AI chat for B2B websites, where there is no cart

With nothing to add to a cart, a good conversation has to be defined some other way. It is a qualified named person, produced quickly, or a clean no.

There is no product grid on your site and nothing to add to a cart, so none of the usual ways of telling a good chat conversation from a bad one apply. The only outcome worth counting is a named person, with a company and a way to reach them, who wants to hear from your sales team and still wants it when the call comes. Everything else the assistant does is either in service of producing that person, or in service of establishing in the first minute that this visitor is never going to be that person. The second half is the one that decides whether your team's week is worth anything, and it is the half almost nobody writes down before launch.

The named person replaces the cart

A conversation that ends with "we do not work with companies under about fifty people, so we are probably not the right fit" is a successful conversation. Nothing was captured, nobody was booked, and that is the point: a sales rep did not spend twenty-five minutes on a call to reach the same conclusion. If you measure the assistant only on conversations captured, you have built something that is paid to keep people talking. An assistant that never says "we are not right for you" is not a lead engine. It is a tax on your sales team, collected one bad meeting at a time.

So keep two counts, not one. Leads captured is the obvious one. The second is conversations where the assistant ruled itself out on a stated reason. If that second number is zero after a month of real traffic, either every visitor is a perfect fit, which is not true, or the assistant is dodging the questions that would have ended things honestly. Both counts come out of reading transcripts, which on a B2B site is a job of maybe twenty minutes a week, because the volumes are small enough that you can actually read them.

The two-minute clock after intent appears

Intent on a B2B site is loud when it arrives. Someone asks what a rollout involves, or who they would be working with, or how to get a quote. The gap between "this looks right" and "I will look at this next quarter" is short, and it closes on its own while nobody is doing anything wrong. A calendar reminder fires. A colleague walks over. The tab stays open for six days and then does not.

Which means every extra turn after intent appears is a turn spent losing it. The moment the visitor signals they want a person, the assistant should stop teaching and start collecting: who you are, what company, what you are trying to do, and where to reach you. Two more well-written answers before the ask is two more chances for the moment to pass. The other half of that clock runs on your side, and it is the half designed last: who gets alerted, how quickly they are expected to move, what happens on a Friday afternoon. Those belong to lead handling and are settled there. The only part specific to B2B is the size of the penalty. In a shop a slow reply costs one order. Here it costs a deal that was, for about ten minutes on a Tuesday, at the top of somebody's list, and will not be again.

Your solution pages are the knowledge, in the wrong shape

Almost everything the assistant needs is already written. It is sitting in the solution pages, the case studies, the implementation overview, the careers page that accidentally explains what you actually do. The problem is that all of it was written as prose to be browsed. Nine hundred words about your approach to onboarding is a good page and a bad answer. The work is cutting it into short specifics: which industries, which regions, minimum volume, which systems you integrate with by name, how many weeks a typical implementation takes, what the first call covers.

How that first draft gets written — a crawler reading your public pages, or a drafting assistant you talk to and correct as it files things — is a general setup question with the same trade-offs here as anywhere. What is not general is the shape you are cutting toward. A good B2B card set is mostly short declaratives with a number or a proper noun in them, because the buyer's real question is nearly always a comparison against a constraint they already have: is this available where I am, does it work with what I already run, does it start above what I can spend. Prose describing an approach cannot be compared against anything. A line naming the systems you connect to can.

Every drafted card gets read by a human before it goes live; the setup page makes that case in general and puts an hour on it. What changes on a B2B site is what you are reading for, which is not typos and not tone. It is a card that states the offer wrongly: a plan name you retired, a market you exited last year, a "starting from" figure that stopped being true. Your site is full of those and a crawler reproduces them faithfully, after which the assistant repeats them to a buyer with no signal that anything is old. Twenty to thirty cards is under an hour of reading, and it is the one step you cannot hand to somebody who does not sell your product, because nobody else in the building knows which sentences expired.

