62 B2B buyer response statistics, from five years of original research

Can a buyer reach a human? The five year study

Base for this section: 4,265 B2B technology websites, each above 10,000 monthly visitors, classified in 2021, again in 2025 and again in 2026 at the same addresses. The 2021 and 2025 figures are observed states with no coder judgment about intent, no estimation and no sampling. The 2026 wave sampled the three big groups and tried every site in the three small ones, so its figures carry confidence intervals and the first two waves do not. It is also a different instrument: 2021 and 2025 classified a site by opening the widget on one visit, while 2026 held a conversation and waited three minutes for a person.

1. Websites with no chat at all rose from 64.0% to 72.3% of the study between 2021 and 2025, 2,731 sites to 3,083. McNemar p below 0.0001, 95% confidence interval on the change +6.7 to +9.8 points.

2. 753 sites removed chat between 2021 and 2025 and 401 added it, a net loss of 352 sites.

3. Sites where a buyer could reach a person rose from 3.7% to 14.1%, 158 sites to 601. Change interval +9.3 to +11.5 points. That figure combines two states and must never be published without the split at items 4 and 5.

4. Staffed chat, a widget with a person scheduled behind it, did not move: 3.3% in 2021, 2.6% in 2025, 140 sites then 111. Change -0.68 points, 95% confidence interval -1.38 to +0.02 points, McNemar p=0.064. The study cannot separate this from no change.

5. Bot to person, a bot that offers to bring one, went from 18 sites to 490, 0.4% to 11.5%. Change +11.07 points, 95% confidence interval +10.10 to +12.03 points, p below 0.0001. The whole of the rise at item 3 is this state.

6. Chat present with nobody behind it fell from 18.5% of the study to 1.6%, 789 sites to 69. Change interval -18.1 to -15.7 points, p below 0.0001.

7. Of the 789 sites running an unstaffed widget in 2021, 391 were switched off and 55 had a person put behind them. A further 174 had a bot put in front offering a person and 138 a bot offering nothing. Removal beat putting a person behind it by 7.1 to one.

8. 560 of the 789, 71.0%, ended 2025 with no human reachable.

9. Bot-only sites moved from 13.8% to 12.0% of the study, 587 sites to 512. p=0.009, change interval -3.1 to -0.5 points. This is the state that covers the entire arrival of modern language models.

10. Across all 1,534 sites in the study that had chat of any kind in 2021, 10.3% could reach a person, 158 sites, and 9.1% were genuinely staffed, 140 sites.

11. Human reachability by sector, 2021 to 2025. B2B: staffed chat 3.11% to 2.54%, bot to person 0.39% to 12.81%, combined 3.51% to 15.35%, across 2,795 sites. B2C: staffed chat 3.80% to 2.85%, bot to person 0.51% to 9.58%, combined 4.31% to 12.43%, across 1,368 sites. 102 study sites carry both sector flags or neither and are excluded. No significance test is published on this comparison. The direction of the combined measure depends on which of the source workbook's sector columns is used to define the groups, and a comparison that changes sign between two reasonable readings is not a finding.

12. The 2021 wave holds 5,007 rows, of which 5,004 carry a complete single classification, across 4,832 distinct domains. Matching to the 2025 wave on domain gives the analysis panel of 4,265.

---

The 2026 wave: what a buyer meets when they actually ask

Base: the same 4,265 sites. Coded 16 to 18 September 2026 by opening the chat, asking a buyer's question, and waiting a fixed three minutes, in US business hours. The three big 2025 groups were sampled at roughly 110 sites each, and every site was tried in the three small ones. Weighted to the 2025 group sizes with a finite population correction.

13. 2026 state shares on the study, with 95% confidence intervals and the 2025 figure beside each. No chat 74.1% (69.1 to 79.1), against 72.3%. Bot only 15.6% (12.1 to 19.0), against 12.0%. Unstaffed chat 6.3% (3.2 to 9.5), against 1.2%. Staffed chat 2.3% (0.4 to 4.1), against 2.6%. Bot to person 1.1%, against 11.5%. Bot to unstaffed chat 0.7% (0.2 to 1.1), against 0.4%. The bot-to-person interval runs below zero on a normal approximation, which is reported rather than clipped because it shows the method at its limit on a share that small.

14. Of the sites whose bot offered to bring a person in 2025, 4 of 88 judged still delivered one in 2026, 4.5%, 95% confidence interval 1.8% to 11.1%.

15. Of the sites that were genuinely staffed in 2025, 22 of 104 judged still reached a human in 2026, 21.2%. A census of the whole group, so no sampling interval. Where the other 82 went: unstaffed chat 33, no chat 32, bot only 15, bot to unstaffed chat 2.

16. A schedule with names on it survived 4.7 times as often as a promise to fetch somebody, 21.2% against 4.5%.

17. A chat widget with nobody behind it returned. Any unstaffed widget was 1.6% of the study in 2025 and 7.0% in 2026, 95% confidence interval 3.8% to 10.2%. The interval excludes the 2025 figure.

