← Back to Blog

AI Sales Assistant for Marketing Sites: What Changed

An AI sales assistant for marketing sites is a pre-login chat widget that answers prospect questions from a curated knowledge base, works out who's on the other end, and hands a qualified lead to your CRM instead of dropping a transcript into a shared inbox. You install it with a script tag, one line in your site's HTML, with no SDK to import and version. It stays fenced to content you approved, and captures contact details on the fields you configure. That fencing is the part worth your attention: the assistant isn't improvising answers about your pricing, and it isn't guessing who's reading.

Last updated: September 2026

The part that actually changed

Marketing-site chat has existed for a decade and most of it is bad. You know the pattern. A bubble slides up, asks whether you have any questions, then either pushes you toward a form or answers something adjacent to what you asked. I've closed more of those widgets than I've ever used.

Two things moved recently, and they moved together.

The first is fencing. External Chat mode replies from a curated knowledge base only, with guardrails, so the assistant declines a question instead of inventing a plausible answer about your roadmap. The second is that the same script tag does double duty: pre-login it's this guardrailed sales assistant, and the moment someone signs in, it switches to a screen-aware mode that knows the current page and route and can act on it — no separate tool, no second integration.

Fencing is what makes it safe to put in front of strangers. The dual-mode part is what stops you from buying and maintaining two chat products. You need both, and most tools have shipped neither.

What a pre-login assistant collects that a form can't

A form collects what the visitor chooses to type. An assistant collects the conversation and the context wrapped around it.

On Total Chat's External Chat mode, that context includes:

• Digital fingerprinting so the session is recognized as new or returning without asking the visitor to identify themselves.
• UTM attribution so the campaign, source and content that brought someone in travel with the lead.
• IP-based visitor de-identification returning company, location and industry, inferred from the network the visitor is on.
• Email-based contact enrichment so once you have an address, name, title and company details come back attached to it.
• Configurable lead capture so you choose which fields matter, and when the assistant asks for them.

Enrichment is built in through EM MCP rather than bolted on from a separate data vendor. That matters for a boring reason: one fewer contract, one fewer integration to keep alive at renewal time.

And be clear-eyed about the IP piece. It tells you the company, not the person. Vendors blur that line constantly, and it's worth knowing which side of it you're buying.

Getting the lead somewhere it can be worked

A lead sitting in a dashboard isn't a lead. It's a note.

On the Pro tier, delivery fires the moment a lead is captured. HubSpot, Salesforce and GoHighLevel receive natively formatted payloads; anything else, including Pipedrive, Zoho and Freshsales, goes through the Zapier or Make universal webhook relay. There's automatic retry with backoff and a delivery log behind it, which reads like plumbing right up until a webhook endpoint starts throwing 500s overnight and you need to know whether the lead was ever sent.

Check that log in your first week. Not because I expect it to break, but because silent webhook drops end with a sales team quietly concluding the chat widget doesn't produce anything, and that's a much harder problem to unwind than a bad URL.

If you want the mechanics of scoring those leads once they land, that's a separate discipline, covered in chatbot lead qualification and scoring.

The setup order I'd actually use

1. Put the script tag on the marketing site. No framework work, no build step, no package to install. The free tier provisions immediately, and an API key and dashboard login arrive by email within minutes with no sales call in between.
2. Feed it your existing site instead of writing articles. Point the zero-code URL crawler at your live pages; it crawls them and generates 3 to 8 knowledge base article drafts for you to approve. Editing drafts beats staring at an empty text field.
3. Set capture fields deliberately. Asking for a phone number in the first message is how you end a conversation. Decide what you genuinely need in order to route the lead, and ask for nothing else.
4. Wire the CRM webhook and send yourself a test lead. Then open the delivery log and confirm it arrived.

Tuning after that happens in the business owner dashboard, live since June 2026, which holds knowledge base curation, conversation analytics, lead management, integration config and webhook delivery logs in one place. Setup specifics live in the docs.

Where this approach falls down

Curated-knowledge-base-only is a tradeoff, not a free win.

An assistant fenced to approved content will sometimes decline a question it could plausibly have handled, and a visitor who wanted a quick answer gets an offer to connect them with someone instead. If your knowledge base is thin, that fence is the only thing your visitors will experience, and a chat that keeps saying “let me find out” converts worse than no chat at all.

The fix isn't clever. The dashboard's top-questions view surfaces what visitors are actually typing, ranked by frequency, so you know exactly what to write next. Re-running the URL crawler against updated pages pulls in fresh drafts for approval too. Neither closes the gap on its own — somebody still has to read the list and act on it.

Fair warning: this needs a few weeks of real conversation volume before that list is worth much. Plan for a thin start.

Marketing site and product are two different jobs

A sales assistant and a support assistant answer to different people, and they should be configured differently.

Pre-login, you're talking to strangers who may be competitors, job applicants or bots, so guardrails and curated answers are the right default. Post-login, you're talking to paying customers inside the product, where the assistant can be far more permissive because it knows who the user is and which page they're on. There it can take someone to a specific page and call out the one element they need, in a single step. Same script tag, two modes, different risk profiles.

Most teams get this wrong by running one configuration everywhere. If you're weighing up the pre-login side specifically, the broader case sits in ai lead generation chatbot for website, and the post-login, screen-aware side is covered in in app ai support assistant.

Frequently asked questions

Is an AI sales assistant different from a chatbot?

In practice, yes. A chatbot follows scripted branches you author in advance and breaks the moment a visitor phrases something unexpectedly. A sales assistant answers from a knowledge base and captures and enriches a lead as part of the conversation rather than handing off to a separate form. The difference shows up most on off-script questions.

Do I need developers to install it on a marketing site?

No. It ships as a script tag you drop into your site's HTML, so anyone who can edit a template or add a snippet through a tag manager can install it. There's no SDK to import into a codebase, no build pipeline to change and no framework requirement. Marketing sites built on static generators or hosted site builders are the easy case here.

Will it invent answers about our pricing or roadmap?

That's what the pre-login fencing is for. External Chat mode answers from a curated knowledge base only, with guardrails, so questions outside approved content get declined rather than improvised. The cost of that design is the opposite failure: it can decline something it could have answered. Keeping the knowledge base current is what shifts the balance.

What happens to a lead once the conversation ends?

It gets enriched and delivered. IP-based de-identification adds company, location and industry; an email address triggers contact enrichment for name, title and company details. On the Pro tier a webhook fires immediately to HubSpot, Salesforce or GoHighLevel with a native payload, or to any other CRM through a Zapier or Make relay, with automatic retry and a delivery log.

---