Flex Tier V2: Smarter Listing Visibility for a Better Marketplace

April 22, 2026

When a new shift is listed on our platform, who sees it first matters. We want to ensure that workers have first access to shifts they prefer, and that companies secure workers they can count on.

With this in mind, we've noticed that for any given shift, hires who have previously worked for that employer tend to perform better, show up more reliably, and require less onboarding overhead. We refer to these hires as repeat workers, and hypothesize that increasing the rate at which repeat workers fill shifts improves show rates, worker retention, and both worker and employer satisfaction, while keeping fill rates stable.

To act on this, we developed Flex Tier, a listing visibility system that prioritizes workers with prior company experience by giving them early access to new listings. Since its deployment, repeat worker rate has remained above 90% and fill rate has held broadly stable. Most importantly, reliability metrics have trended substantially upward, meaning more workers are completing the shifts they commit to and employers are getting the dependable staff they need.

The Flex Tier System

A simple solution is to let companies select their favorite workers and give them a head start on shift enrollment by making the listing visible to them before anybody else. Among workers, this both encourages good performance and rewards those who are actively engaged with the app. However, it’s not immediately clear how much time to allocate for the head start – too much time, and we risk the listing not having enough exposure, potentially leading to the shift being left unfilled. Too little, and we risk offering the shift to the broader pool prematurely, increasing the likelihood of no-shows and inconsistent performance.

We built on top of these ideas in creating our original listing visibility system: Flex Tier V1. This solution relies on statistical models to identify dependable workers and to predict engagement with each listing among all eligible workers. Repeat workers and those classified as dependable are then placed highly in a tier system, where higher tiers receive priority for claiming newly posted shifts. For each listing, we then use the engagement predictions and hand-tuned heuristics to determine how much enrollment time each tier receives.

In late 2024, we rolled out Flex Tier V2. The tier system was expanded from 4 to 12 tiers using refined experience thresholds. The finer granularity of the new tier system allows us to have more precise control over listing visibility and richer insights into workforce patterns. For any given listing, we now generate personalized engagement predictions for each worker instead of forecasting at the aggregate level. We also improved the time allocation step by replacing hard-coded heuristics with dynamic time windows that scale with per-tier engagement and open to the next tier ahead of schedule when observed engagement lags behind projections.

Worker Tiers

When segmenting the workforce based on experience for any given listing, we want to capture nuance while keeping repeat workers as the benchmark. For any given shift, a worker could have experience in several overlapping senses:

Number of shifts they've worked at the same company before
Number of shifts they've worked the same position before
Number of shifts they’ve worked in the same location before
Number of shifts they’ve worked through WorkWhile
Some combination of all of the above

These dimensions are roughly hierarchical — a worker who has worked the same position has also worked through WorkWhile, and one who has worked at the same company has likely done so in the same position. This hierarchy maps well onto a pyramid, with repeat workers occupying its top tiers. We divide the worker pool into 12 tiers:

Pyramid showing the 12 tiers of Flex Tier V2 showing those with varying degrees of company experience at the top

Workers in tiers 1 through 6 (Favorite through Company II) are repeat workers since they have prior experience at the hiring company. The tier boundaries were established through a combination of hand-tuning and data analysis, calibrated to reflect meaningful differences in expected reliability.

For companies large enough that explicitly selecting Favorite workers isn't practical, Flex Tier creates a virtual roster of Regular I workers that can be used to give them first access in place of Favorites, an exclusive enrollment window with a custom expiry, or a listing visible only to them. This helps enterprise customers maintain high repeat worker rates without the overhead of manually curating a Favorite tier.

Determining the probability of worker follow-through

Giving experienced workers early access is only useful if those workers are going to act on it. So the system needs to estimate, for any given worker-listing pair, how likely that worker is to both claim the shift and show up for it.

At the core of Flex Tier V2 is our schedule rate model, a classifier which predicts whether the worker will sign up for the shift. The model is trained exclusively on workers who have been active on the app recently to avoid diluting the signal, and takes into account 171 features spanning:

Company attributes (e.g., number of shifts paid out by the company)
Worker attributes (e.g., number of shifts worked in the same position)
Shift attributes (e.g., time of day the shift is scheduled)
Some combination of all of the above (e.g., distance from the worker to the shift)

The model outputs the probability that the worker $w$ claims the shift $s$ :

P(w \text{ claims } s)

About 98% of worker-listing pairs do not have the worker claim the shift, simply due to workers only being able to work one shift at a time, and eligible workers vastly outnumbering the number of workers needed for any shift. Despite this substantial dataset imbalance, the model achieves a true positive rate of ~42% while maintaining a true negative rate of ~99.9%. Consistent with our hypothesis, three of the strongest predictors relate to a worker's history at the company: the number of shifts previously worked there, the number previously scheduled, and the number previously waitlisted. The leading predictor is distance to the shift – workers are much less likely to claim shifts further from them. Below is a beeswarm plot of SHAP values for the model’s most impactful features:

Beeswarm plot of the SHAP values of the model's nine most impactful features (ordered by sum of absolute SHAP values) and then the aggregated SHAP values of the 162 other features

The plot confirms that distance can be a powerful deterrent, though its effect varies considerably across workers. Company history is a dependable signal in both directions: most workers with little company history are somewhat less likely to claim, while those with more history can be significantly more likely to.

