Drift's customers did not switch. They quit.

A person who opens a chat window and asks for a human is not browsing. They are deep in research, or close to a shortlist with one or two questions they want answered before they hand over their details. They are the most valuable visitor you will get that day.

This is the story of what several hundred companies did about exactly those people. It is told from the outside, with four years between the two photographs. It happens to be about Drift, because Drift is where the install base was. Nothing here turns on one vendor. The pattern belongs to the category.

The story everybody tells about Drift goes like this. A category leader defined conversational marketing, grew fast, got expensive, lost its edge to cheaper and smarter tools, and was eventually absorbed and scheduled for retirement.

The customers, in this telling, went shopping. They evaluated the alternatives, picked a new vendor, and the market redistributed itself. Every vendor with a comparison page has been writing some version of that sentence for two years.

I have been running website classification research on B2B chat since 2021, and that account is wrong in the part that matters. Not wrong about the decline. Wrong about where the customers went.

They did not go shopping. Most of them turned the lights off.

That has different consequences. If a category leader's customers migrate to competitors, the category is healthy and one vendor failed. If they abandon the category on the way out, the failure is broader than one company. The surviving vendors have been congratulating themselves on winning a share war that mostly did not happen.

There is a second thing in this data that I did not expect, and it changes what the exit means. Drift's customers were not running a conversational layer that disappointed them. Almost none of them had ever put a person behind it.

The findings

All panel figures come from a matched panel of 4,265 B2B websites classified in 2021, again in 2025 and again in 2026. The 93.4% comes from a separate 2023 study of 1,685 demo requests and is a baseline rather than a current reading. The Drift rows rest on the 319 study sites running Drift in 2021.

Judged, in that last row, means the 2026 conversation gave a clear answer. Some sites had gone and some would not let us in, so the denominator is the sites we could read rather than the sites we tried. The website panel page defines the 2026 sampling in full.

1. What we measured

We built a matched panel of 4,265 B2B websites classified in both 2021 and 2025. The same sites in both waves, so a site that removed chat was watched removing it. Sites were drawn at random from B2B technology companies each receiving more than 10,000 visitors a month. So this is mid-market and up, with enough traffic that a chat decision is a real decision.

For each site in each wave we recorded three things. Whether chat was present. Which vendor, if we could identify one. And whether a buyer could actually reach a human through it.

That last one is what most industry data leaves out. It turns out to explain everything else.

2. Almost nobody was behind Drift while they were paying for it

Of the 319 study sites running Drift in 2021, 315 had no human reachable. That is 98.7%. 295 of them ran a bot with nobody behind it and 20 ran a widget nobody was staffing. Four could put a buyer in front of a person.

Four.

Picture what that looked like from the buyer's side. A window sits in the corner of the page inviting you to talk to somebody. You open it and type your question. What comes back is a branch of a script, then another branch, then a box asking for your email address.

The invitation was answered by the same thing that issued it, and the answer to talk to us turned out to be that there was nobody to talk to.

Now the comparison that gives it meaning. Across every site in the study that had chat of any kind in 2021, 158 of 1,534, or 10.3%, could reach a human. Among Drift sites it was 1.3%, four sites of 319. Drift's install base was about one eighth as likely to have someone behind the chat as the average site running chat.

I want to be careful about what that does and does not say. It does not say Drift was worse than its competitors. Every vendor in that period sold automation as a way to have the conversation without staffing it, and buyers deployed it exactly that way. Drift is simply where we can see it clearly, because Drift is where the install base was.

Nobody was mis-sold here. The promise was a conversation you did not have to staff, and a conversation nobody staffed is exactly what 315 of these companies installed. The product did what it said on the box. A lie would be easier to fix than that.

But it reframes the question. The interesting question is no longer why these companies left the category. It is what they thought they were leaving.

3. Now there is nothing to open

By 2025, 164 of the 319 Drift sites had no chat at all. That is 51.4%.

Not a different vendor. Nothing.

Slightly more than half of the sites running the category-defining product in 2021 have no conversational layer at all four years later. Add the bot-only row and the unstaffed row, since an unstaffed widget has nobody behind it by definition. That gives 256 sites of 319, or 80.3%, with no route to a person.

A buyer landing on one of those pages today has nothing to open and nobody to ask.

Put the sites that went dark and the ones that kept a bot next to how few of them ever had anybody behind the chat. The exit stops looking like disillusionment. And nothing forced it. Salesloft bought Drift in February 2024 and did not publish an end date until 2026, setting retirement at 31 January 2027, which is still ahead of us. So the sites we found dark in 2025 switched off a product that was live, supported and inside its contract. Nobody sent them a deadline. There was nothing to preserve. The thing had never been answered.

