# Methodology: Book a demo is winning on B2B software homepages

By Cristian Frunze, Founder of Ken · Oct 22, 2026 · [Ken Research](/research) No. 12 · Websites and pricing, Sales teams, Founders and startups

The full methodology behind [Book a demo is winning on B2B software homepages](/research/demo-vs-free-trial-ctas.md): the data, every method, every number and every limitation. The same text as the [methodology PDF](/research/demo-vs-free-trial-ctas.pdf).

Data: Ken's web archive homepage captures and Ken internal data (Ken's dataset of public professional profiles, work histories and company records). Every number below comes directly from the study's outputs.

## 1. Question and pre-registration

How has the main homepage call to action (CTA) shifted between "Book a demo / Contact sales" and "Start free / Sign up", by company size and stage? Do companies that switched grow faster afterwards, and does their sales team change shape?

The hypothesis and plan were pre-registered before any CTA was extracted. Before that, column schemas and two sample rows of each input file were printed to check their layout; no aggregate or outcome was looked at. The study plan records every later change and the reason for it.

Hypotheses in brief: (1) demo/sales is the most common primary CTA in every year; free/self-serve peaks between 2019 and 2022 and then falls; pages with both kinds rise; larger and older companies lean to demo/sales. (2) Demo-to-free switchers do not grow headcount meaningfully faster than matched stayers (8-quarter median gap within plus or minus 3 points). (3) After a demo-to-free switch, SDR/BDR headcount grows less than at matched stayers.

## 2. Data

**Panel.** The shared homepage panel built for the Ken Research study *Half of B2B software homepages now say AI. Did it pay off?*: a seeded simple random sample of 10,000 B2B-software-type company records from Ken internal data (industries: software development; technology, information and internet; computer and network security; data infrastructure and analytics; IT system custom software development; desktop, embedded and mobile computing software products), headquartered in the US, UK, Canada, Australia, Ireland or New Zealand, founded 1950-2018, with 11 to 5,000 people today. For each year one web archive crawl near mid-year (2026: a September crawl).

**Freeze.** This analysis uses the primary-crawl captures for 2019-2026, the point at which those crawls were complete for all 10,000 companies (2026-10-19). Valid primary captures by year: 2019 4,542, 2020 4,523, 2021 4,794, 2022 4,789, 2023 4,629, 2024 4,736, 2025 4,419, 2026 3,636. The 2014-2018 captures and the same-year fallback crawls were still being fetched and are not used, except for a fallback sensitivity (section 6). This study made no web archive requests of its own.

**Headcount.** The shared headcount history built for the Ken Research study *Did AI kill the SDR role?*: quarterly headcount per company, in total and by function, rebuilt from the start and end dates of roles in Ken internal data, 2015 Q1 to 2026 Q1. Functions used: `sdr_bdr` (sales and business development reps) and `account_exec` (account executives).

**Funding.** The shared funding-events extract built for the Ken Research study *When should you hire your first salesperson?*: dated raises from US SEC Form D filings matched to company records, plus the last round recorded on each company record.

## 3. Measuring the primary call to action

**Extraction.** The extractor reads each capture's raw HTML from the panel's archived pages (the panel's text table drops buttons and navigation). A fixed rule set lists the clickable elements (links, buttons, submit inputs) in document order, skipping cookie and consent banners, footers, login links and interface controls (menu, search, close, slider and language controls). The header is the `header`/`nav` region plus anything before the first `h1`; the hero runs from the first `h1` to the next `h2`, at most 60 elements.

**Classes.** Each element's text (and link target, for "Get started") is classified as demo/sales (demo, contact sales, talk to an expert, quote, book a call or meeting, contact us, a phone or email link), free/self-serve (start free, free trial, try free, sign up, create account, join, register, download or install the product) or other. Text over nine words, newsletter, webinar and event sign-ups, recorded or video demos and brochure downloads are not conversion CTAs.

