There are two conversations in conversational marketing, and most budgets go to the wrong one. The first is the chat on your website, serving people who already found you. The second is the one your buyer has with an assistant before they reach you, which decides whether they find you at all. The second one is worth more and gets a fraction of the attention.
Conversation one: the assistant conversation
This is the one that sets your pipeline.
A buyer opens ChatGPT, Perplexity or Gemini and describes a problem in their own words. The assistant asks clarifying questions, narrows the requirement and returns two or three options. That exchange happens entirely outside your systems, produces no analytics, and effectively decides who gets considered.
You cannot participate in it. You can only influence what it has to work with, which means:
- Being resolvable as an entity, so a model knows who you are
- Being extractable, so your answers can be lifted cleanly
- Being corroborated, so the model is confident enough to name you
- Answering the qualifying questions, so you get matched precisely rather than vaguely
That is the whole of AI visibility work, and the sequencing argument is made properly in AI visibility strategy. What matters here is the framing: the highest value conversational marketing work is not building a conversation, it is being available to one.
The full map of what this does to the buying path is in AI-powered customer journeys.
Conversation two: the chat on your site
Not worthless. Narrower than it is sold as.
Website chat works in specific conditions:
A repeated question that navigation answers slowly. Stock, availability, opening hours, delivery timescales, whether you cover a postcode, whether a product fits a model. If the answer is a lookup, chat beats browsing.
Complex qualification. Where finding the right product genuinely requires several questions, a guided conversation beats a filter interface.
Out of hours capture. Not answering, capturing. A person who would otherwise leave.
High-consideration purchases with a clean handover. Where the value is getting somebody to a human faster, and the bot is a router rather than an answerer.
It fails, reliably, in these:
As a replacement for clear information. If the answer should be on the page, put it on the page. Hiding it behind an interaction adds friction for the visitor and removes crawlable text that AI systems could otherwise use to recommend you. That trade is worse than it looks.
As an interceptor. The widget that opens unprompted three seconds after arrival, covering the content somebody came to read.
Ungrounded. A generative bot answering from general model knowledge will eventually state a price you do not charge, a timescale you cannot meet or a service you do not offer. It said it on your behalf, in writing, to a prospect.
As a phone number replacement. For businesses whose buyers want to speak to somebody, removing the number and offering chat reduces enquiries. It is measurable and it happens often.
How to build the on-site one properly
If you are going to do it, four rules.
1. Ground it in verified content. Retrieval from your own maintained material, not free generation. The bot should be able to cite the page it took an answer from.
2. Make it refuse. “I do not have that, here is how to reach somebody who does” is a good answer. A confident wrong one is not. Most commercial harm from chatbots comes from a system that was not permitted to say it did not know.
3. Log every unanswered question. This is the most valuable output the tool produces and almost nobody reads it. It is a list of things your buyers want to know that you have not published, which is a content plan generated by your actual market rather than by a keyword tool.
4. Hand over cleanly. Named person, real timescale, context carried across. A handover that loses the conversation and starts again is worse than no bot.
The overlap: making content work in both conversations
Here is the useful part, and the reason these two topics belong on one page.
The same content quality serves both. A passage that answers one question completely, in plain language, without depending on the paragraph before it, is simultaneously the best material for a retrieval-based chatbot and the best material for an assistant deciding whether to cite you. Both systems lift fragments. Both are defeated by prose that only makes sense in sequence.
So the test is the same in both directions: take any section of your page, remove everything around it, and check whether it still answers something. If it does not, it will not be used by either system.
The technique is covered in how to write content AI models actually cite and what content AI prefers to cite.
The questions are the same too. The unanswered questions in your chat log are the questions being asked of assistants. Publishing proper answers to them serves the bot, serves the buyer, and puts material into the pool the assistant conversation draws from. One piece of work, three returns.
Voice, and where it fits
Voice assistants are conversational by definition and the requirements are stricter, because a spoken answer is one answer with no list to scan.