The model is generally robust to missing data, with the exception of shift-specific features like distance, which have fewer natural proxies. To understand where the model is most vulnerable, we identify the features whose absence most significantly degrades the True Positive Rate (TPR), while adjusting the classification threshold to hold the True Negative Rate (TNR) at 98%:

Plot showing the five features most vulnerable to nullification by their respective drops in TPR (at fixed FPR = 0.02) as well as the median of the TPR drops of the rest of the features

As expected, the three largest TPR drops upon feature nullification correspond to shift-specific features, and the next two largest drops come from features ranked among the strongest predictors by the SHAP analysis.

Next, we consider the probability of a worker showing up if they have already signed up for a given shift. For this we use our reliability model, developed independently of Flex Tier, that estimates the probability that a given worker will attend any shift they've committed to:

P(w \text{ shows up for a claimed shift})

While the reliability model is shift-agnostic, once a worker has committed to a shift, whether they follow through is largely a function of individual reliability rather than shift-specific factors.

Multiplying these two gives us a good estimate of the probability that a given worker will actually be present on shift day:

\begin{aligned} P(w \text{ shows up for } s) &= P(w \text{ claims } s) \cdot P(w \text{ shows up for } s \mid w \text{ claims } s) \\ &\approx P(w \text{ claims } s) \cdot P(w \text{ shows up for a claimed shift}) \end{aligned}

Allotting Time for Each Tier

With per-worker probabilities in hand, the system aggregates up to the tier level to compute the expected turnout for the shift, and uses this to determine how much time the listing should be visible to each tier before cascading down to the next tier. The higher the tier’s expected turnout, the longer its visibility window.

For any tier $T$ , we sum the individual worker probabilities to get the expected number of workers from that tier who will show up for the shift $s$ :

\mathbb{E}[\text{\# workers showing up in tier } T] = \sum_{w \text{ in } T} P(w \text{ shows up for } s)

We now consider each tier's share of the listing’s enrollment time, $R_{T}$ :

\text{time in tier } T = R_{T} \cdot (\text{\# hours notice})

We want $R_{T}$ to scale with the expected turnout so that tiers expected to fill a greater share of the shift are given ample opportunity to claim it before it opens to lower tiers. A natural way to do this is to set $R_{T}$ to the ratio of the expected turnout to the total turnout needed:

R_{T} = \frac{\mathbb{E}[\text{\# workers showing up in tier } T]}{\text{\# workers needed for the shift}}

If a higher tier is expected to fill most of the shift on its own, it gets a longer window. If a tier is thin or its members are unlikely to engage, the window is shorter and the listing moves on quickly. Of course, the system expands visibility rather than replacing it – once a shift opens to a lower tier, higher-tier workers don't lose access.

The one exception to this cascade is that once visibility passes the WorkWhile tiers, the remaining enrollment time is shared between the New and Reserve tiers simultaneously. In 2025, both tiers were roughly six times more likely than workers in WorkWhile I or higher to no-show or reject a shift with short notice. With both groups showing similar reliability, further distinction between them adds little predictive value.

In practice, a fixed schedule cannot adapt to how workers will actually respond to any given listing, so the system adjusts windows dynamically based on observed engagement. If we observe fewer sign-ups than expected by the midpoint of a tier’s window, we open the listing to the next tier ahead of schedule. Up to half of the time allotted for tiers above New can be recovered by triggering these early tier transitions.

The System in Action

To illustrate the system end to end, consider a listing calling for 10 workers with a 48-hour enrollment time:

Flow chart showing the different steps in computing the Flex Tier V2 schedule for a listing with 10 workers needed and a 48-hour notice given

The engagement model factors in both shift-level characteristics – such as time of day and pay rate – and recent worker activity, downweighting workers who have been less engaged in recent weeks. As a result, several top-tier workers receive lower predicted show-up probabilities despite their experience. For example, Worker B has been inactive on the app for the past few weeks which suggests they may be unavailable, and Worker F has historically avoided shifts with the listed start time, so the model assigns a lower claim probability accordingly.

Nevertheless, by summing the expected number of workers per-tier, we predict that we can source 5 of the 10 workers required for the shift from experienced workers, most of whom are repeat workers. Since this represents half of all eligible workers, we set aside the first half of the enrollment time for early visibility by these experienced workers, with the rest going to New workers and those in the Reserve tier.