**Primary conversion CTA** = the first demo/sales or free/self-serve CTA in the hero; if none, the last one in the header; otherwise `no_cta`. Pages with fewer than three clickable elements are `unreadable` (the page is built by scripts that the archive's crawler does not run) and are excluded from shares; their share is reported. **Page mix** records whether the header and hero contain demo/sales only, free/self-serve only, both, or neither.

**Valid capture**: HTTP 200 HTML, not a parked or holding page, page language English or undeclared.

**Validation.** Pages were drawn at random and labeled in a blind check (the "readers" below): each page's cleaned top-of-page markup was read against a written definition, without seeing the classifier output.

| Round | Rule | Capture years | Pages | Agreement | Cohen's kappa |
|---|---|---|---|---|---|
| v1 | First button in the hero (pre-registered) | 2023-2026 | 200 | 0.625 | 0.486 |
| v2 | Primary conversion CTA | 2023-2026 | 200 | 0.775 | 0.645 |
| v3 | Primary conversion CTA, refined (used) | 2023-2026 | 200 | 0.77 | 0.646 |
| v4 | Same rules as v3 | 2019-2020 | 100 | 0.65 | 0.486 |

The pre-registered rule ("first button in the hero") failed v1, mainly because a hero often shows "Learn more" next to "Book a demo" and readers took the conversion button as the main CTA. The measure was changed to the primary conversion CTA and validated on fresh pages. The pre-registered bar of 85% page-level agreement was not met in any round; the study uses the v3 rules and reports the agreement plainly. Reviews of the blind labels flagged a handful of questionable reference labels, so the agreement figures are measured against imperfect labels.

What the trend depends on is whether the class shares are right. On readable pages, v3 (2023-2026): classifier 55.6% demo/sales, 12.9% free/self-serve; readers 51.6% and 14.0%. v4 (2019-2020): classifier 51.8% and 12.0%; readers 51.6% and 18.3%. The classifier's error drifts: it undercounts free/self-serve on older pages and overcounts demo/sales on recent pages. Section 5 gives the trend adjusted for this.

## 4. Company-year table and subgroups

The company-year table has one row per company and year. Size = rebuilt headcount at Q4 of that year (2026: Q1, the last quarter available), in bands 1-10, 11-50, 51-200, 201+; a company with no rebuilt headcount row in that quarter (no profiles at the company then) counts as 0 people, band 1-10. Rebuilt headcount counts only people with profiles in the six covered countries and undercounts older years. Company age = year minus founding year (0-5, 6-10, 11+). Funding stage = whether any dated raise is on record by the end of that year. Subgroup cells with fewer than 50 companies are suppressed.

## 5. Trend results

| Year | Readable homepages | Demo/sales % | Free/self-serve % | No conversion CTA % | Shows both % | Unreadable % (of all) |
|---|---|---|---|---|---|---|
| 2019 | 3889 | 46.2 | 19.4 | 34.4 | 11.8 | 14.4 |
| 2020 | 3887 | 48.8 | 18.8 | 32.5 | 12.4 | 14.1 |
| 2021 | 4175 | 50.0 | 18.2 | 31.8 | 13.1 | 12.9 |
| 2022 | 4148 | 52.7 | 17.1 | 30.2 | 12.6 | 13.4 |
| 2023 | 3986 | 54.4 | 17.0 | 28.5 | 14.1 | 13.9 |
| 2024 | 4052 | 56.0 | 16.6 | 27.4 | 13.9 | 14.4 |
| 2025 | 3686 | 57.5 | 16.1 | 26.5 | 14.9 | 16.6 |
| 2026 | 3207 | 61.2 | 15.5 | 23.3 | 14.1 | 11.8 |

Demo/sales was the most common primary CTA in every year and rose in every year. Among homepages with a conversion CTA, demo/sales made up 70.4% in 2019 and 79.8% in 2026.

