The Rival Review-Shift Read: Durable Decline, Event Spike, Solicited, or Noise

A rival's employer-review rating has moved, and leadership wants to know whether that rival's people are now gettable. This guide is for talent-intelligence analysts, strategy teams, and employer-brand leads who have to answer before anyone re-points sourcing or brand spend. It gives you a sample-size gate, six scored dimensions, and four verdicts, each tied to a named action and a re-read date, so two analysts reading the same profile reach the same call.

What do the four verdicts mean?

A rival's rating shift is one of four things: a durable decline in how current staff rate the place, a spike tied to a dated event such as a layoff, a solicited wave of reviews, or noise. Only a durable decline, or an event spike with visible exits, justifies moving sourcing effort toward that rival.

Terms used throughout:

• Employer-review rating: the average star score on a public employer-review profile, built from individual anonymous reviews.
• Sub-rating: a category score on the same profile, such as senior management, culture, work-life balance, or CEO approval.
• Baseline: the rival's own monthly review count and mean rating over the 24 months before the period you are reading.
• Window: the period whose move you are trying to explain.
• Solicited review: a review written because the employer asked for it, usually as part of a coordinated push.
• Corroboration: evidence outside the review site that people are actually moving, meaning public profile exits and backfill job postings.

The four verdicts:

• Durable decline: current employees rate the employer lower across more than one window, and people are leaving voluntarily. This is the only verdict that supports re-pointing standing sourcing effort.
• Event spike: a dated event, usually a layoff, produced a burst of reviews carried mostly by people who left. The affected cohort is reachable for a while; the people who stayed may not be.
• Solicited: the employer drove a wave of reviews, usually positive. The rating is telling you about an internal campaign, not about staff sentiment.
• Noise: the move sits inside what the sample can produce by chance, or it is a decline with no corroboration. No action beyond a re-read.

Most current advice stops at "set an alert on the star rating". An alert tells you something moved. It does not tell you which of these four things happened, and the four call for opposite actions.

How many reviews does a rating move need before it counts?

Nobody has published a minimum review count for employer ratings, so you have to derive one from the rival's own data. Under reasonable assumptions, a 0.2-star move needs hundreds of reviews per window before it means anything, which most mid-size rivals never accumulate in a quarter.

The thresholds that circulate come from practitioners and give no method. One says profiles with fewer than about 10 reviews are unreliable. Another says fewer than 25 to 30 reviews makes scores unstable. The WSJ's 2019 analysis only looked at companies with at least 100 reviews. None of these answer the actual question, which is how many reviews a given shift needs.

The table below is derived, not published. It rests on an assumed standard deviation (SD) for individual ratings, because no SD has been published either. The single-window figure is what you need to pin one window's mean within the shift. The two-window figure is what you need to say the window differs from the baseline, at 80% power.

Formulas: single-window n = (1.96 x SD / shift)²; two-window n = 15.68 x SD² / shift² per window.

Check it locally. Your export contains every individual star rating, so compute the SD for this rival directly and plug it into the formula rather than borrowing either column. If you cannot compute it, use the SD 1.3 column; voluntary reviewers are polarised toward one and five stars, which pushes the spread up.

One shortcut is legitimate. Leadership sub-ratings move much further than the overall average after a shock: Glassdoor's 2025 study found an 11.3 percentage point drop in CEO approval after layoffs, against a 0.13-star drop in the overall rating. Treat the sub-rating as a leading indicator worth watching. It does not let you skip the gate, because the sample needed for a percentage-approval figure is not published either.

The six dimensions and how to score each

Score every read on the same six dimensions, and record each one as a reading that points toward a verdict, or as neutral. The table is the scoring key; two analysts using it on the same export should record the same readings.

