Chat vendor share, 2021 to 2025, method and dataset

Full figure set: 62 B2B buyer response statistics.

Vendor detection fell from 99.5% of panel sites with chat in 2021 to 75.7% in 2025. Every share figure on this page sits on that rate, and every share movement is published as a bound rather than a point.

The caveat comes before the results because the results are not interpretable without it. Figures were recomputed from source on 14 August 2026.

The caveat that governs every figure here

Base: the 4,265 site matched panel.

The rate fell 24 points. State it as a rate. The raw fall from 1,528 identified sites to 896 reads as fewer installs, when the actual problem is thinner detection.

Part of the fall is fewer sites running chat at all, which the website panel documents directly. Part of it is detection failing on sites that do run chat. The two causes cannot be separated from the data collected.

The consequence is specific. If the 287 sites we could not identify in 2025 skew toward newer entrants, then part of any share rise attributed to an established vendor is a detection artifact rather than a real gain. Those 287 sites decide whether any share figure on this page holds.

None of this applies to counts of whether chat is present at all. Detecting the presence of a widget is a substantially easier problem than attributing it to a vendor, and the website panel's presence and staffing figures do not carry this caveat.

Method

Vendor was detected by examining what each panel site loaded and exposed to a visitor. Detection was attempted on every site in each wave.

Sites where no vendor could be identified are excluded from share denominators rather than assigned to an other category, because assigning them would hide the coverage problem inside a residual bucket. Shares are calculated among sites where a vendor was detected and chat was present in the same wave, so the denominators are 1,528 for 2021 and 896 for 2025.

Brand strings are normalized so the work can be rerun. Strings are lowercased and trimmed. Obvious misspellings are folded, so "Drfit" counts as Drift and "Hubspto" as HubSpot. Product names are folded to their vendor, so Zopim counts as Zendesk, SalesIQ as Zoho, "Live Chat" as LiveChat and "Terminus chat" as Terminus. Cells that name no vendor at all are treated as unidentified rather than as a vendor. A site counts as an install only if chat was also detected as present in that wave.

Two of those rules moved published counts. Folding misspellings adds one Drift site in 2021 and three in 2025, against an earlier internal figure set that used exact string matching. Treating cells that name no vendor as unidentified removes sixteen sites from the 2025 identified base: six record the reviewer's own doubt, one holds a response time belonging to another column, one holds free text describing what the bot said, and eight read "UI", which describes an interface rather than naming a product a reader could look up. The same rule removes nothing from the 2021 wave, where all 1,528 in-scope cells name a vendor, a product or a named assistant.

One judgment inside the identification base is larger than any correction made to it. "Live Chat" as two words, on 44 sites in 2021 and 48 in 2025, is folded into LiveChat. Read instead as a generic description rather than a vendor name, 2021 identification falls to 1,484 of 1,536, 96.6%, and 2025 falls to 848 of 1,183, 71.7%.

The reporting rule: share and count together

Share and absolute install count are always published together, and neither is published alone.

Drift is the case that shows why. Its share of identified vendors rises from 20.9% to 27.8% while its install count falls from 319 sites to 249. Share rises because the pool shrank faster than Drift did. A share figure quoted alone would describe that as growth. The count shows a vendor losing 70 installs inside a contracting market.

This holds for every vendor in the table below and is the most likely way for these figures to be misread.

Results

Shares are of identified installs, 1,528 in 2021 and 896 in 2025.

Intercom is the clearest illustration of the reporting rule. Its install count fell from 307 sites to 182, a loss of 125 sites, while its share of identified installs moved only from 20.1% to 20.3%. The pool that share is measured against fell 41.4%, from 1,528 to 896, and Intercom fell 40.7%, so holding share here means shrinking at the rate of the pool. Sites running chat of any kind fell 23.0% over the same period, 1,536 to 1,183.

HubSpot and Qualified are the only vendors here whose absolute counts rose. HubSpot went from 127 sites to 138. Qualified went from 14 sites to 83, off a base small enough that the percentage change should not be quoted without the underlying numbers.

Drift's share, published as a bound

Bases: 1,528 and 896 for the identified-vendors row, 1,536 and 1,183 for the lower-bound row.

Drift's share change runs from plus 0.2 points on the lower bound to plus 6.9 points on the identified-vendors-only reading. Neither end can be ruled out by this data, so both are published and no share rise is asserted for any vendor.

Vendor retention

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

Working, because this base has been published at three different values. 132 of the 319 Drift sites carry a 2025 brand string. One of those had no chat present in 2025, which gives 131. Two more hold a cell that names no vendor, which gives 130. The retention numerator of 82 is the same on every version.