**Same companies every year** (645 companies with a readable homepage in all eight years):

| Year | Demo/sales % | Free/self-serve % | No conversion CTA % |
|---|---|---|---|
| 2019 | 49.0 | 24.5 | 26.5 |
| 2020 | 48.8 | 24.2 | 27.0 |
| 2021 | 49.0 | 25.0 | 26.0 |
| 2022 | 54.4 | 21.7 | 23.9 |
| 2023 | 55.2 | 21.9 | 22.9 |
| 2024 | 54.9 | 20.9 | 24.2 |
| 2025 | 56.7 | 21.6 | 21.7 |
| 2026 | 60.0 | 19.5 | 20.5 |

**By size** (selected years):

| Year | People that year | Readable homepages | Demo/sales % | Free/self-serve % |
|---|---|---|---|---|
| 2019 | 1-10 | 2626 | 42.8 | 20.2 |
| 2019 | 11-50 | 965 | 52.5 | 17.3 |
| 2019 | 201+ | 60 | 43.3 | 25.0 |
| 2019 | 51-200 | 238 | 59.2 | 17.6 |
| 2022 | 1-10 | 2556 | 49.6 | 17.9 |
| 2022 | 11-50 | 1134 | 57.0 | 15.0 |
| 2022 | 201+ | 101 | 60.4 | 19.8 |
| 2022 | 51-200 | 357 | 59.4 | 17.6 |
| 2026 | 1-10 | 1829 | 54.9 | 18.1 |
| 2026 | 11-50 | 962 | 68.2 | 12.0 |
| 2026 | 201+ | 99 | 67.7 | 23.2 |
| 2026 | 51-200 | 317 | 73.8 | 9.1 |

**By company age:**

| Year | Company age | Readable homepages | Demo/sales % | Free/self-serve % |
|---|---|---|---|---|
| 2019 | 0-5 | 1132 | 40.4 | 26.1 |
| 2019 | 11+ | 1827 | 50.0 | 14.7 |
| 2019 | 6-10 | 930 | 45.9 | 20.5 |
| 2022 | 0-5 | 613 | 48.1 | 20.9 |
| 2022 | 11+ | 2266 | 55.0 | 13.5 |
| 2022 | 6-10 | 1269 | 50.8 | 21.7 |
| 2026 | 11+ | 2487 | 62.1 | 14.4 |
| 2026 | 6-10 | 720 | 58.1 | 19.3 |

Companies founded in 2019 or later are not in the panel, so the 0-5 band empties out after 2023.

**By funding:**

| Year | Dated raise on record by then | Readable homepages | Demo/sales % | Free/self-serve % |
|---|---|---|---|---|
| 2019 | false | 3313 | 46.4 | 18.4 |
| 2019 | true | 576 | 45.3 | 25.3 |
| 2022 | false | 3235 | 53.1 | 15.5 |
| 2022 | true | 913 | 51.3 | 23.0 |
| 2026 | false | 2306 | 60.3 | 13.9 |
| 2026 | true | 901 | 63.5 | 19.6 |

**Measurement-adjusted endpoints**: each endpoint share is scaled by the reader-to-classifier ratio from the validation of its own period (v4 for 2019, v3 for 2026). Point estimates from small samples; the adjusted shares need not sum to 100.

| Year | Class | Measured % | Check | Readers % | Classifier % | Ratio | Adjusted % |
|---|---|---|---|---|---|---|---|
| 2019 | Demo/sales | 46.2 | v4 | 51.6 | 51.8 | 0.996 | 46.0 |
| 2019 | Free/self-serve | 19.4 | v4 | 18.3 | 12.0 | 1.525 | 29.6 |
| 2019 | No conversion CTA | 34.4 | v4 | 30.1 | 36.1 | 0.834 | 28.7 |
| 2026 | Demo/sales | 61.2 | v3 | 51.6 | 55.6 | 0.928 | 56.8 |
| 2026 | Free/self-serve | 15.5 | v3 | 14.0 | 12.9 | 1.085 | 16.8 |
| 2026 | No conversion CTA | 23.3 | v3 | 34.4 | 31.5 | 1.092 | 25.4 |

After adjustment, demo/sales rises by about 10.8 points and free/self-serve falls by about 12.8 points. The direction of both trends is unchanged; the measured demo rise is probably overstated and the measured self-serve fall understated.