One property of this setup is worth knowing before you write anything. The assistant answers from your imported material and the cards you wrote, not from general knowledge about your industry. On a B2B site that is what you want. A gap in the cards shows up as a question it cannot answer, which you can see in the transcript and fix that afternoon, rather than as a confident paragraph about a capability you do not have. It also means the assistant is exactly as accurate as those twenty or thirty cards, which is the real argument for reading them.

Write the four disqualifiers down before you launch

B2B buyers are trying to rule you out quickly, and they do it with roughly the same four questions everywhere. Before you launch, write an honest allowed answer for each one and put it in the assistant's knowledge.

The temptation is to write these vaguely so nobody leaves. Resist it. "We do not serve that market" in the first minute is a feature. The visitor gets their afternoon back, your team never sees the lead, and you have not spent a demo slot discovering something a sentence could have covered. Vague answers do not save the deal. They move the disqualification to a call, three weeks later, with two people on it.

Then test the four before anyone else sees them. Open a shareable test link and ask the questions in the blunt form a stranger uses: do you work with companies our size, we are in Brazil, is that a problem, do you connect to NetSuite. Send the same link to whoever owns pricing and let them try to break it. Ask a few of those questions in the languages your buyers write in, because the assistant replies in the language it is asked in, and a limit that only exists in your English cards still has to hold when the question arrives in Spanish.

Price, the question you cannot dodge

Most B2B sites do not publish price, and almost every visitor asks. If you have not decided what the assistant may say, it will answer from whatever is nearest in your cards, and that is worse than any of the alternatives. Decide which of three shapes you want: a band, a starting point, or an honest "it depends, and here is what it depends on".

The third one is fine, but only if it comes with the actual dependencies. "It depends on seat count, data volume, and whether you need the on-premise deployment" is a real answer that lets a buyer estimate. "It depends on your requirements" is a refusal wearing a helpful voice. Write down the version you are comfortable with, then check it against what your reps say on calls, because an assistant quoting a number your team stopped using is a problem you find out about late. Silence on price gets read as expensive. Evasion gets read as expensive and slippery.

Keep the number in one card rather than scattered through several. When it changes you edit one place, and the per-message traces tell you which card produced a given answer, so you can check after the fact whether the old figure went out to anyone during the week you forgot.

The same company will ask three times

B2B rarely has one buyer. An analyst asks on a Tuesday whether you connect to the warehouse system. Someone more senior asks the following week what implementation looks like and who does the work. Two weeks after that a third person asks about contract length and where the data lives. Three conversations, three concerns, one company, and each person arrives with no memory of the others.

Two consequences. First, what you capture has to include the company or the work email domain, not just a first name, because the domain is what lets your team see three conversations as one deal instead of three cold leads. Second, someone on your side has to look. Reading transcripts once a week tells you which accounts have a committee forming, and which questions keep coming back in the second conversation because the first one answered them badly. That second signal is the useful one: a question a company asks twice, a week apart, through two different people, is a card that reads as an answer and is not.

A good share of your visitors are not buyers

A B2B chat window gets a mix no storefront ever sees. Alongside the buyers there are competitors doing research, students writing something, candidates who arrived through the careers page, and vendors who want to sell you something and have worked out that your chat gets answered faster than your contact form. Most of that is harmless traffic. One part of it is worth deciding about before launch rather than after.

The competitor question is the one owners raise, and it is usually handled badly, by trying to make the assistant work out who it is talking to. It cannot, and it should not attempt to, because the cost of a false positive is refusing a real buyer in the one minute they were interested. The workable rule is about content rather than detection: anything you would publish on a public page can be said to anybody who asks, and anything you would not publish should not be in the cards at all. If a card contains your margin structure, your churn figure, your largest account by name or the price you gave a specific customer last quarter, the problem is not who is reading the chat window. The problem is that the fact is loaded into a system whose job is answering strangers.

