They changed the chat vendor. They did not change who answers.

Say you are choosing a chat vendor this quarter. What you have to go on is review grids, analyst charts drawn up from vendor briefings, and case studies the vendors wrote. All of it describes right now. None of it tells you what happened to the last group of people who made this decision.

The decision is not really about which product is best today. It is about what your website looks like in three years. By then the vendor may have been acquired, or repositioned, or quietly deprioritized the product you bought in favor of the one that sells better.

Nobody publishes that, and the reason is boring rather than sinister. Producing it means watching the same websites for years, committing to a sample before you know whether it will show you anything, and going back to those exact addresses when the interesting thing may have happened somewhere else entirely.

A snapshot tells you what is installed. A panel tells you what a decision turned into. A buyer is asking the second one.

So we watched the same sites. Same addresses, same classification scheme, four years apart.

The findings

Every row above sits on 4,265 B2B websites classified in every wave, drawn at random from B2B technology companies above 10,000 monthly visitors. The traffic floor was set before collection so that no finding could be waved away as describing sites nobody visits. The 2021 and 2025 figures are counts of observed states. The 2026 figures are an estimate built group by group and carry confidence intervals, which are in the notes.

A census in that second-last row means we tried every site in that state rather than a sample of them. Judged means the conversation gave a clear answer, so the denominators are the sites we could read. The website panel page has the sampling.

Holding share while the room empties

Intercom's install count fell from 307 sites to 182 while its share of identified installs moved from 20.1% to 20.3%.

Shares below are calculated among sites where a vendor was identified and chat was present. Base of 1,528 sites in 2021 and 896 in 2025.

The counts carry more information than the percentages here, because the percentages describe a pool that shrank.

Holding share here means shrinking at the speed of the pool the share is measured against. That pool drained faster than the number of sites running chat at all. The arithmetic is in the notes. A vendor could put the flat share line in a board deck and nobody in the room would ask about the sites it lost.

Across the whole study, 753 sites removed chat and 401 added it, a net movement of minus 352 sites. Every share point in that table is a share point of something smaller than it was. Winning this category now means winning a category fewer people want to be in each year.

Two counts rose. HubSpot went from 127 to 138. Qualified went from 14 to 83, off a base small enough that I would not quote the percentage change without the raw numbers next to it.

HubSpot grew installs while almost every other vendor shrank. That is what bundling into a platform people already own tends to look like. Nobody in that group of 138 sites necessarily chose HubSpot chat against a field of alternatives. They chose HubSpot, and chat came with it.

That is my inference and not something the study can show. It records what loaded on a website, not why a procurement decision went the way it did.

Drift's install count fell by 70 sites over the same period, 319 to 249. Its share is the one figure here I will not hand you as a single number. Vendor fingerprints got harder to read between the waves, so the honest answer is a range. Both ends of it are in the notes.

That caveat governs share and nothing else. In 2025 there were 287 sites where chat was clearly present and no vendor could be attributed to it. So the table reports the shares as measured, and this study asserts no real market gain for any vendor on the strength of them. That includes the ones where a gain would make a better story. Detecting that a widget exists is a much easier problem than working out who built it, which is why the category-level movements here are firm and the vendor-level shares are not.

Two in five of them moved, at the outside

257 of the 585 sites where we identified a vendor in both waves had changed vendor by 2025, 43.9%.

Four years. Two in five, and that is the top of the range rather than the middle of it. The mechanics are in the notes, and every correction so far has moved the number down.

From the buyer's side of the glass, all that work shows up as a different logo in the same corner of the same page.

If you are running a selection process right now, that number should change the shape of the process rather than the shortlist. You are not making a permanent decision. Two in five of the sites we could track through both waves remade this one inside four years. It gets remade by you, or by whoever holds the job after you.

So the questions that dominate a typical selection are the wrong size. Feature grids assume permanence. Three-year contracts assume permanence. Deep integration work that only pays back over years assumes permanence.

The questions that survive a 43.9% churn rate are the unglamorous ones. How do the transcripts come out. How does routing get rebuilt when the vendor changes. What breaks in the CRM when the widget is swapped. Who answers the messages, whichever logo ends up in the corner of the box.

A team that can answer those has a bad quarter when it switches vendors. A team that cannot has a bad year.

The vendor they chose is not on the map now

Three vendors ran on 15 or more study sites in 2021 and had no detected installs in 2025.

Zoho went from 15 study sites in 2021 to one in 2025. It misses the table by a single install, which makes the table a conservative count rather than a complete one.

That table is not an obituary column, and it is worth saying exactly what it is.

Detection returning zero means no site in this study was found running that vendor in 2025. It does not mean the vendor stopped trading, lost its customers, or failed. Some of these companies operate substantially outside the segment this study covers. That segment is B2B technology sites above 10,000 monthly visitors.

