OpenAI's Sponsored Agents test lets a person start a labeled conversation from a ChatGPT ad, then follow a link to the advertiser's site. The format is in a limited US test with selected advertisers. Its useful lesson for any business running ads is simpler: the question behind the click matters as much as the click.
You can apply that lesson on a normal website today. Put a chatbot beside a landing page that answers the exact questions raised by the ad, then measure whether the conversation produces a qualified lead. Use this paid-traffic launch check before spending more:
- Write one ad promise and one matching page headline. Check both on a phone.
- Give the chatbot verified answers to the five questions a buyer will ask next.
- Let visitors get an answer before requesting contact details.
- Preserve campaign tags at landing and connect them to chat and lead events where your setup permits it.
- Review transcripts and downstream lead outcomes, not just widget opens.
Use the first four steps to prepare a small campaign test. The fifth determines whether the campaign deserves a larger budget.
Make the ad promise answerable
Imagine an ad for a home renovation company that says, "Get an estimate for a kitchen remodel in Tallinn." A page headed "Quality construction services" breaks the promise before the visitor reaches the chatbot. The page should say what kind of estimate is available, what information the company needs, and when a person will respond. The chatbot can then answer questions about scope, timing, materials, and service area.
Write the ad, headline, first chatbot invitation, and expected answer side by side. If the ad says "same-day estimate," the site and chatbot need a real same-day process. If the team only offers a callback within two working days, change the ad. A conversational interface can make a mismatch feel worse because the visitor can ask about it directly.
Keep the page useful without chat. Some people will read the price range or service terms and leave; others will use a form or call. A chatbot is an additional route for specific questions, not a reason to hide basic facts. If the landing page has several offers, give each ad group a destination whose visible copy matches its promise. One generic bot instruction cannot repair a vague page.
For a product campaign, write down the exact facts that must be current: price and currency, eligibility, exclusions, delivery region, stock or availability, and the offer end date. Store an owner and review date with each fact. Train the bot on the relevant page or source material, and use controlled Q&A pairs for high-stakes wording. Our website training guide covers how to combine those source types. Test a question that the offer does not answer; the bot should say what is unknown and show the next step.
Build the first conversation around the ad
A paid visitor often has one unresolved objection. Give them an opening that helps with it: "Ask about kitchen remodel estimates, timing, or service areas." Avoid a generic "How can I help?" when the campaign already tells you what brought them here. Also avoid pretending to know the visitor's identity or private intent from an ad click.
Here is a realistic exchange for the renovation campaign:
Visitor: Your ad says estimates are free. Do you come to Tartu?
Chatbot: We offer free initial estimates for projects in Tallinn and Harju County. I don't have a confirmed service area for Tartu. Would you like me to pass your location and project details to the team so they can check?
Visitor: Yes. It is a small kitchen, about 12 square meters.
Chatbot: Please share an email address if you want a reply. I will include Tartu, the kitchen size, and your service-area question.
That exchange answers a real question, states the limit of the source, and asks for contact information only after the visitor chooses follow-up. The submitted lead should carry the question and stated project details, subject to your consent and retention policy. A silent widget open is no lead. An email address without a serviceable location is a contact, but it may not meet the campaign's qualification rule.
Define that rule before launch. For this example, a qualified lead might be a unique person who requests a quote, supplies a reachable contact method, describes a kitchen project, and has an address in the confirmed service area. A Tartu inquiry stays in a separate "service area to confirm" bucket until a human checks it. Do not let a model guess a location or silently convert an uncertain inquiry into a sales-ready lead.
The lead generation guide explains broader capture and qualification flows. Paid campaigns add a constraint: every extra question consumes traffic you paid to acquire. Start with the minimum details needed for a useful next action. Ask about budget only if a person will use the answer to route the inquiry; do not turn the chat into a long survey.
Keep campaign context without collecting the conversation
Google Analytics recommends consistent utm_source, utm_medium, and utm_campaign parameters on custom campaign URLs. Choose stable values in advance, such as utm_source=google, utm_medium=cpc, and utm_campaign=kitchen_estimate_tallinn. Add utm_content when two creative variants need separate reporting. Keep these fields free of names, email addresses, search text that identifies a person, or other personal information; Google's PII guidance warns against placing it in campaign parameters.
The tagged landing visit, chatbot start, lead submission, and eventual sales result are different events. Your website analytics or CRM needs an approved way to relate them. Do not assume a chatbot platform automatically stores UTMs with each conversation or exports them to a CRM. If that link is unavailable, report page-level campaign performance and manually review a sample of conversation records. Mark the attribution gap rather than inventing a precise lead source.