18. Across all three waves, staffed chat stayed at about one site in forty: 3.3% in 2021, 2.6% in 2025, 2.3% in 2026.

19. The machine sweep that ran before the conversations missed chat on 45 of 155 live sites it had called empty, 29.0%. Every machine-detected count of chat presence in this wave is a floor, never a point.

20. 50 of the 93 domains that were no longer live redirect to a different company, so part of the movement out of chat is consolidation rather than retreat. A further 21 of the 93 only blocked the sweep and still exist. Sites that were not live are excluded from their group denominator, which assumes a site that died behaves like the survivors in its group.

---

Chat vendors, 2021 to 2025

Every share figure in this section sits on detection coverage, which fell sharply between the waves. Quote the rate, not the raw count, and never quote a share without the install count beside it.

21. Vendor detection fell from 99.5% of sites with chat in 2021, 1,528 of 1,536, to 75.7% in 2025, 896 of 1,183. That leaves 287 sites in 2025 where chat was clearly present and no vendor could be attributed.

22. Drift ran on 319 study sites in 2021 and 249 in 2025.

23. Drift's share of identified installs went from 20.9% to 27.8%. On the assumption that none of the 287 unidentified sites run Drift, it went from 20.8% to 21.0%. Somewhere between flat and up 6.9 points, and this data rules out neither end. No vendor in this study gets a share rise attributed to it.

24. 315 of the 319 Drift sites, 98.7%, had no human reachable in 2021, while they were still paying for it. 295 ran a bot with nobody behind it, 20 ran an unstaffed widget. Four could put a buyer in front of a person, 1.3%, against 10.3% across all sites with chat.

25. By 2025 the 319 Drift sites had gone: no chat at all 164, bot only 85, human reachable 63, unstaffed widget still running 7. 256 of the 319, 80.3%, had no route to a person.

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

27. The 48 Drift sites that moved went to Qualified 19, HubSpot 12, Intercom 9, Terminus 2, and six other vendors at one site each.

28. 257 of the 585 sites where we identified a vendor in both waves had changed vendor by 2025, 43.9%. This is a ceiling rather than a point: a site recorded under two spellings of one vendor counts as a change until normalization catches it, and every correction so far has moved the number down.

29. Three vendors with 15 or more installs in the study in 2021 were not detected on a single site in 2025: LiveAgent 60, SnapEngage 22, LivePerson 17. Absence from this study is not a business failure. Several of these companies sell largely outside the segment it covers.

30. Vendor install counts, 2021 to 2025: Intercom 307 to 182. HubSpot 127 to 138. Qualified 14 to 83. Zendesk 183 to 52. LiveChat 130 to 52. Shares of identified installs: Intercom 20.1% to 20.3%, HubSpot 8.3% to 15.4%, Qualified 0.9% to 9.3%, Zendesk 12.0% to 5.8%, LiveChat 8.5% to 5.8%.

---

What happens when a buyer asks for a demo

Base for this section: 2,528 B2B companies, each above 10,000 monthly visitors, each sent one demo or contact sales request through their own website form during 2023. Fieldwork is three years old. Treat these as a baseline rather than a current reading.

31. 798 of the 2,528 demo forms could not be submitted at all, 31.6%. A further 45 errored. This is an upper bound on breakage: a form that defeated our operator may yield to a determined buyer.

32. After the forms that could not be submitted, the usable base is 1,685 companies, 66.7% of the 2,528 approached.

33. 93.4% of the B2B companies we asked never put a human in front of the buyer. That is 1,573 companies out of 1,685. The 95% confidence interval runs from 90.3% to 96.1%.

34. The complement: an estimated 6.6% did put a human in front of the buyer, about 112 of 1,685, 95% confidence interval 3.9% to 9.7%. That is a modeled expectation, not 112 companies we could name. About one in fifteen.

35. The interval on the 93.4% covers sampling error in three group rates and nothing else. The largest group holds 1,060 of the 2,102 readable emails, was hand read 30 times and returned no human replies, so the estimator treats its rate as certain when it is not. If the true rate in that group is 3% rather than zero, the headline is 91.5%. If it is 5%, it is 90.3%.

36. Collection stored up to two emails per company and stopped there. Ten of the seventeen human replies sat at the second email, and in six of those ten the first email was an auto receipt or a bulk send. A third email would have found further human contact, costing about one point on the observed rate and no more than 2.4 points.

37. 153 of the 1,685 companies, 9.1%, produced no email with a readable body, so no hand-classified group applies and the estimator counts them as never having answered. If those 153 answered at the rate of the other 1,532, the headline falls from 93.4% to 92.7%.

38. 545 of the 1,685 first replies arrived inside five minutes, 32.3%. This count is a timestamp subtraction and rests on no judgment about who wrote anything.

39. Sixteen of the 545 fast first replies fell into the hand classified sample and none was written by a person: fourteen automated acknowledgments, two marketing sends. Those sixteen are not a random draw, so no rate can be built on them.

40. What the design does support: at least 342 of the 545 fast replies, 63%, sit in the group carrying bulk mail infrastructure, the group that returned no human replies across 30 hand reads.

