The chat vendors caused their own demise. Five years on, nobody has made the decision they skipped.

If you have bought chat software, or you are about to, there is a second decision sitting inside that purchase and almost nobody makes it. Five years of this data says that second decision is the one that separates the sites a buyer can get an answer from, and it has nothing to do with the vendor.

Live chat began as a staffed channel. A person sat with the queue open and answered whoever turned up. That was the whole product, and it worked because a buyer with a question got the question answered by somebody who knew the answer.

Then the category got sold, and what was being sold quietly changed. A chat vendor today sells an install. A snippet, a playbook, routing rules, a case study from a company two sizes up. And a demo where the widget catches a visitor at eleven at night and books a meeting while everyone is asleep.

The argument running underneath all of it is that the software is the answer. For one question it is. Nobody needs to be awake to put a meeting on a calendar.

The question nobody in that meeting gets asked is what happens to the buyer in the meantime. They typed a question at eleven at night. A slot on Thursday is a date, not an answer, and the pitch is built so that the difference never has to come up.

There was a second piece of advice attached to the pitch, and it did more damage than the first. Put a gate in front of the conversation. Ask for an email address before the chat begins, and at least you capture the lead when nobody is available.

It was sold as qualification, on the reasoning that a serious buyer will hand over an address and a time waster will not.

You can read the argument off the contents page of the category's own canonical text without opening it. Conversational Marketing, by David Cancel and Dave Gerhardt, Wiley, 2019.

Chapter 4 replaces lead capture forms with conversations. Chapter 6 adds messaging to the website and starts capturing more leads. Chapter 8 qualifies leads through conversation. Chapter 9 filters out the noise. Chapter 10 builds a lead qualification chatbot. The subtitle is the promise in one line: how the world's fastest growing companies use chatbots to generate leads 24/7/365.

Read those in order and notice what is missing. The form comes off, and the thing replacing it captures, qualifies and filters. Answering the buyer's question is not one of the steps.

The mechanic is still shipping today. Drift's own product documentation, now hosted by Salesloft, describes a lead qualification node with a qualification floor. A visitor who fails a required condition is "automatically disqualified and routed out via the 'unqualified' offramp, saving your sales team time".

So somebody arrives wanting to know whether the thing integrates with their stack. They are not ready to say what they intend to spend. They get offramped.

What a gate actually filters is anybody who wanted an answer before they were ready to be contacted, which on a high-intent page is most of them. It also removes the last real advantage chat had over a form, which was that you could ask something without identifying yourself first.

And it was an easy thing to agree to, because it made the staffing question disappear. If the box captures a lead whether or not anybody is there, nobody has to decide who is there.

The contract gets signed. The widget goes live on a Tuesday. The meeting does not reconvene.

No shift schedule. No name against the hours. No response time anybody would be uncomfortable missing. The item on the plan said "chat," chat is installed, and the item is closed.

I have done exactly this, so I am not writing from a position of having always known better. I bought the software and filed the operational question as solved. It was not solved. It was never asked.

Two decisions sit inside that purchase. The first is to install the software. The second is to put somebody behind it, on a schedule, accountable for answering. Only the second is ever experienced by a buyer on your website. You can complete the first without starting the second, and the software will keep working perfectly the whole time.

Meanwhile the person on the other side of the glass is not looking for a widget. They are looking for an answer to a question specific enough that the pricing page did not cover it.

We watched 4,265 B2B websites across five years and three waves to see how that ended. The configuration the category sells is software installed with nobody behind it. It fell from 789 of those sites in 2021 to 69 in 2025, then came back to 7.0% of the study in 2026.

The configuration it stopped talking about is a widget with a person scheduled behind it. That was on 140 sites in 2021, 111 in 2025 and about 2.3% of the study in 2026. It has not moved in five years.

Chat did not lose to AI, and it did not lose to buyers who wanted to be left alone. What was sold was an install, the staffing decision was left for later, and later has not arrived.

The findings

Base for every row above: the 4,265 matched sites in the study, in all three waves. The 2021 and 2025 figures are counts of observed states. The 2026 figures are an estimate built group by group, and their intervals are in the method section.

Judged in those base columns means the conversation gave us a clear answer. Some sites were gone and some would not let us in, so the denominator is the sites we could actually read. It is defined on the website panel page.

The study rows are website states. The response rows come from the demo study, whose fieldwork ran in 2023. They are a baseline, not a current measure of the market.

More of them deleted it than answered it

789 study sites were running a widget with nobody behind it in 2021. By 2025, 391 had removed chat entirely and 55 had put a person behind it. A further 174 had put a bot in front of it that offered to bring one. Base: the 4,265 matched sites in the study.

