Methodology
The momentum score answers one question: compared with the other companies we track, how strongly are this company's observable signals moving right now? It is not a valuation, a quality rating or investment advice.
1. Which companies are tracked
We track 89 AI agent companies in five lists. A company qualifies when its main product is an AI agent, an agent platform or infrastructure built for agents, and it is privately held. Companies confirmed acquired or shut down by a dated public source are dropped from the rankings and keep a page that says so. When a data vendor's acquired or shut-down flag contradicts a dated public source, the source wins and is linked on the company page. A company can sit in more than one list (for example a category list and the relaunch list). The universe is reviewed every week; additions are noted in the archive.
2. The five signals
- Headcount change, 3 months. Percent change in employees, from Aviato company data (LinkedIn-derived headcount history). As-of date shown on each company page.
- Web traffic change, 3 months. Percent change in estimated monthly website visits, from Aviato company data.
- Tranco rank movement, about 30 days. Change in the domain's position in the Tranco research ranking of top sites, comparing the latest daily list with the list about 30 days earlier. A domain outside the ranking has no value for this signal.
- Recency of the last funding round. Days since the most recent dated round. Rounds come from PredictLeads financing events (with the article each was found in) or Aviato company data; the source link is shown next to every round.
- Launch and discussion activity, 90 days. Number of Hacker News stories in the last 90 days whose link points to the company's domain (Algolia HN API).
Entity checks
Data vendors sometimes match a domain to the wrong company. When a vendor record clearly belongs to a different company (different name, website or LinkedIn page), we drop that vendor's numbers for the company and show n/a, with a note on its page, rather than publish another company's headcount or traffic.
Funding rounds
The last round is the most recent dated round in either source. PredictLeads records the date it found an event separately from the round date; only real round dates are used. The amount is shown only when a second, independent record (the other provider, or the linked article's headline) agrees within 10%; otherwise the amount shows n/a.
3. The score
For each signal, every company with a value gets a percentile rank against all tracked, active companies (0 = lowest, 100 = highest; ties share the midpoint). The momentum score is the plain average of the percentiles the company has. All signals have equal weight; we do not tune weights.
A company needs at least 3 of the 5 signals to be ranked. Companies below that are still listed, alphabetically, as unranked. Missing data is shown as n/a and never filled in or estimated.
Ranks within a list sort by score (ties alphabetical). Because percentiles are computed across the whole universe, scores are comparable between lists.
4. What the score does not capture
- Revenue, retention, customer quality or product quality.
- Private signals such as pipeline, burn or term sheets.
- Lag: headcount and traffic data can trail reality by weeks; each figure shows its own as-of date. Large companies move slowly in percent terms, so small teams can swing further in either direction.
- Recently relaunched companies: trailing signals mix the old and new product. Read their trend against the relaunch date on the company page.
5. Editorial rules
- No company pays to be included, ranked or described. No company can edit its numbers; corrections need a public source and change only the field in question.
- Descriptions are one neutral line about what the product does. The "What the data says" notes are generated by the same rules for every company.
- Every number is stored with its source and the date it was fetched, in the data file behind the site.
6. Update schedule
Signals are refetched and the lists rebuilt once a week. Each weekly snapshot is kept in the archive. This page records every change to the method; the current version is 1.0.
Change log
- v1.0 (November 12, 2026): first published version.