Recognition stopped starting, not spreading

Recognition happens between two people. Someone notices what a colleague did and tells them so, by name. We hold the record of that moment, and we have six years of those records across 64 organizations, which makes it possible to watch how often people actually do it.

They do it much less than they used to. The share of employees who recognized a colleague in a given month fell from 0.632 in 2020 to 0.294 in 2026.

Two explanations come up first. People have grown less generous, or workplace relationships have thinned out to the point where colleagues no longer notice each other's work. Both sound reasonable. The data supports neither.

Recognition travels. One person recognizes a colleague, and that colleague recognizes someone else. We measured the two halves of that chain separately and they moved very differently. Spreading weakened a little. Starting collapsed. The people who do still recognize colleagues do it as often, and to as many people, as they did in 2020.

Reading a six-year panel

Two things have to be controlled before a year-over-year comparison means anything. Both are general problems in longitudinal workplace data rather than quirks of this dataset, and each one on its own is large enough to reverse the answer.

The first is lifecycle. Participation follows a predictable arc: people are busiest in their first few weeks, settle by around month six, then hold steady. Any year that happens to contain a lot of recently onboarded people reads high. Every series here is tenure-matched, comparing people at the same point in their own history rather than at the same point on the calendar.

The second is composition. The set of organizations contributing data shifts from year to year, and organizations differ enormously in how much recognition they generate, so a pooled average across years partly measures who was present. Every series here holds the panel fixed, and the central finding is repeated on a balanced panel of organizations observed at both ends of the window.

Fewer people give, and the givers have not changed

Split recognition into two questions: how many people give at all, and how much each giver gives. Only the first one moved.

Every year, givers send between 3.2 and 4.1 recognitions a month to between 2.3 and 2.7 different colleagues. Nobody got stingier and nobody's circle got smaller. That rules out the simplest version of the weakening-relationships story, which would have turned up in both of those numbers and turns up in neither.

Starting collapsed faster than spreading

To pull the two halves apart, take employees who have given nothing for four weeks and watch what happens next. A few of them receive recognition that week. Most do not. The gap between those two groups is how much recognition propagates. Returns to whoever did the recognizing are thrown out, so a quick thank-you does not count as spread.

Unprompted giving fell 60%. Triggered giving fell 38%. The ratio between them climbed from 2.29 to 3.59, which looks like spreading got stronger. It did not. The ratio rose because its denominator fell faster.

The same people, before and after

A decline like this is often just composition, meaning different people in the later period rather than changed behavior. Restricting to the 766 individuals we see as quiet in both eras takes that off the table.

The gap holds. Inside the same individuals, starting still falls about 1.6 times faster than spreading. The drops are smaller than in the full population because these 766 are the committed core, people who stayed engaged for four years or more.

The receipt rate is worth a look on its own. It fell 29%. Fewer people start, so fewer people receive, so fewer get triggered, so fewer start. The multiplier still works. It has less and less to work on.

The same companies, before and after

The same worry applies one level up, at the organization rather than the person. The panel is not the same set of organizations in 2026 as in 2021, so a decline could be nothing more than which ones were present. The strictest test available keeps only organizations observed at both ends of the window. That leaves 10 organizations and 5,205 employees, measured in 2021 through mid-2022 and again in 2025 through mid-2026.

Across 180,402 quiet person-weeks the numbers move the same way and in the same order. Ten organizations is a small panel, and the gap between starting and spreading is narrower here, about 1.3 times rather than 1.6. It never closes and it never reverses.

New joiners never pick up the habit

If people learn recognition by watching colleagues do it, then fewer chances to watch should show up in whether new arrivals ever start. Hold tenure fixed at months 6 to 17, so every cohort is measured at the same point in its own life, and each new cohort starts at roughly half the old rate.

Of the people who joined in 2020, 62% picked up the habit. Everyone who joined from 2022 on sits between 30% and 33%. The ones who do pick it up send as much as ever, between 3.17 and 4.08 a month. They behave like they always did. There are half as many of them.

What this means

Most recognition programs aim at the wrong half of the chain. Campaigns and reminders that push existing givers to give more are working on the part that never broke.

The scarce thing is the occasion. Someone has to be close enough to a colleague's work to see what they did, and prompted enough to say it out loud. That part happens between two people and no tool does it for them. Software can ask the question and carry the message, but if the moments are not there, there is nothing to carry. Making more of them is mostly a question of how a team runs its week.

Limitations

• We do not know the cause. Two unexplained breaks fall inside this window and show up across almost every company at once, one in participation in the first half of 2022 and one in recognition in the second quarter of 2023. We cannot yet separate changes in how organizations work from changes in the environment where these interactions get recorded. Either would produce collapsing starts with spread left intact.
• We only see recognition that gets recorded. If colleagues still recognize each other but do it somewhere we cannot see, part of this is a change of venue rather than a change of behavior, and we cannot rule that out directly. It fits the pattern badly, though. A move to another channel would pull observed starting and observed spreading down together. It would not open a gap between them, leave volume per giver flat for six years, or make each new cohort start at half the old rate at matched tenure.
• The multiplier is a correlation. Receiving recognition is not random. Recipients are better connected than non-recipients to begin with, and that gap may have widened, so we cannot put a causal size on the spread effect.
• Cohort and period cannot be separated. Every cohort is measured at matched tenure but necessarily in a different calendar year, which is the standard age-period-cohort problem. Long-tenured employees also decline as they age, from 0.638 to 0.404.
• Early levels rest on few organizations. The pre-2021 panel covers five or six organizations, so the early levels are shakier than the post-2021 trend.
• Holding the panel fixed has its own cost. It is what removes the composition artifact, but it also means the early years are seen through the subset of organizations present throughout. The balanced panel is the strongest check available on that, and a stronger one is not constructible from this data.