Once that line holds, the rest sorts itself out. A competitor asking about your integrations learns what your website already told them, slightly faster, and the twenty minutes they spent in a chat window is twenty minutes nobody on your team spent on a call with them. That is not a loss worth designing around.

What it should refuse outright

Some questions should end the assistant's turn immediately and put a human in the seat: contract terms, redlines, security questionnaires, data processing agreements, procurement paperwork, anything legal, anything that ends in a signature. Not because a model cannot produce fluent text on those subjects. Because fluent text about your liability cap, your subprocessors or your indemnities is a statement your company may be held to.

The line worth drawing is between stating a fact and negotiating one. "Yes, we sign DPAs, and the person who handles that is on our security team" is a fact you can safely write into the assistant's knowledge. Anything past that is a handoff, phrased as help rather than as a wall: take the details, say who will follow up, get out of the way. Write that handoff line yourself, because on a procurement question the wording is the last thing the buyer reads before deciding whether this company is organized.

Where this is a small lever, or the wrong one

If your pipeline starts with a referral and a phone call, chat on the website is not going to change your quarter. A site getting two hundred visits a month, mostly from people who already know your name, does not have enough top of funnel for an assistant to sort. The same goes if your sales team cannot follow up within a day: a fast qualified lead that sits for a week is a slow lead with extra steps, and you have added a system to maintain for nothing. And if nobody at your company will sit down and write the four disqualifier answers, do not launch. An assistant with vague answers to the four questions that matter is worse than no assistant, because it produces leads that look real.

This helps a specific shape of B2B site. High enough traffic that reading every conversation yourself is impractical. A wide top of funnel where a real share of visitors are not a fit, so sorting has value. Buyers who do their reading outside the hours anyone on your team is awake, which on a site with international traffic is most of the reading. Long pages that answer the question somewhere in paragraph nine. If three of those four describe you, an assistant earns its place. If none do, spend the effort on the pages instead, and set business hours with an honest away message rather than pretending to be staffed.

Common questions

What should an AI chat assistant do on a B2B site with no products to search?

Produce a qualified named person for your sales team, or rule the visitor out quickly. With no catalog and no cart, the only outcome worth measuring is a named contact with a company and a way to reach them, captured while they still want the call. The second valuable outcome is an honest no: a visitor told in the first minute that you do not serve their region, size or price band is a conversation that saved a demo slot.

Should the assistant give a price if we do not publish prices?

It should say something, because silence on price is read as expensive. Pick one of three shapes and write it into the assistant's knowledge: a band, a starting point, or an honest "it depends" that names the actual dependencies, such as seat count, volume or deployment type. What you should not do is leave it undecided, because then the assistant answers from whatever is nearest in your cards. Check whatever you write against what your reps currently quote on calls.

How does the assistant learn what we sell if it is all in long solution pages?

Two routes, and choosing between them is a general setup question rather than a B2B one: a crawler can read your public pages into a first draft, or a drafting assistant you talk to can write the cards from what you know that the site never said. What is specific to B2B is the shape you are cutting toward — short declaratives carrying a number or a proper noun, not paragraphs describing an approach, because a buyer is comparing you against a constraint rather than reading you. Review the draft before launch either way, since the failure that matters is a card stating your offer wrongly and then repeating it to a buyer.

Should an AI assistant answer contract or security questionnaire questions?

No. Contract terms, redlines, security questionnaires, data processing agreements, procurement and anything ending in a signature should trigger an immediate handoff to a person. The assistant can state a fact, such as confirming that you sign DPAs and naming the team that handles them, but it should never negotiate one. Fluent text about your liability cap or subprocessors is text your company may be held to.

Is AI chat worth it for a low-traffic B2B website?

Often not. If your deals start with a referral and a phone call, and the site gets a couple of hundred visits a month from people who already know you, there is not enough top of funnel for an assistant to sort. It is also the wrong tool if nobody can follow up on a lead within a day. It pays off on B2B sites with real traffic, a wide top of funnel where many visitors are not a fit, and buyers self-educating outside your business hours.