The study records decisions, not opinions, which is the only reason any of this is worth publishing.

Every one of those companies had already paid for the software. The install was done, the cost was sunk, and in front of each of them sat a single operational question: is anybody accountable for answering this.

560 of them answered no. 391 switched the widget off entirely. 138 moved to a bot with no human path. 31 were still sitting there unstaffed four years later.

55 put a person behind it.

174 took a third option, which is the one I missed for four years. They put a bot in front of the widget that offered to bring a person. On a one-visit classification that reads as a route to a human, and I counted it as one.

A company that had already bought chat was seven times more likely to delete it than to put a person behind it. And three times more likely to buy a bot that promised one.

To a buyer those two outcomes land the same way. A box that nobody answers and no box at all both end the conversation before it starts.

The study records what those sites did. It does not record what was said in the rooms where they decided, so the next part is my reading.

What I think happened is that the purchase felt like the decision. The staffing conversation was never scheduled, because the pitch was built so it did not have to be. Four years later somebody looked at a widget nobody had answered since the quarter it went live. They made the only call left to make.

For four years I read the rest of that movement as the market splitting into two ends, one staffing properly and one giving up. It did not. Staffed chat went from 140 sites to 111. That is 0.68 points down, with a 95% confidence interval of 1.38 down to 0.02 up, which the study cannot separate from no change.

What grew was the bot that offers to fetch somebody, from 18 sites to 490. The full state-by-state movement is on the website panel page.

In 2026 we went back and asked. The wave opened each widget, put a buyer's question to it and waited three minutes. Of the sites whose bot had offered a person in 2025, 4 of 88 judged delivered one. What the other bots did instead is in the bot handoff study.

Of the sites that were genuinely staffed in 2025, 22 of 104 judged still reached a human. Of the 104 we could judge, 33 are now running a widget with nobody behind it.

So the category did not shrink because buyers stopped wanting conversation, and it did not grow at the configuration the vendors had stopped talking about.

It shrank at the configuration they were selling and replaced it with a cheaper version of the same skipped decision. Even the companies that did make the decision mostly did not hold it. Four in five of the sites staffed in 2025 no longer reach a person.

Accountability does not hold by itself. Somebody owns it or it decays. That is what the third wave added to the first two.

We looked and could not tell

766 of the companies in the 2023 demo study had chat software installed and 919 did not. Nothing in the data says whether owning it changed a buyer's odds of being answered. Base: 1,685 usable submissions.

The obvious question is whether the group that had already bought the category was any more likely to put a person in front of the buyer. I cannot answer it from this data.

Not because the two halves came back level. Because the study was never built to separate them. It was built to measure one rate precisely, and it does that. Split that base in two and the answer disappears into the margin.

Settling the chat question would take another round of hand classification that nobody has done, including us. So the honest sentence is that we looked and could not tell, which is not the same sentence as we looked and found nothing. Anyone claiming that owning chat software makes a measurable difference to whether a buyer gets answered, in either direction, is working from a comparison this kind of study cannot support.

What is precise is the headline, and it covers both halves at once. Better than nine in ten of these companies never put a person in front of the buyer.

Speed was not the missing part. 545 of the 1,685 first replies arrived inside five minutes, 32.3%. Of the fast replies we read by hand, not one was written by a person. They were automated acknowledgments and marketing sends. Speed was purchased. Attention was not.

What landed in the buyer's inbox that fast confirmed the message had been received. It did not answer the thing they had asked.

The cohort that bought hardest had nobody there

315 of the 319 study sites running Drift in 2021 had no human reachable through chat, 98.7%. Base: 319 Drift sites in the study.

Drift was the category: the conference, the vocabulary, the playbook, the sales team that would tell you your form was costing you pipeline.

A buyer who opened the chat box on one of those sites was not going to reach a person. The company was paying for the flagship product of the category at the time.

The baseline across every site in the study running chat of any kind in 2021 was 10.3% human reachable, or 158 of 1,534 sites. So the cohort that bought the most opinionated product in the category was about one eighth as likely as the average site with chat to have anybody there to have a conversation.

I do not think that is a product failure, and I would say the same if the logo were mine. It describes what gets sold. Drift sold the argument that speed to response wins deals. A marketing team can act on that alone, with a budget line and an install.

What that argument does not do is create a person. Staffing needs a headcount conversation, a coverage schedule and an owner. It also needs someone senior enough to say that answering buyers is a job, rather than a thing people do when they have a free minute. None of that fits in a software contract.

By 2025, 164 of those 319 sites had no chat at all. More than half of the cohort that had bought hardest ended up with nothing. On 98.7% of those installs there had never been anybody behind the widget, so there was nothing there to lose.