One group looks at first like it cuts the other way. Sixty-three of the 319 sites could reach a human in 2025, and most of those had been bot-only in 2021.

I reported that as an upgrade for four years. Split into its two states, 4 of the 63 were a staffed chat with a person scheduled behind it. The other 59 were a bot offering to bring one. Ninety-four percent of the upgrade was a promise.

When we came back in 2026 and asked the bots for a person, the study-wide version of that state delivered one on 4 of 88 judged sites. So a fifth of Drift's base did not quit and did not upgrade either. It replaced a widget nobody answered with a bot that offered to fetch somebody. The transcripts from those conversations are in the bot handoff study.

The decomposition here is computed from the 2021 workbook's own chat brand field, restricted to the matched panel. That field identifies 321 cohort sites against the 319 the vendor analysis identifies. The within-cohort split does not turn on that difference, and the published cohort figures stay on the 319.

4. Fewer sites ran it, and the share story does not survive a question

Drift's absolute install count fell from 319 sites to 249.

Its share of the vendors we could identify appears to rise, because the pool it was measured against shrank faster than it did.

I am not going to lean on that. It does not survive the first question an analyst will ask: how much of the 2025 chat market could we not identify at all? Our detection coverage fell hard between the two waves, and the sites we could not attribute are the ones that decide the answer.

On the most generous reading Drift gained share while dying. On the strictest it was flat. I cannot tell you which, and neither can anyone else without vendor-side data. Both readings are published in the method below, with the coverage arithmetic behind them.

So the widely repeated line that Drift was gaining share when it was shut down may be true. Our data does not establish it.

Our data does establish two things. Drift remained the most installed vendor we could identify in both waves, while its install base fell by 70 sites. And the buyers leaving it were not going to competitors in any volume.

5. Hardly anybody left for somebody else

Of the 130 Drift sites where we could detect a named vendor in both waves, 82 were still on Drift in 2025 and 48 had moved.

Where they went:

That is a small base and the ordering is worth more than the gaps. It is at least consistent with the wider study, where Qualified grew from 14 sites to 83 and was the biggest gainer of the period.

What matters about the table is how short it is. Forty-eight sites of the 130 we could track through both waves moved to a competitor. One hundred and sixty-four of the 319 went dark. The competitive reallocation everyone wrote about is real, and it is roughly a third the size of the exit.

6. This is a category story, not a Drift story

Look only at Drift and you would conclude its customers were unusually disillusioned. They were not. They behaved like the rest of the study, and the 2026 wave says the same thing about the staffing decision that the first two did.

Across all 4,265 matched sites, the share with no chat at all rose from 64.0% in 2021 to 72.3% in 2025, and stood at 74.1% in 2026. In raw counts, 753 sites removed chat between the first two waves and 401 added it. The 2026 figure carries a 95% confidence interval of 69.1% to 79.1%. That is too wide to settle whether anything moved after 2025. The retreat across the whole five years is the subject of Five years of AI hype, and B2B websites went backwards.

Underneath that total, the composition changed more than the headline did.

Sites where a human could be reached rose from 3.7% to 14.1%, 158 sites to 601. Staffed chat did not move: 140 sites, then 111. The whole of the rise was the bot-handoff state, from 18 sites to 490. The staffed-chat change is 0.68 points down, with a 95% confidence interval of 1.38 down to 0.02 up. The study cannot separate that from no change.

Unstaffed widgets fell from 18.5% to 1.6%, 789 sites to 69. This is where most of the category's decline came from, and what left was overwhelmingly the unattended.

Faced with a choice between putting a person behind the thing and removing it, the market chose removal by seven to one. Most chose silence when the buyer asked a question. The third option, a bot that offers to fetch somebody, was taken three times as often as putting a person there. That is the same behavior as the Drift cohort, at a different scale.

Drift's base had almost nobody behind it, and then more than half of it disappeared. The study's unattended widgets were mostly switched off.

Neither group was abandoning a conversation. They were abandoning a fixture nobody had ever answered.

Vendor churn ran heavy in the same period. 257 of the 585 sites where we detected a brand in both waves had changed vendor by 2025, 43.9%. Read that as a ceiling rather than a point. The sites that moved, and who was answering after they did, are mapped in the vendor churn study.

Several established vendors we found in 2021 had no detected installs left by 2025. Disappearing from our study is not the same as disappearing from the market. I would not report those as business failures.