What each dimension proves, and what it looks like when it lies:

• Volume against its own baseline. Proves something happened in the window. It lies when a hiring surge, an award-season push, or a layoff all produce the same spike shape. Volume never decides a verdict alone.
• Sub-rating concentration. Proves where the move sits. A drop concentrated in senior management and CEO approval is the classic early shape after a shock. It lies when a single solicitation email lifts CEO approval, which is exactly where drives aim.
• Text theme clustering. Proves what reviewers are reacting to. Themes that recur across consecutive months point to structure; themes that appear in one burst point to an event. It lies when the burst is generic: short, interchangeable text is a commonly cited solicitation tell, but it has not been validated publicly, so record it and do not count it.
• Reviewer tenure and role mix. Proves who is talking. This is the dimension that most often separates an event spike from a durable decline. It lies when former employees review long after leaving, which spreads an event's tail across later windows.
• Solicitation fingerprints. Proves a campaign. It lies when a genuine hiring wave produces happy new joiners.
• Corroboration from departures and postings. Proves people are moving, which is what leadership actually asked. It lies when exits are layoffs rather than choices.

The evidence gets stronger as you move from the surface of the profile toward what people actually do.

How do layoffs move a rival's rating, and for how long?

A layoff knocks the overall rating down fast and keeps it down for at least six months. Glassdoor's 2024 report found the overall rating fell four percent on average, from 3.66 to 3.49, in the 30 days after a layoff, and stayed that low 180 days later, across more than 100 companies with major layoffs.

The 2025 study, which combined Glassdoor reviews from 2021 to 2025 with layoffs.fyi entries and WARN Act filings, adds the detail you need for scoring:

• Layoffs lowered ratings by 0.13 stars overall and by 0.16 stars among employees who stayed.
• The median employer with layoffs dropped from the 56th percentile to the 41st.
• CEO approval fell 11.3 percentage points.
• Employees looking for jobs on the platform rose 40% after layoffs.
• Current-employee ratings had not fully recovered after 24 months.

Sources do not fully agree on the shape of recovery. One account of the 2024 report says ratings stabilise for most categories over time, while culture and work-life balance were still falling at five months. Read that as a reason to re-check sub-ratings at the 180-day mark rather than the average alone.

The size of the drop depends on where the rival started.

The best-rated firms fell 11 times as far as the poorly rated ones. The likely mechanism is expectation: a strong reputation sets up a bigger gap between what staff believed and what happened. The practical consequence is counter-intuitive. A large drop at a highly rated rival is weaker evidence of a durable decline than the same drop at a mediocre one, because big post-layoff drops are normal for highly rated firms.

The 0.16-star figure for stayers matters too. Layoffs do not only produce angry ex-employee reviews; they can start a real decline among the people who remain. That is why an Event spike verdict always carries a fixed re-read, and why it can convert to Durable decline on that re-read.

How to detect solicited employer reviews

Solicitation shows up in the rating distribution and the calendar before it shows in the mean. The documented fingerprints come from journalism, chiefly the WSJ's 2019 analysis, not from peer-reviewed research, so treat them as tells to count rather than proof.

The WSJ found more than 400 of roughly 8,500 companies with unusual single-month review spikes. The tells it documented, which I count toward the Solicited reading:

1. A single-month volume spike. For scoring, the window's monthly count exceeds every month in the 24-month baseline and matches no layoff or event.
2. A five-star share jump. In spike months, five-star reviews made up 45% of reviews, against 25% in the six months either side.
3. October timing. More than a quarter of spikes fell in October, about triple the share expected if spread evenly, because that is when Best Places to Work awards are calculated.
4. Manager titles. Reviewers in spike months disproportionately listed manager titles.

Two or more tells, or a public alert, scores the dimension toward Solicited. Commonly cited tells that are not established publicly as valid signals: short generic text, and first-time reviewers clustered within a few days. Record them in notes; do not count them.

The platform's own rules give you a fifth, decisive input. Glassdoor allows one review per person per company per year and bans asking friends, family, or anyone else to write reviews, along with incentives. When it has evidence an employer artificially boosted its rating, it shows an "Inflated Reviews" alert on the profile and review pages. Glassdoor Alerts cover only two cases: legal action against reviewers and inflated reviews. Its awards methodology also checks the consistency of review volume across the eligibility period, which is the same test you are running.