41. Median wait to a first reply of any kind was 19.7 minutes, across all 1,685 first replies, with a quarter inside 4.1 minutes and a quarter beyond 393.9 minutes.

42. On a different and clearly separate population: strip out bulk sends, role sender addresses and unattended out-of-hours mail and 531 first replies survive, among which the median wait is 111.7 minutes, a quarter inside 24.5 and a quarter beyond 911.3. Only one of these two medians belongs in any given sentence.

43. Median lag was 765 minutes for the replies a person wrote against 281 minutes for automated acknowledgments. Base: 17 human and 162 acknowledgment verdicts. Direction, not measurement.

44. The three fastest replies in the entire hand classified set landed at 0 seconds, 3 seconds and 5 seconds. All three were machines.

45. 499 of the 1,685 first replies, 29.6%, arrived outside the sender's own working hours: before 08:00, from 18:00, or at a weekend in the sender's local time taken from the email header offset.

46. 984 of the 1,685 companies, 58.4%, sent exactly one email and never a second, inside a capture that stopped at two. 701 sent two.

47. Replies arrive as late as 133.6 hours after the form was submitted for a first reply, and 149.2 hours for a second email. There was no fixed observation window: collection was bounded by the two-email cap, not by elapsed time.

---

What the reply actually contained

48. The one feature that separated a human reply from an automated acknowledgment was the sender reporting an act they had personally performed on this specific submission, and naming what it turned up. Present in 14 of 17 human replies and 5 of 162 acknowledgments.

49. Six features that separated nothing, on bases of 17 human replies and 162 acknowledgments: asking a question, 71% in both classes. "Thank you for your interest", 18% against 17%. A named personal sender address, 17 of 17 against 152 of 162. A booking link, 41% against 38%. The buyer's company name in the body, 24% against 17%. Speed, which runs backwards.

50. 785 of the 1,685 first replies demand a further step from the buyer, 46.6%, and 52.6% of the 1,493 with a readable body. A floor, because link tracking rewrote some scheduler destinations out of the record.

51. Between 3.9% and 27.8% of first replies sent a scheduling link, 66 to 468 of 1,685. The lower end counts only links still visible in the record; the upper adds every reply whose text says a calendar link was supplied. On visible links alone, 34 companies sent one and asked the buyer nothing at all, 2.0%.

52. Among the 545 fast first replies, 221 carried an imperative call to action with a link, 40.6%.

53. Question rates come from three different instruments and must never be mixed. 82.6% of readable emails contain a question mark anywhere. 31.4% pass a strict rule requiring a question mark, a second person pronoun, at least three words and no match against fifteen template patterns such as "Questions?". That is 661 of 2,102, and 428 of 1,685 first replies, 25.4%. 71% were judged to ask a question by a person reading them, in both the human and the acknowledgment class. The 71% is a comparison between two classes on one instrument and is not a population rate.

54. First replies run to a median of 136 words, a quarter under 87 and a quarter over 218, range 10 to 1,363. Base: the 1,493 first replies with a readable body.

55. Across all 2,102 readable emails: bulk mail infrastructure 47.4%, role sender address 29.0%, imperative call to action with a nearby link 30.0%, booking link 4.3%, form echo 4.0%, specific reference beyond merge fields 2.7%, claims a personal action 5.9%. The booking figure and the specific reference figure are both floors. Detecting a specific reference requires reading an email for meaning, which the automated pass does badly: a hand census of the same corpus found 55 emails reporting a phone call the sender had placed, the clearest form of the feature there is, and the automated pass flagged none of them. The true share is at least 5.2%.

56. Between 8% and 26% of companies sent at least one marketing email that never acknowledged the request, best estimate 17%, 289 of 1,685.

---

Form friction

57. 1,192 of the 1,685 forms demanded a phone number, 70.7%. 493 did not.

58. 292 forms asked for three fields or fewer. 170 asked for eight or more.

59. 52 of the 1,685 forms used a CAPTCHA, 3.1%.

60. 766 of the 1,685 companies had chat software installed and 919 did not.

61. This study cannot tell you whether a phone field, form length, a CAPTCHA or having chat installed changes whether a person answers. It was built to measure one rate precisely. Estimating each subgroup honestly gives intervals 6.5 to 10.6 points wide that overlap heavily, and eleven of the eighteen cells contain no human verdicts at all. Settling the chat comparison alone would take roughly 880 further hand classifications, and the form length comparison cannot be settled at any sample size we could reach. We publish no subgroup comparison of the never-answered rate and no significance test on one.

62. 144 domains appear in both studies, 8.5% of the demo base and 3.4% of the website study, so the two are close to independent samples of the same market and their bases must never be mixed.

---

How to quote these

Every figure above carries its base and its year. The two studies have different bases and different fieldwork dates and should never be combined into one sentence.

The demo response study is 2023 fieldwork. The website study is 2021, 2025 and 2026. Where a figure is a bound rather than a point, both ends are printed and both should travel.

Prior audits of inbound response, and what this set measures that they did not, are on the demo response study page.