LivePerson in particular has a business this sample was never built to see. None of these three is a business failure, and this table is not evidence that any of them is.

The narrower thing it says is this. A buyer who picked one of them in 2021, from that segment, ended up with something no longer visible among their peers four years later. That was not a foolish choice at the time. LiveAgent on 60 sites was an ordinary mid-market decision. Ordinary decisions are the ones that disappeared, which is the reason to plan around churn at all.

Where Drift's leavers went

Of the 130 Drift sites where we identified a vendor in both waves, 82 were still on Drift in 2025, 63.1%.

The 48 that left went in several directions at once. Qualified and HubSpot took the largest pieces, then Intercom, then Terminus, then six other vendors on one site each. The split is in the notes.

Direction only. Forty-eight sites, the top two destinations on 19 and 12 and the rest in single figures, and the same detection problem applies at both ends of every move. They show that departing sites scattered instead of consolidating behind one successor. They will not carry a calculation of market transfer. I would rather say that plainly than let someone build a slide out of a base this small.

A larger number sits next to it. Of the 319 sites running Drift in 2021, 164 had no chat at all by 2025. 164 sites quit against 48 that switched, and those two counts sit on different bases, set out in the notes. On the common base of the 130 we could track to a named vendor in both waves, 48 switched. The quitting is the larger movement on any reading of it. The whole cohort is followed in the Drift aftermath study.

A new box, the same empty corner

789 sites were running a widget with nobody behind it in 2021. Companies that had already paid for the software removed it rather than put a person behind it by seven to one.

Stand on the buyer's side of the glass and compare these products honestly. A box appears in the corner. It says hello. There is a field to type in. Whatever is behind that field, the buyer's experience of it is nearly identical from one vendor to the next. The buyer cannot see your integration architecture, your routing rules, or the sophistication of your intent model.

What actually differs is what happens after the box appears. Whether it was configured for a real conversation or shipped with the demo copy still in it. Whether anyone maintained it after the person who set it up changed roles. Whether anyone was on the other end.

That last one is where the study is unambiguous. Of the 789, 391 switched the widget off and 55 put a person behind it. A further 174 put a bot in front of it that offered to bring a person. In 2026 that configuration delivered one on 4 of 88 judged sites. Where all 789 ended up is set out on the website panel page.

Those 789 sites all bought software. The software worked, in the sense that it loaded and rendered and collected. The thing that decided whether it produced anything was whether a person was assigned to it, and that was never a property of the vendor.

So the 43.9% has an obvious shape to it. A widget underperforms. The vendor gets the blame. A new vendor is selected on a better feature grid. And the new box goes into a company where nobody's job description contains the sentence "answer the people who ask." The switch was real work. The result was a fresh logo on the same silence.

Across four years, as many as two in five of the sites we could track changed supplier. Over the same four years, the share with a person scheduled behind the widget went from 140 sites in 2021 to 111 in 2025. That is 3.3% of the study against 2.6%, a change the study cannot separate from zero.

At both ends of those five years, about one site in forty had somebody scheduled to answer.

What did grow was a bot offering to bring somebody, from 18 sites to 490. In 2026 we asked those bots for a person and got one on 4 of 88 judged sites. The sequence those bots ran instead is in Ask a B2B chatbot for a person and it asks for your details first.

So the market worked hard on the question it could answer with a purchase order, and left the other one exactly where it was.

That is what a buyer meets. Not a slower answer: no answer. And nothing on the front of the page to tell them in advance which kind of site they have landed on.

You can test this on your own site in about ten minutes. Open your chat widget from a browser your company does not recognize, ask what a buyer would ask, and time the reply. If nothing comes back, you have learned the thing the feature grid cannot tell you, and no vendor on the grid fixes an empty chair.

Method

B2B technology websites were drawn at random in 2021, each above 10,000 monthly visitors. Each was classified by how a buyer could reach a human. The 2021 wave holds 5,007 rows, of which 5,004 carry a complete single classification, across 4,832 distinct domains. Matching those to the 2025 wave on domain gives the analysis panel of 4,265 sites, reclassified on the same scheme.

Every finding here rests on that 4,265. Because the same sites appear twice, a site that removed chat was observed removing chat. It was not inferred from a shifting sample. Full method is on the website panel page and the chat vendor share page. Every figure quoted here sits with its base in the full figure set. The rest of the study's findings are in The Buyer Wait Time Report 2026.

Two limits beyond the detection caveat. Each site was seen once per wave. And an install is not a usage figure: a detected vendor means the software was present, not that anyone was behind it.

Notes and bounds

Detection coverage. Every percentage in the share table depends on being able to tell whose software loaded on a page, and between the two waves that got substantially harder.