## 6. Sensitivity: fallback crawls

Using the fallback crawl where the primary crawl has no valid capture (fallbacks fetched by the freeze date only), the series barely moves:

| Year | Readable homepages | Demo/sales % | Free/self-serve % |
|---|---|---|---|
| 2019 | 3889 | 46.2 | 19.4 |
| 2020 | 3887 | 48.8 | 18.8 |
| 2021 | 4723 | 49.5 | 17.9 |
| 2022 | 4723 | 52.7 | 16.9 |
| 2023 | 4550 | 53.9 | 17.0 |
| 2024 | 4521 | 55.8 | 15.9 |
| 2025 | 4087 | 57.4 | 16.2 |
| 2026 | 3973 | 60.6 | 15.1 |

## 7. Switchers and growth

**Design.** Switch year t = the first year whose primary conversion CTA differs from at least two earlier readable captures that all agree; it counts only if the next readable capture (if any) keeps the new class. Stayers have the old class in at least two earlier readable captures, in no other class through t+1, and readable old-class captures in both t and t+1 (the pre-registered wording; the AI homepages study's rule, a capture in t or t+1, is a sensitivity below). With captures from 2019, switches can be dated 2021-2025. The matched event study follows the AI homepages study: anchor = Q4 of t-1; exact strata of industry group x founding bucket x headcount band at the anchor; nearest neighbour on the two prior four-quarter log growth windows, up to 3 stayers with replacement. Pre-registered: caliper 0.10 log points, at least 5 people at the anchor. Exploratory (not pre-registered): caliper 0.20, at least 3 people. Pre-trend check: joint test of the -8..-1 quarter gaps has p >= 0.05, or every gap is within 0.02 log points. Bootstrap 95% intervals, 1,000 draws, seed 59. A synthetic test with planted effects recovered them (headcount +0.087 log points vs 0.0875 planted at 8 quarters; SDR -0.264 vs -0.2625).

| Switch | Switchers (2021-2025) | Matched, pre-registered | Matched, two-capture confirmation | Matched, exploratory | Stayers used |
|---|---|---|---|---|---|
| Demo/sales to free/self-serve | 104 | <50 | <50 | 70 | 151 |
| Free/self-serve to demo/sales | 146 | <50 | <50 | 61 | 79 |

Under the pre-registered design fewer than 50 switchers were matched in each direction, so the pre-registered growth and sales-team comparisons are **not assessed**. The exploratory setting reaches more than 50:

| Switch | Pre-trend (headcount) | 4-quarter median gap, pp [95% CI] | 8-quarter median gap, pp [95% CI] | 8-quarter mean gap, log points [95% CI] |
|---|---|---|---|---|
| Demo/sales to free/self-serve | p = 0.0355, max gap 0.033 (fails) | 3.45 [-0.65, 4.58] (n=70) | -2.06 [-19.69, 4.07] (n=55) | -0.0371 [-0.1735, 0.1041] |
| Free/self-serve to demo/sales | p = 0.033, max gap 0.0587 (fails) | 0.0 [-7.69, 5.13] (n=61) | -4.35 [-16.67, 9.52] (n=53) | -0.0708 [-0.2314, 0.0663] |

| Switch | SDR/BDR, 8 quarters, log(1+n) gap [95% CI] | AE, 8 quarters, log(1+n) gap [95% CI] | Any dated raise within 24 months, % (switchers vs stayers) | Site gone by 2025-26 |
|---|---|---|---|---|
| Demo/sales to free/self-serve | -0.0185 [-0.1414, 0.0795] | 0.08 [-0.0473, 0.2283] | 10.91 vs 7.27; gap 3.64 [-3.95, 13.64] | <50 cohorts 2020-2023 |
| Free/self-serve to demo/sales | 0.0977 [-0.0284, 0.2308] | -0.1227 [-0.3228, 0.0444] | 11.32 vs 10.38; gap 0.94 [-11.64, 12.26] | <50 cohorts 2020-2023 |