Scale matters for calibration. In one case reported second-hand (the WSJ original was not retrieved), Guaranteed Rate's CEO approval went from 43% to 83% after a solicitation email. Platform-wide, five-star reviews rose from 17% of all reviews in 2013 to 28% by the end of 2018. That drift is why you compare a rival with its own surrounding months, never with an old platform norm.

The confound to rule out every time is a hiring surge. Glassdoor itself notes that a positive uptick can correspond with one. If a matching wave of job postings exists and the new reviews come from recent joiners, score the volume dimension neutral and do not call Solicited.

That polarisation has a second consequence. Without any drive, a profile's extremes are inflated at both ends, so a rising one-star share on its own is a noisy decline signal. Pair it with reviewer mix before you score it.

Turning six readings into one verdict

Apply the readings in a fixed order, and stop at the first rule that fires. Order matters because the cheap, decisive tests (sample size, alerts, event dates) should rule out the common explanations before you spend hours on text and departures.

Decision order:

1. Gate. Window n below the single-window figure for the observed shift: Noise. Stop.
2. Solicited. An Inflated Reviews alert, or two or more counted tells with no matching hiring surge: Solicited.
3. Event spike. A volume spike within 30 days of a layoff or WARN filing, the read is within 180 days of that event, and former employees carry the drop: Event spike.
4. Durable decline. Gate passed at the two-window figure, current employees carry the drop across at least two consecutive windows, leadership or culture sub-ratings lead, themes recur, and corroboration shows voluntary exits or backfill postings: Durable decline.
5. Anything else: Noise. If rule 4 failed only on corroboration, write "uncorroborated decline" in the record.

The mandatory corroboration rule is the opinionated part of this standard. Leadership asked whether people are gettable, and a rating is a mood; exits are behaviour. The best evidence linking the two is Sull, Sull and Zweig's study of 34M profiles and 1.4M reviews, which found a toxic culture was by far the strongest predictor of industry-adjusted attrition, around 10.4 times as predictive as pay in most accounts (Bloomberg reported 12.4). Their attrition measure included layoffs, though, and how often a sustained decline is followed by visible profile departures within two quarters is not established publicly. Until it is, you check exits yourself.

Corroboration is the slow step, because it means finding people who left a named employer in a named period and checking where they went. Refolk handles that as a plain-English search across the public profile graph, so the exit count arrives with dates and destinations rather than a manual trawl.

Write the read down in a fixed shape so a second analyst can check it line by line.

RIVAL:                      READ DATE:            SNAPSHOT FILE:
OBSERVED SHIFT: ___ stars over window ___ to ___
BASELINE: 24-month mean ___ ; monthly review rate ___
GATE: window n ___ ; measured SD ___ ; single-window n ___ ; two-window n ___
GATE RESULT: fail / provisional / pass
ALERTS ON PROFILE: none / inflated reviews / legal action
EVENTS: layoffs.fyi ___ ; WARN filings ___ ; none

D1 VOLUME VS BASELINE: near baseline / spike on event / spike no event / hiring surge
D2 SUB-RATINGS: leadership-led / broad / leadership up in one month / flat
D3 TEXT THEMES: recurring across months / one-burst event / one-burst generic / mixed
D4 REVIEWER MIX: current carry / former carry / manager-title heavy
D5 SOLICITATION TELLS COUNTED (0 to 4): ___ ; uncounted notes: ___
D6 CORROBORATION: voluntary exits + backfills / exits at WARN dates / none

RULE THAT FIRED (1 to 5):
VERDICT: Durable decline / Event spike / Solicited / Noise
ACTION:
RE-READ DATE:
SECOND READER: ______  AGREES: Y / N  IF N, WHICH DIMENSION:

Step-by-step: running the read

The procedure takes roughly one working day per rival, most of it in text clustering and corroboration. Sources disagree on where to start: WSJ-style analysis goes volume first, while Glassdoor's own guidance on reading reviews starts with filtering and the text. I start with volume and dates because they are cheap and frequently end the read at rule 1 or 2, but the text and the exits decide anything that gets further.