Set a short, written attribution rule before comparing campaigns: which landing session counts, how long a later lead may be associated with it, what happens if the person returns through another channel, and whether a duplicate contact counts again. Be explicit about cases you cannot join across devices or consent states. Store a campaign identifier rather than a raw URL when possible. Limit who can inspect transcripts and remove personal details from analytics events.
The following event contract is a useful handoff between marketing, the site owner, and whoever manages the CRM. Event names are a proposed implementation, not events Agentkit emits by default. Google Analytics documents custom events and a recommended generate_lead event that you can use when a lead is actually created.
| Event | Fire when | Required fields | Do not count as |
|---|---|---|---|
landing_view | Tagged page loads after consent rules allow measurement | Campaign ID, landing variant, session ID | A unique person or an ad click |
chat_start | Visitor sends the first message | Session ID, landing variant | A qualified lead |
generate_lead | Contact submission succeeds | Lead ID, session ID, campaign ID if available | A closed sale |
lead_qualified | A human or recorded rule confirms fit | Lead ID, rule version, reviewer or decision source | Every submitted contact |
sale_confirmed | CRM records a completed purchase or signed contract | Lead ID, value if permitted, confirmation date | A booked meeting |
Use one identifier per deduplicated lead. If a visitor retries a form, the second submission should update or retry the first lead, not create a new conversion. For events sent to analytics, keep the transcript and email address out of event parameters. A product analytics event can say that a qualified conversation occurred without exporting the conversation itself.
Calculate what the campaign bought
Suppose a hypothetical campaign spends $900 and the ad platform reports 1,200 clicks. Your consent-aware site analytics records 1,100 eligible tagged landing sessions. From those sessions, 280 visitors start a chat. Thirty-eight unique contacts submit details; after reviewing duplicates and the service-area rule, 24 qualify. Six attend an estimate call and two become customers within the reporting window.
The chatbot start rate among observed landing sessions is 280 ÷ 1,100 = 25.5%. The qualified-lead rate among chat starters is 24 ÷ 280 = 8.6%. Cost per qualified lead is $900 ÷ 24 = $37.50, and observed cost per customer is $900 ÷ 2 = $450. Show the counts beside each percentage. The 100-click difference between the ad platform and site sessions is a measurement gap to investigate, not 100 proven lost visitors. Refunds, later sales, and repeat contacts could change the customer figure.
Now suppose a second ad variant creates 350 chat starts but only 15 qualified leads on the same spend and comparable traffic. Its chat-start rate looks stronger, yet its cost per qualified lead is $60. Read a transcript sample before changing the page: perhaps the creative attracts people outside the service area, or perhaps the bot asks for an email before explaining the estimate. Both lead to different fixes. Our KPI definitions make the difference between engagement and confirmed outcomes explicit.
For a cleaner comparison, split comparable traffic between a page with the chatbot and a page with the same offer and a conventional contact route. Keep the ad promise, location targeting, price information, and follow-up speed as similar as possible. Track qualified leads and later outcomes for both. If the sample is small, report the counts and uncertainty; do not call a few extra chats a conversion lift. Review whether either path excluded visitors who prefer a phone number or a form.
Review the expensive failures first
Spend your weekly review on conversations that could change the campaign decision. Read the chat from a paid visitor who asked about an offer, a qualified lead who never reached a salesperson, a contact marked unqualified, and a session that ended after an unclear answer. Check each against the landing promise and the source material the bot used. Correct an outdated price or missing service-area page at its source before changing the prompt.
Test the paths that often break after launch: a mobile visitor with the keyboard open; a campaign URL that redirects and drops UTMs; a form submitted twice; a visitor who declines tracking; a chat that cannot answer an exclusion question; and a handoff outside business hours. Confirm that the CRM receives the lead once and that someone can act on the context. If you cannot complete an end-to-end test, pause the claim that the funnel is measurable.
When the bot answers correctly but qualified leads remain low, inspect the ad's promise and audience before adding more chatbot questions. When many qualified contacts fail to become meetings, inspect the human follow-up and the availability you advertised. Conversation logs are valuable here because they preserve the exact objection the landing page did not settle. They still need a recorded review decision and a downstream sales outcome to explain performance.
A useful chatbot landing page keeps the promise made by the ad, gives a buyer a fast answer, and makes the next step clear. The campaign earns a larger budget only when the resulting leads meet a defined qualification rule and reach the promised follow-up. In Agentkit, you can train a website chatbot on source pages and Q&A pairs, collect leads, and review conversation logs; connect campaign and sales outcomes through your own approved measurement setup.
Build your chatbot for free → No credit card required.
