# Methodology: Half of B2B software homepages now say AI. Did it pay off?

By Cristian Frunze, Founder of Ken · Oct 16, 2026 · [Ken Research](/research) No. 9 · Websites and pricing, AI and work, Founders and startups

The full methodology behind [Half of B2B software homepages now say AI. Did it pay off?](/research/ai-homepages.md): the data, every method, every number and every limitation. The same text as the [methodology PDF](/research/ai-homepages.pdf).

Data: Ken internal data and Ken's web archive. All numbers below come from aggregate tables.

## Question

What share of B2B software homepages mention AI, year by year from 2019 to 2026? Did companies that put AI on their homepage grow headcount or raise money faster afterwards than similar companies that did not?

## Data

- **Companies.** Ken internal data company records (Ken's dataset of public professional profiles, work histories and company records). The frame is every record with an industry in the software and internet group (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), headquarters in the United States, United Kingdom, Canada, Australia, Ireland or New Zealand, a website domain, a founding year from 1950 to 2018 and a current size band of 11 to 5,000 employees: 58,208 records. After dropping 406 records with no usable domain (or a social network or other platform domain) and 6,778 records whose domain is shared with another record, 51,024 remain. The panel is a seeded simple random sample of 10,000 of them (US 7,164, GB 1,382, CA 808, AU 479, IE 95, NZ 72).
- **Homepages.** Ken's web archive, one crawl per year nearest mid-year: June 2019, June 2020, June 2021, July 2022, June 2023, June 2024, June 2025 and September 2026. For each company and crawl we look up the capture of the domain's root page (`/`, with or without `www` and https), keep the newest status-200 HTML capture, and follow one redirect within the same domain. A redirect to another domain is recorded as moved. Pages are parsed to visible text (title, meta description, H1, H2/H3, main text, and navigation/footer text kept separately). Parked pages (domain for sale, suspended, coming soon) are excluded.
- **Headcount.** A shared headcount history (built for the Ken Research study *Did AI kill the SDR role?* on 2026-10-17): quarterly headcount per company from 2015 Q1 to 2026 Q1, counting people whose role at the company spans the quarter end, rebuilt from work histories in Ken internal data.
- **Funding.** A shared table of funding events (built for the Ken Research study *When should you hire your first salesperson?*): US private raises from the SEC's quarterly Form D data sets (2008-2026, funds, SPVs, real-estate and financial issuers removed), matched to company records by normalized name and state, plus the dated last round on each company record.

Capture coverage of the panel (primary crawl lookups):

| Year | Captured | Not captured that crawl | Moved to another domain | Classified (not parked) |
|---|---|---|---|---|
| 2019 | 4,598 | 5,402 | 0 | 4,592 |
| 2020 | 4,597 | 5,233 | 170 | 4,579 |
| 2021 | 4,868 | 4,921 | 211 | 4,847 |
| 2022 | 4,867 | 4,899 | 234 | 4,846 |
| 2023 | 4,709 | 5,085 | 206 | 4,680 |
| 2024 | 4,832 | 4,921 | 247 | 4,804 |
| 2025 | 4,493 | 5,262 | 245 | 4,484 |
| 2026 | 3,715 | 6,285 | 0 | 3,704 |

Moves to another domain cannot be seen in 2019 or 2026, so those two years show none.

## Classifying homepages

1. **Terms.** "AI" or "A.I." as an upper-case standalone word; artificial intelligence; machine learning; deep learning; neural network; generative AI; GenAI; LLM; large language model; ChatGPT; GPT-3, 3.5, 4, 4o, 5; AI-powered, -driven, -native, -first, -enabled, -based, -assisted; AI agent, assistant, copilot. "Copilot", "agentic" and "AI agent" alone are treated as ambiguous. A word inside a longer matched phrase counts once.
2. **Labels.** *Mentions AI* (broad): any term anywhere on the page, including navigation and footer. *AI as the main claim* (strict): a term in the page title, meta description or H1, or three or more terms in the first 1,500 characters of the main text. A sensitivity version of broad excludes mentions found only in navigation or footer.
3. **LLM check.** Borderline captures (a broad match with at most one unambiguous term, or a match only in navigation or footer) were sent to a small language model at temperature 0 with short snippets only: title, meta description and up to three 220-character windows around the matches. It answered none, mention or core. A "none" removed the match; the model never added a label. It checked a seeded sample of up to 150 borderline captures per year in each of two random halves of the panel. Veto rates:

| Year | Borderline captures | Checked | Veto rate (%) |
|---|---|---|---|
| 2019 | 174 | 174 | 5.2 |
| 2020 | 216 | 216 | 4.6 |
| 2021 | 249 | 249 | 3.6 |
| 2022 | 252 | 252 | 4.0 |
| 2023 | 280 | 280 | 3.9 |
| 2024 | 327 | 300 | 2.0 |
| 2025 | 339 | 300 | 3.0 |
| 2026 | 360 | 355 | 2.5 |
| all | 2,197 | 2,126 | 3.4 |