At the midpoint of the Regular I window, say 3 workers (A, B, C) have signed up against an expected threshold of 3.5 (2 Favorites + 1.5 projected Regular I). As part of a dynamic acceleration system – which continuously compares actual sign-ups against projected targets – falling below this threshold automatically triggers an early transition to Bench tiers.

Results and Observations

Because Flex Tier directly influences the listings workers see, we cannot backtest its effects on historical data. So, we shipped Flex Tier V2 gradually, with the first companies to gain access in late November 2024, and the full rollout by June 2025. We consider three metrics that reflect worker supply and demand:

Fill Rate: the ratio of workers filled to workers needed at the shift’s start
Show Rate: the ratio of workers filled to workers scheduled at the shift’s start
Schedule Rate: the ratio of workers scheduled to workers needed at the shift’s start

Below we examine each of these rates over a 27-month window approximately centered on the initial release, along with the proportion of workers requested across shifts listed using Flex Tier V2, and excluding short-notice shifts listed less than 12 hours before shift start:

Plot showing the fill, schedule, and show rates as lines and the ratio of listings posted using Flex Tier V2 as bars over the 27 months from January 2024 to April 2026. There is a shallow but consistent decline in fill and schedule rates from April to September 2025 followed by a recovery by 2026.

All three rates remain broadly stable over this period, with brief dips in schedule and fill rates after full deployment in April 2025 that normalize after a few months. Mature, high-liquidity markets and high-volume companies saw mixed results, with some showing a small lagged dip, and some showing no disruption at all. For example, Phoenix & Tucson — our fifth-largest market — averaged a 97% schedule rate post-deployment and showed no discernible decline around the time of rollout, as shown below:

The vertical dashed line shows that the only notable dip in schedule rate over this timeframe occurred when Flex Tier V2 adoption was still minimal, making it an unlikely cause. This inconsistency suggests demand-side or regional factors may have coincided with the rollout, and further analysis would be needed to isolate the system's contribution to schedule rate from other confounding factors.

We also consider three metrics that reflect worker reliability and worker experience levels:

Reliability Rate: the ratio of workers who successfully completed their shifts to workers filled
Experienced Worker Rate: the ratio of workers in at least WorkWhile I to workers filled
Repeat Worker Rate: the ratio of workers in at least Company I to workers filled

We examine each of these rates over the same timespan and exclude short-notice shifts as before:

Plot showing the repeat worker, experienced worker, and reliability rates as lines over the 27 months from January 2024 to April 2026. Repeat worker rate and experienced worker rate see slight increases and reliability rate sees the most pronounced increase.

All three rates trend upward following the rollout, consistent with our hypothesis that a higher repeat worker rate is correlated with more reliable workers. Repeat worker rate remains above 90% after deployment. The standout result is reliability rate, which rose from an average of 91% pre-deployment to 95% post-deployment, a meaningful gain that speaks directly to the quality of placements on the platform. Taken together with the restabilization of fill rate, these trends make a strong case that prioritizing repeat workers is good for the marketplace, and that Flex Tier is an effective way to do it.

Directions for Further Study

Encouraged by these results, we are actively exploring several directions to extend and refine the system:

Shift-level flexibility. Different industries and roles have different priorities: a company with an event listing calling for 100 workers may favor speed over selectivity, while one requiring 5 specialized workers may prefer to wait for proven workers even at the risk of a lower fill rate. Scaling tier enrollment windows and tailoring acceleration behavior to shift type and company preference is a natural next step.

Granular tier acceleration. The current tier acceleration trigger is at the midpoint of a tier's window based on a single threshold. A more granular approach might incorporate dynamic checkpoints that adapt to shift type, manual override controls for use by operations teams, or models that predict each worker's response latency – the time between seeing a listing and claiming it – to make acceleration decisions more precise.

Worker development pathways. The tier system is designed to prioritize experienced workers, but this comes with a tradeoff: new workers see fewer listings early on, particularly for desirable shifts. This limits their ability to build experience, and could discourage them before they have a chance to establish themselves on the platform. Organic growth in listing volume partially offsets this but does not directly address it. We could intervene by occasionally surfacing listings to new workers ahead of schedule, or by developing signals that identify promising newcomers early to place them in a fast-track tier between the New and WorkWhile tiers. Either approach would help retain new workers and build a steady pipeline of experienced workers.

System interplay. Flex Tier interacts with other shift-level systems. For instance, our bonusing system, which dynamically incentivizes workers to claim shifts, directly overlaps with tier transitions. Higher-tier workers may respond differently to bonuses than lower-tier workers, motivating research into how bonuses should be tuned differently across tiers. Because bonus changes affect worker engagement in real time, the acceleration logic would benefit from accounting for current bonus levels when deciding whether to trigger an early tier transition. Making each system more aware of the other's behavior could improve overall system coherence.