All data in this report is drawn directly from Lusha’s live signal database — monthly headcount tracking, executive-move events, and funding signals for 25 named AI-native companies, January 1 to June 22, 2026. This is a focused sample of named companies, not a full population census. Source: Lusha live signal database, pulled June 22, 2026.
Every data vendor tells you records decay. Almost nobody shows you what that actually looks like on real accounts, in real time, with the numbers attached. So we ran a simple test on our own signal data: take 25 companies, freeze what a CRM would have known about them on January 1, and measure what was still true by late June.
The aggregate number is striking, but it’s not the most useful finding. The useful finding is how the decay happened — because the three patterns we found fail in different ways, and the most damaging one is invisible to the way most teams monitor their accounts.
The numbers at a glance
Part One: The slow leak — decay that never trips an alert
Perplexity’s headcount fell from 1,633 in January to 1,437 by May — a 12% net contraction, with declines in four of the five months. Here’s the part that matters: not one of those monthly declines exceeded 5%.
A team monitoring this account with a standard alert threshold — “notify me if headcount moves 10%” — would have received nothing. Five alert checks, five passes, and at the end of it the account is 12% smaller and whatever growth assumptions the January plan made are wrong. And this wasn’t a story anyone could have caught in the news instead: no layoff was announced, and public layoff trackers list no known layoffs for the company in this period. The slow leak is the most dangerous decay pattern precisely because it’s built to slip under both alert thresholds and press coverage: no single data point looks like news, and the trend only exists if something is watching the trend.
Worth saying plainly: nothing here means Perplexity is in trouble — headcount falls for many reasons, including deliberate ones, and this report has no visibility into why. What it means is narrower and more useful: any account plan that assumed growth was wrong within a quarter, and no threshold alert would have said so.
Part Two: The step change — when a record snaps back to reality
Cursor’s tracked headcount record shows 53 employees in January 2026. One month later: 1,988. That is not hiring — no company staffs up 1,935 people in four weeks. What happened is simpler and more damning: the record was frozen. Public reporting puts the company at roughly 60 employees in March 2025, growing to nearly 2,000 within a year — one of the fastest scale-ups in software. The tracked profile held the early-2025 figure while the real company 30x’d underneath it, then snapped to reality in a single update. The same class of record-identity problem surfaced elsewhere in this sample: the company Writer still returns under its pre-rebrand legal name, Qordoba, and Cursor itself is legally Anysphere, Inc., doing business as Cursor — a naming split that makes profile fragmentation easy.
Here’s why this matters more than a data-trivia footnote. A CRM that synced this account in January carried a record saying 53 employees — an SMB, routed to an SMB rep, priced for an SMB deal. By June, the reality was a roughly 2,400-person company that SpaceX had agreed to acquire for $60 billion. Every downstream decision built on the January record — segmentation, territory, tiering, deal sizing — was wrong, not by degrees but by category. And unlike the slow leak, this failure announces itself loudly in the data and is still routinely missed, because nobody re-checks a field that was “verified” five months ago.
Part Three: The ghost signal — old news wearing a new timestamp
The third pattern isn’t about the account changing. It’s about the feed. Of 66 executive-move events detected across this sample, 43 carried a usable event date — and 3 of those 43 (7%) described something that happened more than a year before it resurfaced. The oldest: a Vercel executive departure from July 2021, surfacing in a May 2026 feed — 4.8 years stale. The other two were roughly two years old (Harvey, Character.AI).
All three events were factually accurate. The person really did leave, on roughly the date described. What’s broken is the timestamp — a re-crawl, a retrospective article, or a syndicated republish makes old news look current to any automated pipeline, and to any rep skimming an account feed. Act on it and you open a call with “congrats on the new role” to someone two years into the job. Most vendors selling signal data don’t publish this failure mode. We’d rather you know it exists, because the fix is cheap once you do: check the event date, not the article date, before a signal triggers anything.
Part Four: What the three patterns mean for how you monitor accounts
1. Thresholds catch step changes and miss slow leaks — you need both a threshold and a trend check. Perplexity never tripped a 10% alert while losing 12%. A simple additional rule — “flag any account down three consecutive months, regardless of size” — catches the leak the threshold can’t see.
2. Treat impossible changes as identity events, not data errors. A record that jumps 1,000% in a month isn’t wrong — it’s telling you the entity changed: a rebrand, a merger, a profile consolidation. Route those to a human for re-verification instead of letting a sync silently overwrite the old record or, worse, reject the new one.
3. Date-check every signal before it triggers outreach. The 7% ghost rate in this sample is small, but outreach built on a two-year-old “new hire” signal doesn’t just waste a touch — it tells the prospect you’re running on autopilot. Event date versus publish date is a one-field check.
4. Assume a six-month-old snapshot of a fast-moving account is wrong until proven otherwise. In this sample, that assumption would have been right 23 times out of 25.
Where this goes next
This is one sample over one half-year, in the fastest-moving category in software — AI-native companies are the stress test, not the average, and decay in a portfolio of regional manufacturers will run slower. But the three patterns aren’t category-specific. Slow leaks, identity events, and ghost signals exist in every account base; the AI-native sample just compresses them into a timeframe where you can watch them happen. As more GTM teams wire signals directly into automated outreach, the cost of each pattern goes up: an automation acting on a ghost signal or a pre-rebrand record doesn’t just log a bad data point, it sends the email. The teams that win with signal automation will be the ones that build these three checks in before scaling the sends, not after the first embarrassing sequence goes out.
Methodology and data notes
Data source: Lusha live signals API — monthly headcount tracking, executive hire/departure events (with publish and effective dates), and funding/investment events. Pulled June 22, 2026.
Sample: 25 named AI-native companies across foundation model labs, AI coding tools, AI customer service platforms, and AI-native infrastructure vendors. A sample study, not a census.
Material change definition: a company counts as materially changed if, between January 1 and June 22, 2026, at least one of the following occurred: detected headcount moved 10% or more in any tracked window; at least one executive hire or departure was detected with a 2026 event or publish date; or at least one funding, strategic investment, or asset investment event was detected. 23 of 25 companies met at least one criterion; 11 met all three.
Perplexity figures: monthly headcount signals show declines in January (-5%), February (-2%), March (-3%), and May (-3%), with April essentially flat (+4 employees). Net January-baseline-to-May: 1,633 to 1,437, a 12% contraction. An earlier version of this analysis described the decline as five consecutive months; the month-by-month data shows four declining months and one flat month, corrected here.
A note on absolute headcounts: different workforce data providers report different absolute headcount levels for the same company, because methodologies differ — employee-profile counts, payroll-derived estimates, and full-time-only counts all produce different baselines. The reliable signal in this report is the trend within one consistent tracking method, not the absolute level. Lusha’s month-over-month figures are internally consistent, which is what trend detection requires.
Ghost signal definition: an executive-move event whose publish date and effective date are more than 365 days apart. 3 of 43 dated events (7%) met the threshold.
Privacy and compliance: all figures are company-level, drawn from public signal data. No individual contact details are included or derivable. Full privacy documentation at lusha.com/trust-center.
Corrections and right of reply: this report describes what Lusha’s tracking detected, using the stated methodology — it makes no claims about any company’s performance, health, or intent beyond the detected figures. If you represent a company featured in this report and believe a figure is inaccurate, or want to add context, contact us at lusha.com/contact and we will review, correct where warranted, and note material corrections on this page. One correction has already been made and is documented in the Perplexity figures note above — that’s the standard this report holds itself to.