**Agreement.** On a separate seeded random sample of 300 captures from all years, the model labelled each page from its title, description, H1, first 600 characters and match windows. Agreement with the word rules: broad 98.0% (Cohen's kappa 0.944), strict 96.3% (kappa 0.756); the rules labelled 9.3% of the sample strict against 7.0% for the model, so the strict rule leans slightly generous.

## Adoption results

Share of homepages (%) with 95% Wilson intervals. "Balanced" is the 878 companies with a readable homepage in all eight years.

| Year | Companies | Mentions AI | AI main claim | Mentions AI, no nav/footer-only | Balanced: mentions | Balanced: main claim |
|---|---|---|---|---|---|---|
| 2019 | 4,592 | 11.7 (10.8-12.6) | 4.6 (4.0-5.3) | 11.3 | 13.4 | 4.9 |
| 2020 | 4,579 | 14.0 (13.0-15.0) | 5.3 (4.7-6.0) | 13.4 | 15.4 | 4.9 |
| 2021 | 4,847 | 15.4 (14.4-16.4) | 5.7 (5.1-6.4) | 14.8 | 16.7 | 4.6 |
| 2022 | 4,846 | 15.8 (14.8-16.9) | 5.3 (4.7-6.0) | 15.2 | 17.8 | 4.7 |
| 2023 | 4,680 | 19.3 (18.2-20.5) | 6.9 (6.2-7.7) | 18.5 | 21.0 | 5.9 |
| 2024 | 4,804 | 26.3 (25.1-27.6) | 10.7 (9.8-11.6) | 25.2 | 30.2 | 10.7 |
| 2025 | 4,484 | 33.0 (31.7-34.4) | 15.2 (14.2-16.3) | 31.4 | 37.2 | 17.0 |
| 2026 | 3,704 | 49.0 (47.4-50.6) | 25.8 (24.4-27.2) | 46.6 | 52.2 | 26.3 |

Excluding thin pages (under 50 words of visible text, often script-rendered sites):

| Year | Mentions AI | AI main claim |
|---|---|---|
| 2019 | 12.6 | 4.9 |
| 2020 | 14.7 | 5.5 |
| 2021 | 16.4 | 6.0 |
| 2022 | 17.1 | 5.6 |
| 2023 | 20.9 | 7.4 |
| 2024 | 28.6 | 11.5 |
| 2025 | 36.6 | 16.8 |
| 2026 | 51.7 | 26.9 |

## Linking the change to growth

- **Adopters.** A company's change year is the first year its homepage is labelled AI, after at least two earlier readable homepages labelled not AI, with the change in 2020-2025. **Controls** had at least two earlier non-AI homepages, were still not AI through the year after, and had a readable homepage in the change year or the next.
- **Matching.** For each change-year cohort, adopters are matched to up to three controls, with replacement, inside exact strata of industry group (software; internet; security, data and other), founding period (up to 2000, 2001-2010, 2011-2014, 2015-2018) and rebuilt headcount band at the anchor (1-10, 11-50, 51-200, 201+), on both prior four-quarter headcount growth windows, within 0.10 log points on each. The anchor is the fourth quarter of the year before the change; companies need at least 5 people at the anchor and headcount two years before it.
- **Event study.** Log headcount relative to the anchor, adopter minus the mean of its controls, for quarters -8 to +8, averaged over adopters; 95% intervals from 1,000 bootstrap resamples of adopters. Pre-trend check: joint Wald test that the eight pre-period gaps are zero, and the largest absolute pre-period gap (passes when p >= 0.05 or every gap is within 2%).
- **Outcomes.** Headcount growth over 4 quarters (cohorts 2020-2025) and 8 quarters (cohorts 2020-2024); any dated raise from 1 January of the change year to 31 December of the next year (cohorts 2020-2024); no readable homepage in the 2025 or 2026 crawls (cohorts 2020-2023).