The same question gets answered on a consumer site

Staffed chat in 2025 was on 2.85% of consumer-facing sites and 2.54% of B2B sites. Bases: 1,368 consumer-facing sites and 2,795 B2B sites, with 102 sites excluded for carrying both sector flags or neither.

For four years I published this comparison on the combined human-reachable measure, where consumer-facing sites led by about five points. That version is withdrawn.

Its segment bases cannot be reproduced from the source workbook. On the bases that can be, the combined measure puts B2B ahead in 2025, because B2B carries more of the bot handoff. What survives is the gap on staffed chat, and it is under a point in both waves.

So the honest version of this section is smaller than the one it replaces. Consumer-facing sites are very slightly more likely to have a person scheduled, and both groups drifted down together. I publish no significance test on it. A comparison that changes sign depending on which of the workbook's sector columns defines the groups is not a finding.

The explanation I am about to give is an inference, labeled as one. On these numbers it explains a gap under a point rather than a gap of five. The study records what sites do and not why.

I do not think consumer-facing teams care more about their buyers. I think they get told whether it worked. A team that staffs chat on a Monday can read the conversion difference by Wednesday. If staffing helped, the number moves inside a window short enough that the person who argued for it is still in the room.

B2B closes over quarters, through an attribution model nobody in the building genuinely believes. Staff chat in March and the deals it influenced close somewhere between November and never. They get there through a marketing-qualified lead stage, then a sales-accepted lead stage, and then an opportunity a rep will tell you came from a referral.

There is no clean way to prove the staffing did it. When a visible, quantified cost is argued against an invisible, contested benefit, the cost wins the meeting. It wins it in most quarters, at most companies, whichever side is right. Not because anyone is being unreasonable. The person defending the salary line has a number, and the person defending the coverage has a story.

So the question stops being settled by evidence and starts being settled by whoever is most confident. That is a bad way to decide anything, and it is how most B2B companies have decided this.

The obvious objection

The strongest argument against all of this is that buyers do not want to talk to you.

It is a real preference and Gartner has measured it twice from one fieldwork window in late 2025. 67% prefer a rep-free experience, on a base of 646 buyers. 70% prefer buying that is entirely digital and self-service, and 69% go to a rep to validate AI-generated insights, both on a base of 645. Reading those as largely the same people is inference rather than a cross-tabulation, and the identification tax sets out the bases and the arithmetic.

So the objection only lands if rep-free means human-free. It does not. What buyers refuse is the process. The qualification call before the answer. The sequence they get entered into for downloading a PDF. The discovery meeting that exists to fill in a CRM field. What they still do, at a rate of 69% in the May 2026 release, is find a human. Someone to sanity check something before they put their name on a decision.

They do not want to be handled. They want a question answered without becoming a lead. Those are different things, and the industry has read the first as permission to stop doing the second. Read "buyers prefer self-service" as "so nobody needs to be available" and you build exactly the experience the 93.4% figure describes. Fast automated acknowledgment, no person, no answer.

What actually has to change

If you take one thing from this: the tool is not the variable. Accountability is.

Three questions decide whether the software will do anything. Whose name is against answering, by hour, including the hours your buyers are awake and your team is not. What response time is that person accountable for, as a number somebody would be uncomfortable missing. Where does the miss show up, in a report a named person reads on a schedule.

If you cannot answer those, buying better software will not help. On the evidence of the last five years, the likely end of it is that the widget comes off the site.

If you can answer them, you may not need to buy anything. The 55 sites that put a person behind their widget did not need a different vendor. They needed a schedule with names on it. Four in five of the sites staffed in 2025 no longer reach a person. So they also needed somebody to keep owning it after the quarter it went live.

Before any of that, find out which side of the study you are on. It takes ten minutes and you need no data you do not already have.

1. Open your own site at 7pm on a phone your company has never seen, on mobile data rather than the office network.
2. Ask the chat a question a real buyer would ask, one specific enough that your pricing page does not answer it.
3. Start a timer. Note what comes back first, and whether it answered the question or asked you for something.
4. Do the same through your demo form, from an address nobody there recognizes.
5. The next morning, read both replies. Mark each one against a single test. Could this exact message have been sent, unchanged, to anybody else who did what you just did? If yes, no human answered you.
6. Write down two numbers. How long until a person, and whether a person ever came at all.

Most teams have the first number on a dashboard already. Almost nobody has the second, which is the one that decides whether the software you are paying for is doing anything.