**Sensitivity to the stayer rule:**

| Switch | Exploratory setting with the AI homepages study's stayer rule (capture in t or t+1) |
|---|---|
| Demo/sales to free/self-serve | 72 matched; pre-trend p = 0.0966 (passes); 8-quarter median gap -0.89 [-17.27, 2.54] |
| Free/self-serve to demo/sales | 65 matched; pre-trend p = 0.3999 (passes); 8-quarter median gap 3.85 [-12.19, 12.15] |

Under the pre-registered stayer rule the exploratory comparison **fails the pre-trend check** in both directions: switchers and matched stayers were not on parallel paths before the switch (free-to-demo switchers grew more slowly in the two years before). Under the AI homepages study's stayer rule it passes, but the free-to-demo gap changes sign. Neither direction shows a clear growth, sales-team or funding gap, the intervals are wide, and the comparison is fragile, so it supports no conclusion about growth. At best these are associations around the switch year in a matched comparison, not causal effects.

## 8. Hypotheses against the results

1. **Trend: largely supported.** Demo/sales is the most common primary CTA in every year (also after measurement adjustment), free/self-serve falls from 2019 and pages with both kinds rise. Whether free/self-serve peaked between 2019 and 2022 cannot be judged without the pre-2019 years; within 2019-2026 its highest year is 2019. Larger companies lean to demo/sales (1-10 people lowest in every year), and companies over 10 years old more than younger ones; the 201+ band is small and noisy.
2. **Growth after a demo-to-free switch: not assessed** under the pre-registered design (fewer than 50 matched). The exploratory comparison fails the pre-trend check and its interval is wide, so the prediction (a gap within plus or minus 3 points) can be neither confirmed nor rejected.
3. **Sales-team shape: not assessed** under the pre-registered design. Exploratory: no detectable SDR/BDR change after a demo-to-free switch, which does not support the predicted decline, but the same pre-trend failure applies; free-to-demo switchers show no clear SDR/BDR or AE gain.

## 9. Verification

An independent re-analysis, written from the study plan and the data extracts alone without seeing our code, recomputed every headline number with its own code (trend, same-company panel, size bands, switch counts, matched comparisons, validation agreement, adjusted shares). All matched once two definitions were aligned: the stayer rule (the study code first used the AI homepages study's rule; it now follows the pre-registered wording, which changed the exploratory results reported here) and the size band of company-years with no headcount row. Re-extracting 300 random pages from the raw captures reproduced the stored classification on 300 of 300.

## 10. Limitations

- The classifier agrees with the readers on 77% of 2023-2026 pages and 65% of 2019-2020 pages, below the pre-registered 85%, and its error drifts over time (section 3). Shares are adjusted in section 5; switch detection is noisier than the shares, which tends to push growth gaps towards zero and adds spurious switches.
- The panel is today's company records: companies that closed or were acquired before 2026 are missing (survivor bias), and all companies were founded by 2018, so the panel ages over time. The same-company and by-age tables address this.
- About one homepage in seven is unreadable because it is built by scripts; if script-built sites differ systematically, shares among readable pages are biased.
- Ken's web archive does not capture every site every year; a valid primary capture exists for 36% to 48% of panel companies per year.
- 2026 comes from a September crawl and an index that holds only status-200 pages, three months later than the other years.
- Industry labels do not separate B2B from consumer software; "technology, information and internet" in particular includes consumer apps. Many panel companies are IT services firms, whose main CTA is "Contact us".
- Rebuilt headcount covers people in six countries with profiles, undercounts older years and the most recent quarters, and is flat (zero growth) for many small companies.
- Funding covers US Form D filings and the last round on company records only.
- The 2014-2018 years are not included; they would show whether self-serve peaked before 2019 and add earlier switch cohorts.

## 11. Sources

- Ken's web archive, homepage captures June 2019, June 2020, June 2021, July 2022, June 2023, June 2024, June 2025 and September 2026.
- Ken internal data: company records and work histories (Ken's dataset of public professional profiles, work histories and company records).
- US SEC Form D data sets, via Ken's shared funding-events extract.
- Ken Research shared extracts: the homepage panel (the AI homepages study), the headcount history (the SDR study) and the funding events (the first sales hire study). Matching design from the AI homepages study.