7. The people who use chat are the ones you can least afford to lose

A buyer who asks for a human is at the sharp end of the process, and the 315 Drift sites with nobody behind the widget were set up to miss exactly that buyer.

Gartner's buyer research comes at it from the other side. 67% prefer a rep-free experience, on a base of 646 buyers, and 69% go to a rep to validate what AI told them, on a base of 645. Taking those as largely the same people is inference rather than a cross-tabulation, and I work through the pairing in the identification tax.

The measured version of that behavior is a demo request, which is the identical act with a form attached. In a separate 2023 study of 1,685 B2B companies that received one, 93.4% never put a human in front of the buyer. That fieldwork is from 2023, so it is a baseline rather than a current reading.

So consider what the 256 sites with no route to a person are doing. These are companies above 10,000 monthly visitors. They paid to put that buyer on the page. Drift was sold to them explicitly as the way to catch high-intent visitors in the moment. Across the cohort, 98.7% of those installs had nobody behind them. Then the widget came off.

Here is where I would change the conclusion most people reach. The instinct is to say you are handing your best buyers to competitors. We did not follow the buyers, so I cannot claim that, and I think the truth is worse.

A delayed decision does not transfer. It dissolves. Tversky and Shafir, 1992 found that the tendency to defer a decision, keep searching, or take the default option rises when the offered set is enlarged or improved. Their subjects were choosing from a set, not waiting for a reply. Carrying one onto the other is mine, not theirs. On that reading, a buyer left waiting is a buyer whose option set keeps growing, and postponement gets more likely rather than less.

Losing a bake-off at least means the category made a sale and your CRM records where it went. This does not. You funded the demand, the buyer raised a hand, nobody came, and the purchase never happened for anyone.

That is the cost, and it shows up nowhere.

8. A better script had nowhere to hand off to

Bot-only sites went from 13.8% to 12.0%. Down 1.8 points, across a window that contains the entire arrival of modern language models.

The four years covered here are a natural experiment on the previous generation of automation. Companies bought scripted chat, ran it unattended, and delivered a verdict: they switched it off rather than staff it.

The obvious reading is that the bots were bad. They were designed around a talk track someone had to write in advance, they ran out of script quickly, and the buyer hit a wall. On that reading, better language models fix it, and the category recovers on quality. If that were happening, bot-only share should be climbing. It fell.

I have no data on conversation quality and I am not going to pretend otherwise. What the data does say is that 98.7% of those installs had nobody behind them while they were running. So what came off the page was not a conversation that disappointed anybody. It was one that could never be finished: the script ran out and there was no one to hand to. That last step is my reading, not a measurement.

Which leaves two readings, pointing opposite ways. If the script was the problem, a better script fixes it. If the problem was that the script had nowhere to hand off to, a better script produces a more articulate dead end. Nothing in four years of this data suggests the staffing decision changed. The companies deploying AI chat now are the same companies that deployed the last generation unattended.

9. What this means commercially

Everyone competing for what is left of this category is running a migration play. Comparison pages, switching offers, import your history. That play is aimed at the 48 sites that moved and the 82 that stayed to the end, out of the 130 we could track to a named vendor in both waves. A real market, and a small one.

The larger group is the 164, and the several hundred more across the study who removed chat and did not replace it. They are not in a bake-off and they are not reading comparison pages. They concluded the category was not worth staffing, and on the evidence of how they were staffing it, they were right.

The 249 sites still running it in 2025 now have the deadline the 2021 cohort never had, because retirement is set for 31 January 2027. This cohort removed chat at seven times the rate it staffed it when nothing was forcing the question. I would expect the forced version to run the same way, and that is a prediction rather than a measurement.

We have first-party evidence on what answering is actually worth, drawn from our own customer base. I will publish it separately rather than mix it into research about other people's websites.

10. Method, notes and bounds

The sample. 4,265 B2B websites classified in both 2021 and 2025, matched at site level. Drawn at random from B2B technology companies above 10,000 monthly visitors. It does not represent the whole web, small business, or non-technology B2B. Everything here is observational, taken from what is visible outside a website.

The demo response study, 2023. One figure here does not come from the study: the 93.4% who never put a human in front of a demo requester. That comes from a separate 2023 study of 1,685 B2B companies. Treat it as a baseline rather than a current reading.

Detection coverage, which is the caveat that matters. In 2021 we identified a vendor on 1,528 of 1,536 sites with chat, a detection rate of 99.5%. In 2025 we identified 896 of 1,183, which is 75.7%. Those denominators are the vendor analysis's own count of sites with chat.