Method

Two studies. The response figures come from 2,528 B2B companies, each above 10,000 monthly website visitors. Each was approached once during 2023 through its own demo form, from an unrecognized address. That leaves a usable base of 1,685 after 798 unsubmittable forms and 45 errors.

Those readings are a 2023 baseline, not a current measure of the market. The widget figures come from a matched panel of 4,265 websites classified in 2021 and again in 2025. Full method is on the demo response study page and the website panel page. Every figure quoted here, with its base and its year, is in the full figure set. All three waves are reported in The Buyer Wait Time Report 2026.

Two limits matter here. Email only was captured, so a company that telephoned is recorded as never having answered. And the demo study was built to measure one overall rate precisely, at a 6.6% human reply rate on 1,685 companies. It cannot resolve differences between types of site. We publish no subgroup comparison of the never-answered rate and no significance test on one.

Notes and bounds

The never-answered rate covers 1,573 of the 1,685 companies, which is 93.4%, and it carries a 95% confidence interval from 90.3% to 96.1%.

On the chat-ownership split: 766 of the 1,685 had chat software installed and 919 did not. Split that base in two and the intervals come back between 6.5 and 10.6 points wide, overlapping each other heavily. Settling the question would take roughly 880 further emails classified by hand.

Inside the 315 Drift sites with no human reachable, 295 ran a bot with nobody behind it and 20 ran a widget nobody was staffing. Four of the 319 Drift sites could reach a person, 1.3%.

The same staffed-chat gap existed in 2021, on the reproducible bases: 3.80% of consumer-facing sites against 3.11% of B2B sites.

The two Gartner releases state bases of 646 and 645 buyers on the same fieldwork.

Questions

Are you claiming the vendors killed their own market?

I am claiming the sequence, and I am reading the cause. The sequence is measured. The deployment the category sold fell from 789 sites to 69. Companies that had already paid removed it rather than put a person behind it, by seven to one. Three times as many bought a bot that promised a person instead. And the deployment the category stopped selling has not moved in five years.

The reading, that the pitch made the staffing question skippable, comes from me and from having bought chat that way myself. A study like this cannot record what a vendor said in a sales meeting. If you have a better explanation for why paying customers deleted a product they had already bought, it fits the same rows. I would like to read it.

Is this just an advert for your company?

Fair question, and test it against the piece rather than my assurance. Start with the least flattering finding in this research for a company that sells chat. 789 sites had already bought it, and 391 of them deleted it, against 55 who put a person behind it. That sits in the piece at the same size as everything else.

The second least flattering is that I published a figure for four years that counted a bot's promise of a person as a route to a person. That made this market look better than it is, and made my own argument look weaker than it was.

Both corrections are in the piece. The third is that I cannot show you that owning the software changes whether a buyer gets answered. I have said so, rather than filling the gap with a comparison the data will not carry.

The closing argument is that if you can solve accountability with your own team you should do that and buy nothing. What I will accept is that I chose which questions to research, and I chose ones where the answer suits me. Read the method, check your own site, decide from that.

We have chat and we do answer it. Does this apply to us?

Then you are in the group the study counted as staffed, about one site in forty in every wave, and the argument is not aimed at you. Check rather than assume, and check again in a year.

Of the sites this study recorded as staffed in 2025, 22 of 104 judged still reached a person in 2026. Coverage tends to be true during the hours the person who believes in it is at their desk. It tends to stop being true when that person changes roles.

Then check one more number, because it is the one nobody looks at. Something may sit in front of your chat: an email field, a name, a qualifying question. Measure how many people abandon at that step, rather than how many complete it.

Your reporting almost certainly counts the completions. In our own client work we have seen drop-off at a pre-chat ask run very high on exactly the traffic you least want to lose. That is a first-party observation from our own book of business, not a finding from this research. We will publish it separately and label it as such. Measure your own and you will not need ours.

Is bot-only a legitimate answer to this?

Bot only was 13.8% of sites in 2021, 12.0% in 2025 and 15.6% in 2026. The first step is a real fall, 1.8 points on a 95% confidence interval of 3.1 down to 0.5 down. The second step is not separable from no change.

It works when the buyer's question has a documented answer. The Gartner pairing is where it runs out. In the May 2026 release, 69% of buyers go to a human to validate what AI told them. With no human path behind the bot, that validation happens on a competitor's site or not at all.

Why does removal beating staffing prove anything about accountability?

Because every one of those outcomes was available to companies that had already paid. The software cost was sunk for all 789. What separated the ones who removed it from the ones who put a person behind it was not budget for the tool. It was whether anyone would own the answering.

Seven to one for removal, and three to one for a bot that promised a person over hiring one. What the study cannot tell you is what was said in those rooms, so the reading of why is mine, not a measurement.