Methodology — how The AI Index scores companies

Methodology

How The AI Index scores companies

Two cohorts. Two composite scores. Updated weekly. Below is exactly how each is computed.

In short: we track two groups of companies — AI-native challengers and legacy SaaS incumbents — and score how fast each is rising or falling, using real headcount, live hiring momentum, funding, and layoff data. Scores re-run every Monday. Here's exactly how the math works.

August 2026 — the LinkedIn-primary era. The index originally ran on People Data Labs' monthly headcount history. That backbone is now frozen (16 years of history, through July 2026) and the live signal is the weekly LinkedIn-tracked headcount series — validated every Monday, never more than a week old. Sections below marked “frozen” describe data that no longer refreshes; everything else is live.

Rise score (0–100)

For each AI-native company in the Risers cohort:

rise_score =  0.50 · z(linkedin_momentum_28d_%)   // weekly LinkedIn series, short window
            + 0.35 · z(linkedin_momentum_84d_%)   // trend window
            + 0.15 · z(funding_events_90d, weighted_by_amount)   // frozen Jul 2026

then min-max normalize to 0–100 across the cohort for the week.

Cut score (0–100)

For each legacy SaaS incumbent in the Cuts cohort:

cut_score  =  0.40 · z(layoff_severity_90d)          // pct × recency × ai_attribution (news-sourced)
            + 0.30 · z(-linkedin_momentum_28d_%)     // weekly LinkedIn shrinkage
            + 0.20 · z(-linkedin_momentum_84d_%)     // trend window
            + 0.10 · at_risk_category_bonus

then min-max normalize to 0–100 across the cohort for the week.

Column note for API users: headcount_30d_pct and headcount_90d_pct carry the LinkedIn 28-day and 84-day momentum respectively (the field names predate the era change). headcount is the latest LinkedIn-tracked count — a different ruler from the frozen PDL history, so levels across the two series are not directly comparable. When the weekly series is too thin to trust (a scrape outage), the board holds the prior week's ranks rather than re-ranking on noise.

Live headcount momentum (LinkedIn)

PDL's headcount data is authoritative but it lags ~2 months and only moves monthly — so on its own, the index would sit still between refreshes. To make the board reflect who is hiring or shrinking now, we layer in a second, faster headcount signal: a weekly LinkedIn-derived headcount series, tracked independently for every tracked company. Its week-over-week change is thelinkedin_momentum_%term in both formulas above.

It began in June 2026 at a deliberately small weight and was promoted to the primary signal in August 2026 when the PDL backbone froze. Two principles keep it honest:

Source-separated. The LinkedIn series is never mixed into the PDL headcount history. LinkedIn counts public profiles, PDL counts employees — different rulers — so they live in separate series, and only the week-over-week change feeds the score, never the absolute level.
Validated, accepted-only. Every weekly reading is checked before it can score: zero or implausible values, namesake-company mismatches, and impossible week-over-week jumps are rejected. A company with no accepted reading simply gets no LinkedIn term that week, rather than a fabricated one.

Why a company's rank can move while its headcount looks flat: the headcount % shown on a company page is PDL's slower, official figure. The composite rank also reflects the faster LinkedIn momentum trend — so the two can legitimately diverge for a week or two until PDL catches up.

The pressure vs. efficiency split

A falling headcount at a legacy SaaS company isn't always the same story. Sometimes it's a company being disrupted by AI-native competitors. Sometimes it's a company using AI to run leaner — cutting roles while revenue grows. Both end up on the Cuts side of the index, but they mean very different things for investors, employees, and the market.

For every Cut, we classify the last 180 days of layoff announcements:

Under pressure — layoffs announced with no AI-efficiency rationale in the majority of cases. Genuine AI disruption victims.
AI efficiency — majority of recent layoffs explicitly attribute the cuts to AI productivity gains. Company restructuring around AI, not being displaced by it.
Attrition only — contracting via hiring freezes or attrition, with no announced layoffs.

Next: weight the composite by revenue-per-employee trajectory (PDL inferred revenue, already ingested) and 30-day stock moves for public companies. A company shedding roles while revenue rises is a very different investment case from one where both are falling.

Does the index predict? (backtested July 2026)

We backtested the index's headcount backbone — the trailing-90-day change signal, deliberately excluding layoff terms so layoffs can't predict themselves — across Jan 2024 – Feb 2026 at monthly grain: 2,574 company-months. The question: do the companies the signal flags keep moving the same way over the following 90 days?

CUTS — P(sheds ≥5% of staff in the NEXT 90 days)
  cohort base rate                     6.8%
  flagged (worst-quintile decline)    23.1%   ×3.4 lift
  flagged two months running          24.7%   ×3.7 lift
  everyone else on the board           2.6%

RISERS — flagged growers averaged +28.7% next quarter
  vs +8.7% for the rest (3.3× the growth rate)

Honest limits: this is within our tracked cohort (selected in 2026, with hindsight), on PDL history as revised today, with overlapping 90-day windows (so observations aren't independent), and it validates the headcount backbone rather than every published rank. Which is why we don't ask you to trust the backtest: since July 2026 the flagged set is logged forward, in public, every Monday — and scored when the outcome lands. See the live track record.

Guardrails

Headcount % changes are capped at ±50% to limit single-data-error blast radius.
The LinkedIn momentum signal is validated per weekly reading — zeros, namesake mismatches, and implausible jumps are rejected — so one bad scrape can't move a score.
Companies with <90 days of tracked history get a null score, not a forced zero.
Companies with <50 employees are flagged "small base" and shown in a separate sub-bucket so noise doesn't dominate.

Data sources

Live headcount (primary): a weekly LinkedIn-tracked headcount series, banked every Monday, validated per reading (accepted-only). Drives the momentum terms, the displayed current headcount, and — from August 2026 — the track record's measurements.
Historical headcount (frozen): People Data Labs Company Enrichment — 16 years of monthly history through July 2026, when the subscription ended. Kept source-separated from the LinkedIn series (different rulers; levels are not directly comparable). Powers the historical charts and the published backtest.
Layoffs (live): extracted from the continuously-classified news corpus, with AI-attribution where the reporting supports it. The PDL-derived departure-spike detector contributed through July 2026 and is now frozen with the rest of that backbone.
Funding (frozen July 2026): PDLfunding_details — amount, date, round type, last refreshed in the final July 2026 harvest. Its trailing-90-day scoring term empties naturally as events age out; the history remains on company pages, labeled.

Refresh cadence: the LinkedIn-tracked series banks every Monday and scoring re-runs the same morning, folding in the fresh momentum plus any layoff events extracted from the news corpus during the week. The frozen PDL backbone does not refresh; occasional calibration re-harvests may update the frozen layers and are noted here when they happen. The "updated weekly" line in the footer refers to the scoring pass.