The matched panel's count is 1,534 and 1,182, a difference of two sites and one. The vendor numerators have not been recomputed against it, so the detection figures are published on the base they were derived from.

That leaves eight unidentified sites in the first wave and 287 in the second. Some of that is genuinely fewer installs to find. Some of it is that vendor fingerprints change, and newer entrants are harder to recognize than established ones.

Those 287 sites decide the share question. They do not touch the exit finding, which rests on whether chat was present at all rather than on whose chat it was.

Drift's share, both bounds. Its share of identified vendors appears to rise, from 20.9% of the 1,528 identified installs in 2021 to 27.8% of the 896 in 2025.

The gain runs from 6.9 points on the identified-vendors-only reading to 0.2 points if none of the unidentified sites run Drift.

The leaver table rests on 48 sites. Of the 130 Drift sites where we could detect a named vendor in both waves, 82 were still on Drift in 2025, which is 63.1% retention. Forty-eight sites is a small base. The gap between 19 and 12 is not something to build a market model on. Read the destinations as direction only.

The study-wide change. The rise in sites with no chat at all, 64.0% to 72.3%, is tested with McNemar's test, which is the standard test for a before-and-after change measured on the same subjects. It returns p below 0.0001, meaning a change this size would almost never turn up by chance if nothing had really moved.

Vendor churn is a ceiling. 257 of 585, or 43.9%, is a ceiling rather than a point. A site recorded under two spellings of one vendor counts as a change until the normalization catches it, and normalization only moves the number down.

Three vendors with 15 or more installs in 2021 had no detected installs by 2025: LiveAgent with 60 sites, SnapEngage with 22 and LivePerson with 17.

The Gartner reading. The two releases state two bases on the same fieldwork, 646 buyers and 645 buyers. Gartner publishes no cross-tabulation of the two groups. So treating them as the same buyers is inference.

Full method is on the website panel page, the chat vendor share page and the demo response study page. Every figure here, with its base and the caveat that travels with it, is on the full figure set. The study and the demo study are reported side by side in The Buyer Wait Time Report 2026.

11. Questions

What happened to Drift?

Salesloft acquired it in February 2024 and has set its retirement for 31 January 2027, per Salesloft's Drift sunset FAQ. In this study its install base fell from 319 sites in 2021 to 249 in 2025. The corporate history is the part everybody tells. The part that is wrong is where the customers went. Most of them did not move to a competitor. 164 of the 319, slightly more than half, had no chat at all by 2025, and 256 had no route to a person of any kind. Of the 130 sites we could track to a named vendor in both waves, 48 moved. The exit ran roughly three times the size of the migration.

You could not identify the vendor on 287 sites in 2025. How do you know the 164 that went dark are not simply sites you failed to detect?

Because those are different classifications. The 164 are sites where no chat element was present at all. The 287 are sites where chat was clearly present and we could not tell whose it was. Detecting whether a widget exists is a much easier problem than identifying who made it, and the 287 are counted as having chat.

The honest limit is narrower. A chat product that leaves nothing visible until it is triggered, or one that only runs inside a logged-in application, would be recorded as no chat. I do not believe that accounts for 164 sites. But it is the boundary of what looking at a website from outside can tell you.

Is this just saying chat is dead?

No, and for four years I answered this question with the wrong figure. I used the combined human-reachable measure, which rose steeply, and concluded that the category had shrunk at the bottom and grown at the top.

Split properly, staffed chat went from 140 sites to 111, a change the study cannot separate from no change at all. What grew was a bot offering to bring a person, from 18 sites to 490. We tested that offer by conversation in 2026. It produced a person on 4 of 88 judged sites.

Meanwhile the widget with nobody behind it had fallen from 789 sites to 69. In 2026 it came back to 7.0% of the study.

So chat is not dead. The thing that is nearly absent, and has been the whole time, is a person scheduled to answer it.

Is 48 leavers enough to say anything about Qualified?

For magnitude, no. The ordering agrees with the study-wide figure, where Qualified went from 14 sites to 83 and was the largest gainer of the period. Two thin signals pointing the same way is worth reporting and not worth modeling.

Does this mean Drift's customers were unhappy with Drift specifically?

I do not think so, and the timing rules out the easy alternative. The exit pattern in the Drift cohort tracks the study-wide pattern almost exactly. Salesloft bought Drift in February 2024 and did not publish an end date until 2026, setting retirement at 31 January 2027. The sites we found dark in 2025 had no deadline to react to. They switched off a product that was still running.