them AI-generated and says to treat low scores as a place to start reviewing (Verify Responses).
Where the credits go
Lead Capture bills a scan on every visitor message, including messages with nothing new to record, and a Standard account pays two credits per scan where Premium pays one (lead usage guide). The setting that makes the agent useful to marketing is therefore also the one that moves a Standard account toward Premium, or toward the extra-queries add-on, soonest.
So before you pay for a month, do the free half. Take the twenty inbox questions and check whether your website answers each one in writing. Every gap you close there is a page the agent can cite later, and a page that helps the visitor who never opens the chat. If the answers exist and the questions keep coming, the agent has a job; if they do not exist, it will mostly say "I don't know".
FAQ
Why does one conversation show up as several leads?
Because Lead Capture saves each new detail as its own record: a name in one message, the email in the next (export guide). Rows from one visitor share the same project and query identifiers in the CSV, and Zapier sends a session ID with each lead.
Will it record which ad brought the visitor?
Only if the tag is in the browser URL when the visitor chats. The agent collects the listed UTM and click-ID parameters from the URL and does not ask for them (UTM guide), so campaign links need UTM parameters on the landing page itself.
Do visitors see the verification panel?
No. Verify Responses is a builder-only tool; visitors get the normal chat (Verify Responses).
ABOUT THE AUTHOR
I am Joris van Huët, an enterprise interim CMO. 15+ years, 50+ organizations, including ING, P&G, Nestlé, BNP Paribas, WeTransfer and Vinted. I take on innovation projects (venture building, design sprints), agentic marketing (setting up and running AI agents) and hands-on multi-channel work. See the track record.
I wrote and published this with AI assistance, and I answer for it. Claims about my own experience are limited to the track record above, and a statistic links to its source or is labelled as an example. I sell interim and fractional CMO work, which is why this site exists. How this site is written.
A senior operator doing the work: campaigns, channels, CRM and AI agents, live from day one.