Notes on the steps that go wrong most often:

• Step 1. The snapshot is not optional. See the retroactive-removal failure mode below.
• Step 5. Split ratings by current and former employee before you look at themes. Themes from former employees in the 30 days after a layoff will dominate any cluster and make the whole profile look structural.
• Step 6. Search exits by function, not just by company. If the themes say engineering leadership and the exits are all in sales, the corroboration is weak for the function leadership cares about. In Refolk, that is one query per function and period, which keeps the step inside its time budget.

How this read goes wrong

Most bad calls come from one of seven traps, and each has a specific check. Run all seven against the record before the verdict leaves your desk.

Volume spike read as solicitation

A hiring surge produces the same single-month spike, mostly from current employees, often positive. Check: are the reviewers recent joiners, and do job postings show a matching surge? If yes, volume is neutral and Solicited is off the table.

Using 0.6 as an adjustment

Analysts subtract 0.6 stars from a rival's rating to "remove" a drive. This confuses an MTurk experiment with Glassdoor's observational result of 0.035. Check: before using any adjustment figure, find out which population it came from. For drives, use the five-star share.

Retroactive removal

Glassdoor can remove reviews from the period in which a violation occurred, so the baseline itself changes after the fact. A rival's rating can then appear to fall when what actually happened is that inflated reviews were deleted. Check: snapshot the history on the day of the read, and if a later export disagrees with the snapshot, compare the two before scoring a new decline.

Layoff spike called a durable decline

Former employees reviewing right after a cut look exactly like a structural slide on a chart. Check: separate current from former employees. The decline is durable only if current-employee ratings stay depressed, and you should expect them to sit lower for at least 180 days after any layoff, so re-read at that point rather than calling it early.

Small-sample wobble

A 0.2-star move on a profile with a few dozen reviews in the window is noise under any reasonable SD. Check: count n before quoting the delta, every time. Per the gate table, a 0.2-star shift needs hundreds of reviews per window.

Departures confounded with layoffs

The best published link between culture and attrition measures job changes for all causes, layoffs included, and so does a naive exit count. Check: separate exits dated around WARN filing dates from exits outside them. Only the second group corroborates a durable decline.

October seasonality

Award-season drives inflate volume every year, and a rival that runs one annually will spike every October. Check: compare October with the same month in prior years. An October spike that matches previous Octobers is the rival's normal pattern, not news.

Leadership asked whether people are leaving. A rating is a mood; only exits answer the question.

Who should run the read, and how to keep it current

Whoever owns attrition questions should run the read, and in many organisations that is not the person watching the rating. The pool of dedicated talent-intelligence analysts is small, so this standard is written to be run by employer-brand staff too, with the corroboration step as the guard against reading ratings as behaviour.

These are exact-title matches only, so they undercount people doing the work under other titles. Even so, the ratios are telling: the US has 3.4 times as many Talent Intelligence profiles as the UK, where AMS is the top employer with 2, and US Employer Brand profiles outnumber US Talent Intelligence profiles 2.0 to 1. Employer-brand teams are paid to see ratings, not attrition, so they will be tempted to stop at the star average. The decision order prevents that: Durable decline cannot fire without exits.

Keeping the read current:

• Hold every re-read date. Durable decline at 6 months, Event spike at 180 days after the event, Solicited after the next October, Noise when window n clears the gate. Glassdoor's 2025 data supports a 6-to-12-month horizon for post-layoff effects.
• Re-derive the gate each time. The SD changes as the profile grows; recompute it from the new export rather than reusing the old threshold.
• Diff against the last snapshot. Removed reviews show up as differences between snapshots, not as anything the profile announces.
• Re-check the platform rules. Review limits, alert types, and award methodology can change. Re-read the authentic-reviews policy, the Alerts page, and the current awards methodology once a year, and update the tells if they change.