The 48 sites that left went to:

The five rows sum to 48 on a base of 130. Zendesk is one of the six single-site destinations.

These destination counts are direction rather than magnitude. The base is 48 sites, most destinations are in single figures, and the detection caveat applies at both ends of every move. The reading they support is that departing Drift sites went to more than one place, with Qualified and HubSpot taking the largest shares. The reading they do not support is a quantified market transfer.

Churn

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

That is a ceiling rather than a point estimate. A site recorded under two spellings of one vendor counts as a change until the normalization catches it, and the normalization is a mechanical key plus an explicit alias map, so any spelling it misses is still counted as a change. Of the 257 changes, 28 end at a vendor seen on exactly one panel site and 39 end at a vendor absent from the 2021 wave, which is where an uncaught alias would sit. Every correction made to this figure so far has moved it down.

Those 585 sites are the subset of the panel where detection succeeded twice. Sites where detection succeeded in only one wave cannot contribute a churn verdict and are excluded. A site that changed to a vendor we could not identify does not appear in this count at all, which is the detection caveat operating on a churn figure.

Vendors that were not detected in 2025

Three vendors with 15 or more installs on the panel in 2021 were not detected on a single panel site in 2025.

The base here is vendors, not installs. Zoho is not in this table: it fell from 15 sites to 1, so it thinned rather than disappeared.

Detection returning zero is not the same as a vendor ceasing to exist. It means no panel site was detected running that vendor in 2025. For LivePerson in particular, which operates outside the segment this panel covers, the correct reading is absence from these 896 identified installs rather than absence from the market.

The Drift cohort

319 panel sites were running Drift in 2021, identified by brand with chat present. Their state in that year, while the product was in use:

315 of the 319 had no human reachable, 98.7%. That is the 295 running a bot with nobody behind it plus the 20 running a widget nobody was staffing, an unstaffed widget having nobody behind it by definition. The 295 is the bot-only subset and is not the count of sites with no human reachable. Any sentence about reachability takes the 315.

Across all panel sites with chat of any kind in 2021, 10.3% were human reachable, 158 of 1,534 sites, of which 140 were genuinely staffed and 18 were a bot offering to bring a person. The Drift cohort figure is 1.3%, and it rests on four sites.

The 1,534 is the matched panel's own count of sites with chat. The 1,536 used as a denominator in the vendor detection rows on this page is the vendor analysis's count. The two differ by two sites, the vendor numerators have not been recomputed against the panel count, and the vendor rows are therefore published on the base they were derived from rather than restated onto one they were not.

Where those 319 sites were in 2025:

256 of the 319, 80.3%, had no route to a person by 2025. That is the 164 with no chat, plus the 85 running a bot, plus the 7 still running an unstaffed widget, which by definition has nobody behind it. Note that Drift's 2025 install count is also 249, and the two are different quantities.

Of the 63 that became human reachable, 60 came from bot only.

What this page cannot support

No share movement is a point estimate. Every share figure here depends on detection coverage that fell 24 points between the waves, and that dependency cannot be tested from the data collected. Share movements are published as bounds and no vendor is asserted to have gained share.

It carries no outcome data of any kind. A detected vendor means the software was present on the site. Nothing here records whether the widget was staffed, how much traffic it handled, what the site paid, what it returned, or why any site switched. Destination counts describe where sites went, not why.

It cannot be crossed with the human reply rate by site type. The demo response study was powered to measure one overall rate and cannot resolve differences between site types, so no vendor cohort should be given a never-answered rate of its own.

Detection is not neutral across vendors. Vendors differ in how identifiably they load. A vendor that is harder to fingerprint is undercounted in both waves, and if its detectability changed between waves the effect appears as share movement.

The 2026 wave measured how weak automated detection is. A machine sweep ran before the 2026 conversations to find chat, and it missed chat on 45 of 155 live sites it had already called empty, 29.0%. That is a direct measurement of the effect this caveat describes, on the same panel, and it is why every machine-detected count of chat presence publishes as a floor. The 2021 and 2025 vendor and presence work was done by a person opening each widget rather than by a sweep, so the 29.0% is not a correction to those waves. It is evidence that the caveat is real and large.

Panel scope. These figures inherit the panel's construction: sites above 10,000 monthly visitors, drawn in 2021, matched to 2025. They describe vendor share on that panel, not any vendor's total install base or revenue.

How to cite this

GTM Clarity, Chat vendor share, 2021 to 2025, method and dataset, 2026. Website data collected 2021 and 2025. Research directed by Terry Wilson, fieldwork by the GTM Clarity research team.

The vendor findings are reported in The Buyer Wait Time Report 2026.