The detection rate fell 24 points. In 2025 there were 287 sites where chat was clearly present and no vendor could be attributed to it. Vendor fingerprints change over time, and newer entrants are generally harder to identify than established ones. So those 287 sites are unlikely to be a random draw from the market.

They probably skew toward vendors that are harder to see. Which means they sit directly on top of every share figure in the share table.

Drift's share, both bounds.

Best case, Drift gained 6.9 points of share. Worst case it gained 0.2 points, which rounds to nothing. Both of those are consistent with the data collected, and there is no way from inside this dataset to tell you which one is true. So we publish both and assert neither.

This caveat governs share only. It does not touch the counts of whether chat is present, whether a widget was staffed, or whether a human was reachable. Detecting that a widget exists is a much easier problem than working out who built it.

The pool the shares sit in. Identified installs fell 41.4%, from 1,528 to 896, and Intercom fell 40.7%, from 307 sites to 182, a loss of 125 sites. Sites running chat of any kind fell 23.0% over the same period, 1,536 to 1,183.

The churn rate is a ceiling. A site recorded under two spellings of one vendor counts as a change until the normalization catches it. Every normalization pass so far has moved the number down, never up.

Automated detection is weaker than a human look, and the 2026 wave measured how much. A machine sweep ran before the 2026 conversations and missed chat on 45 of 155 live sites it had already called empty, 29.0%.

Any machine-detected count of chat presence is therefore a floor. Every presence figure in the 2021 and 2025 waves was produced by a person opening the widget, not by a sweep. That is why those are counts rather than floors.

Part of the churn is consolidation. Of the 93 study domains we could not reach in 2026, 50 redirect to a different company. A vendor losing an install sometimes lost it because the customer was acquired, not because it switched. A further 21 of the 93 are still trading and simply blocked the sweep.

2026 intervals. No chat 74.1%, 95% CI 69.1% to 79.1%. Any unstaffed widget 7.0%, 3.8% to 10.2%. Staffed chat 2.3%, 0.4% to 4.1%. Bot to person 1.1%, minus 0.7% to 2.9%. That interval runs below zero on a normal approximation, and is reported as computed rather than clipped at zero.

The vendor-analysis base. The vendor detection rate of 99.5% falling to 75.7% sits on 1,528 of 1,536 sites in 2021 and 896 of 1,183 in 2025. Those are the vendor analysis's own count of sites with chat.

The matched panel's own count of sites with chat is 1,534 in 2021 and 1,182 in 2025. The two differ by two sites and one site. The vendor numerators have not been recomputed against the study count. So the vendor figures are published on the base they were derived from, rather than restated onto a base they were not.

Where the 48 Drift leavers went. Qualified 19, HubSpot 12, Intercom 9, Terminus 2, and six other vendors on one site each.

Bases that differ. The 48 Drift leavers come from the 130 sites where a vendor was identified in both waves. The 164 sites with no chat at all by 2025 come from the full 319 cohort.

The 43.9% churn rate sits on the 585 sites with a vendor identified in both waves. The staffed-chat and bot-to-person shares sit on the full 4,265. The 4 of 88 and the 22 of 104 sit on the 2026 sample and census of their own 2025 groups, not on the study.

Questions

Does 43.9% churn just mean the market is competitive and healthy?

Competitive markets churn while growing. This one churned while contracting, losing a net 352 sites from a study of 4,265 sites. Two in five of the remaining buyers changed supplier inside a shrinking pool. I read that as closer to dissatisfaction than to competition, and it is a reading rather than a measurement: two observations per site cannot show whether a switch settled anything.

Are the vanished vendors out of business?

No, and I am not going to imply otherwise. Zero detected installs means nobody in this study of B2B technology sites above 10,000 monthly visitors was found running them in 2025. Several of these companies sell mostly outside that segment, so their absence here is absence from this sample. The finding is about what happened to buyers in this segment, not about the health of those companies.

Why publish share figures at all if the detection rate fell 24 points?

Because the counts are informative even where the percentages are contested. And because suppressing the table would leave the field to numbers published without any caveat at all.

Every share figure here carries its base and, for the vendor where it matters most, both bounds. Drift is somewhere between up 0.2 points and up 6.9 points, on an install count that fell from 319 sites to 249. Anyone quoting 27.8% without that range and that count is quoting a number this data does not support.

We are mid-selection right now. What do we do with this?

Assume you will do this again. Spend the negotiation on exit terms, transcript export, and contract length rather than on the fifteenth feature comparison. Then settle the question this study keeps running into, which is who answers, on what hours, with what target response time, before you sign anything. If that name does not exist, the vendor choice is not the decision in front of you.