How changes are found

StartupHistory follows the homepages of venture-funded startups. Each record is built only from the company's own public website, never from model memory or press coverage.

Sources

• History: homepage versions kept by the Internet Archive, usually months apart.
• Live checks: our own visits to the homepage. New companies are checked weekly; unchanged pages are checked less often (up to every 30 days) and changed pages again within 3 days.

First recorded website

A record can start with one capture. Its messaging, labeled customer and investor logos, prices, compliance claims and funding statements are shown as observations on that date, never as newly acquired customers or launches. We use the earliest visible capture with readable facts and never fill it with facts from a later date. Baselines do not contribute to change counts or topic trends.

What the pages say

Most changes are read straight from the company's own pages and quoted: the homepage headline, customer and investor logo walls (only when the page labels them, e.g. “Trusted by”), traction numbers (“25,000+ companies”), prices, free sign-up or demo-only buttons, compliance badges, funding and partnership lines, and the careers page or job board. Live checks also read the pricing, customers and careers pages. A change is a difference between two versions of the same page; logos that did not load are never counted as removed.

How it describes itself

Each record shows the company's homepage headline over time, quoted, with the dates each version was seen. A new line starts when the headline is rewritten (fewer than half of its words kept). This is a record of messaging, not a judgement: a rewritten headline can be a new market, a new product or just new copy.

Interpreting a change

A structured classifier also reads two consecutive versions and answers fixed questions (same company? comparable? did the product or target customer change?). Its reading is published only as a hedged sentence attached to a quoted headline rewrite, when the pages are comparable and the answer is “yes” with probability of at least 0.85. StartupHistory does not label pivots: in an audit most classifier “pivots” were rewordings, and a new pitch cannot be told from a slogan refresh by its words alone. Every change is a fixed sentence, never free-form model text, and the JSON record says how each was found (basis).

Dating

A change happened somewhere between the two dates shown: the last version that still showed the old page and the first that shows the new one. Live checks narrow this to about a week.

Version and corrections

Current method: messaging-v1. Records can be withdrawn or corrected; when the method changes, the version changes with it.