The practical consequences: answers need to be short enough to be spoken, phrased as complete statements rather than fragments, and attached to a clearly identified business. Local and transactional queries dominate. The detail is in voice search optimisation and speakable schema.
What to measure
Chatbot dashboards default to engagement metrics, which reward the wrong behaviour. Engagement can rise because the bot is failing and people are rephrasing.
Measure instead:
| Metric | Why |
|---|---|
| Resolution rate | Conversations where the question was actually answered |
| Escalation quality | Proportion reaching a human with context intact |
| Unanswered question log | The content plan your market wrote for you |
| Enquiries with and without chat | The only number that settles whether it helps |
| Deflection from the phone | Watch for calls falling without enquiries rising |
And for the conversation you cannot see, measure citation share on a fixed prompt set, monthly. That is the metric with pipeline attached, and the method is in how to measure AI search visibility.
The honest recommendation
If you have budget for one of the two, spend it on the assistant conversation.
A chatbot improves the experience of people who already found you. AI visibility work decides how many people find you at all. The first is an optimisation on existing demand and the second changes the size of it, and at present the second is far less contested because most competitors are still buying widgets.
If you have budget for both, do the visibility work first and the chat second, grounded in the content the visibility work produced. That ordering is not a preference, it is a dependency: the content that makes a good grounded chatbot is the same content that gets you cited, and building it once serves both.
MarGen’s pricing is published, the method is published, and the results are documented in the case studies.
Conversational Marketing AI: Common Questions
What is conversational marketing AI?
It usually means AI chat on your own website, handling questions and qualifying enquiries. That is only half the picture. The other half is the conversation your buyer has with an assistant before they reach you, which decides whether you are on the shortlist at all. The second one has more commercial effect and receives far less investment.
Do website AI chatbots actually improve conversion?
Sometimes, in narrow conditions. They work when they answer a genuine repeated question faster than navigation can, and when there is a clean handover to a person. They fail when they intercept visitors who already knew where they were going, when they cannot answer anything specific, and when they replace a phone number that was working.
Should a chatbot be trained only on my website content?
It should be grounded in your own verified content rather than answering from general model knowledge, because an ungrounded bot will confidently invent prices, timescales and capabilities. Ground it in a maintained source, make it say it does not know rather than guess, and log every unanswered question.
Does a chatbot help with AI search visibility?
No, and this is a common misunderstanding. A chatbot serves visitors who already found you. AI search visibility decides whether they find you. If anything, chat that hides answers behind an interaction reduces the crawlable text on your site, which makes those answers less available to the systems doing the recommending.
What is the biggest mistake in conversational marketing?
Optimising the conversation on your site while ignoring the one that happens before it. A business can have an excellent chat experience and never be mentioned when a buyer asks an assistant who to use, in which case the chat is serving a smaller and smaller share of the market and the metrics look fine.
How do I make my content work in a conversational context?
Write answers that survive being lifted out of the page. Each section should answer one question completely, in its own words, without depending on the paragraph above it. Conversational systems retrieve fragments, so a passage that only makes sense in sequence is a passage that cannot be used.
Should I use AI to write chatbot responses?
For drafting, yes. For live generation without grounding, no. An unrestricted generative bot on a commercial site will eventually state a price you do not charge or a capability you do not have, and that answer was given on your behalf. Retrieval from verified content with a refusal path is the safer architecture.
What should I measure on conversational tools?
Resolution rate, meaning conversations that ended with the person’s question answered, and the proportion that escalate cleanly to a human. Not engagement. A chatbot that increases engagement while resolving nothing is adding friction, and the log of unanswered questions is usually the most valuable output the tool produces.
Where to Go Next
The conversation before yours: AI-powered customer journeys · AI visibility strategy · how to appear in ChatGPT answers
Content that works in both: how to write content AI models actually cite · what content AI prefers to cite · voice search optimisation
Related: AI agents for marketing · what is AIO · what is SXO
Have it delivered: MarGen is a UK GEO agency with published pricing, a published method and documented results.