Matched adopters: strict 260 (cohorts 2020: <50, 2021: <50, 2022: <50, 2023: <50, 2024: 83, 2025: 120), broad 500 (cohorts 2020: <50, 2021: <50, 2022: 55, 2023: 72, 2024: 166, 2025: 160). Cohorts under 50 are not reported separately but stay in the pooled estimates. Pre-trend: strict p = 0.597, largest pre-period gap 0.9% (passes); broad p = 0.7192, largest pre-period gap 0.4% (passes).

Headcount gap, adopters minus matched controls (mean gap converted from log points to %; medians in %; 95% intervals):

| Definition | Horizon | Adopters | Mean gap (%) | Median growth, adopters (%) | Median growth, controls (%) | Median gap (points) |
|---|---|---|---|---|---|---|
| strict | 4 quarters | 260 | -1.3 (-5.1 to 2.7) | 0.0 | -2.5 | 2.5 (-3.8 to 4.9) |
| strict | 8 quarters | 140 | -8.2 (-16.9 to 1.3) | -6.9 | -4.3 | -2.6 (-8.6 to 5.5) |
| broad | 4 quarters | 500 | -2.1 (-4.9 to 0.4) | 0.0 | -0.6 | 0.6 (-0.3 to 1.7) |
| broad | 8 quarters | 340 | -3.4 (-8.7 to 1.7) | -1.4 | -1.8 | 0.4 (-6.0 to 3.9) |

Before matching (selection check): strict, 334 adopters; prior four-quarter growth versus the median non-adopter of the same cohort: median gap 0.000 log points (mean 0.002); next eight quarters: median gap -0.095 log points (186 adopters); broad, 682 adopters; prior four-quarter growth versus the median non-adopter of the same cohort: median gap 0.000 log points (mean 0.028); next eight quarters: median gap -0.053 log points (464 adopters).

Funding within 24 months:

| Definition | Adopters | Raised, adopters (%) | Raised, controls (%) | Gap (points) |
|---|---|---|---|---|
| strict | 140 | 9.3 | 8.4 | 0.8 (-5.4 to 6.4) |
| broad | 340 | 7.9 | 7.1 | 0.9 (-2.3 to 4.5) |

Site survival: 10.9% of adopters (broad definition, cohorts 2020-2023, 174 companies) had no readable homepage in the 2025 or 2026 crawls, against 16.4% of matched companies (gap -5.5 points, 95% interval -11.1 to 0.6). On the strict definition: 12.3% against 15.2% (57 companies).

## Limitations

- **Correlation, not cause.** The design is a matched comparison with an event study. It rules out that adopters were already growing faster in the two years before the change (the pre-trend check), but not that something else changed at the same time, such as a pivot, new funding plans or a market shift.
- **B2B is inferred from industry labels.** The industry labels do not separate B2B from consumer software; "technology, information and internet" includes consumer apps.
- **Survivor frame.** The sample is drawn from today's company records, so companies that disappeared before 2026 are missing, especially in early years. Size band and founding year are as recorded today.
- **Crawl coverage.** Ken's web archive reads about half of these homepages in any one crawl, small sites less often. Sites that block crawlers are missing. The 2026 point is a September crawl, about three months later in the year than the others, so 2025 to 2026 spans about 15 months.
- **Rebuilt headcount** counts only people whose current profile is in the six countries, only current profiles, and undercounts the latest quarters because people add new jobs months late. Both groups are measured the same way, so this mostly cancels in the comparison, but levels are low and small companies move in steps of one person.
- **Funding** covers US raises filed on Form D and matched by name and state, plus the last round on company records; raises before a later last round can be missed. It undercounts for both groups alike.
- **Word rules.** A homepage can say AI without the product using it, and the reverse. The two definitions bracket this; the model check found the rules rarely wrong on mentions and slightly generous on main claims.
- **Small samples for growth.** The growth estimates rest on hundreds of matched adopters, most of them from 2024 and 2025, so they describe the generative AI wave more than earlier years.

## Sources

- Ken internal data: company records and work histories (Ken's dataset of public professional profiles, work histories and company records), snapshot 2026-04-30.
- Ken's web archive, crawls June 2019, June 2020, June 2021, July 2022, June 2023, June 2024, June 2025 and September 2026. 
- U.S. Securities and Exchange Commission, Form D data sets. https://www.sec.gov/data-research/sec-markets-data/form-d-data-sets
- OpenAI, "Introducing ChatGPT", 30 November 2022. https://openai.com/index/chatgpt/
- Ken Research shared extracts: the headcount history (built for the SDR study) and the funding events table (built for the first sales hire study).