1. The data at a glance

The corpus was built in stages, and it helps to see the whole funnel at once: from every facility we found, to the reviews eligible for analysis, to the sample, to the slice that has been read and labeled so far.

About one in ten reviews (9.7%) produced no valid labeled statement - typically very short or non-evaluative reviews ("Great hospital", an address correction, an emoji). Everything that follows is based on the 425,667 statements that passed all validity checks (98.8% of what the model returned).

2. The taxonomy: 29 topics in 7 arenas

Every statement is filed into a taxonomy that was not designed in advance - it was discovered from the reviews themselves (batches of reviews were read by a model, recurring themes consolidated, and the result manually refined), then frozen before labeling began. This is the complete map of what American hospital reviews talk about:

3. How one review becomes data

The method is easiest to understand on a single real example. The model reads a review, extracts every evaluative statement it contains verbatim, and files each statement into one of the 29 topics with its own sentiment. Below is one actual review from the corpus - quoted verbatim, typos included - and everything the pipeline extracted from it:

Behind each of those topic chips sits the discovery pipeline. The table traces the staffing statement's topic from raw model output to the final taxonomy - every value below is a real artifact of the run:

Every statement in this notebook went through exactly this path, and each one must pass four mechanical validity gates (the phrase quotes the review verbatim, the topic exists in the frozen taxonomy, the topic id matches, and the item resolves to a review of its batch) before it counts - details in section 11.

4. Where the hospitals are

The optional dot overlay shows one facility with labeled reviews per point - the full sample is covered, coast to coast, Alaska and Hawaii included.

The sample tracks where Americans live: dense along the coasts, the Great Lakes, Texas and Florida, sparse across the Mountain West and Plains. Because reviews were sampled per facility - with a floor of 20 per facility and a cap of 200 - small rural hospitals are genuinely represented rather than drowned out by large urban systems, which is what makes state- and region-level comparisons meaningful in later analyses.

5. How people rate, how much they write, and what they say

Star ratings, review length, and the sentiment of the labeled statements are the three most basic descriptions of the corpus - and each one already tells a story.

The J-shape is extreme. 31% of reviews are one star, 58% are five, and the three middle ratings together account for barely one review in nine. People write about hospitals when something was very good or very bad - rarely in between. The average (3.54) is therefore a number almost nobody actually gave.

Length follows the mood. The typical review is short - a median of 37 words - but the distribution has a long tail of very long reviews, and those are overwhelmingly accounts of things going wrong: one-star reviews average more than twice the length of five-star ones, and two-star reviews - the rarest kind - are the longest of all. The dots inside the density curve make the shape concrete: a dense cluster of brief praise, and a thinning trail of long, detailed complaints.

The statements are nearly balanced. Unlike most consumer categories - where public reviews skew heavily positive - hospital statements split 54% positive to 45% negative. The public hospital record is as much a complaint channel as a thank-you channel, which is exactly what makes it analytically useful.

6. The record over time

Two questions: how much do people write, and how has the tone changed? (Reviews are shown by the quarter they were published; 2026 covers January–June.)

Volume has tripled. The sample contains about 15,000 reviews from 2019 and over 46,000 from 2025 - mirroring the general growth of Google reviewing. The pandemic left a sharp scar: in the spring of 2020, quarterly volume collapsed to less than a third of its normal level - and then recovered within two quarters. People stopped reviewing hospitals during the first lockdown, but only briefly.

The tone has clearly improved since 2022. Negative share hovered around half of all statements from 2019 through 2022 (peaking at 52.4% in 2022), then fell steadily: 46.8% in 2023, 45.1% in 2024, 40.4% in 2025, and 37.3% in the first half of 2026. The star ratings tell the same story independently - the quarterly mean climbed from about 3.2 in 2019 to about 3.8 in 2025.

An improvement this large deserves suspicion rather than applause. Two boring explanations could produce it without any hospital getting better: new, calmer facility types entering the review record, or the mix of complaints shifting. The next section takes the trend apart.

7. Why is the tone improving?

Before decomposing the trend, look at where the improvement lives. The table below tracks the negative share of each of the seven experience arenas year by year - green means friendlier than the corpus average, red means harsher.

The table splits the hospital record into two different worlds:

• The human arenas are healing. Staff, Communication & Nursing - half of the whole conversation - improved from 38% negative in 2019 to 23% in 2026, and Access & Timeliness fell from its 2022 peak of 64% to 45%. These two arenas are the improving tone.
• The administrative arenas are frozen. Financial & Administrative has been above 90% negative in every single year - no pandemic effect, no recovery, no trend. Care Process & Coordination has barely moved from its high-70s plateau. Whatever healed the human side of hospital experience has not touched billing desks or paperwork.

So "the tone is improving" is really "the people topics are improving." With that established, two decompositions pin down the rest: is it the same facilities getting better, or new ones entering calmer - and which topics carry the change?

The two decompositions give the trend a much more solid footing:

• The improvement is real within facilities. Even the veterans - the 6,375 facilities reviewed continuously since 2019 - improved from 53.9% negative in 2022 to 39.3% in the first half of 2026. New facilities do enter the record calmer (the 2022+ cohort runs roughly 8 points friendlier), so part of the aggregate trend is composition - but composition is not the main story.
• The recovery has a name: the front of the house. Of the 12-point drop in negativity since 2022, shorter wait times alone account for 3.6 points and friendlier front-desk encounters another 2.3 - together roughly half of the entire improvement. Scheduling, bedside manner, nursing responsiveness and staffing add the rest. And notably, no topic got worse. This looks like the post-pandemic normalization of exactly the pressure points that COVID-era hospitals were blamed for: queues, chaos at intake, and overwhelmed staff.

The change is audible in the reviews themselves - the same topic, three years apart:

8. Facility types: where the pain concentrates

Google's own facility categories let us split the record by care setting - general hospitals, emergency rooms, urgent care, psychiatric facilities, and so on. (A few spelling variants of the same category were merged; types with fewer than 1,000 directional statements are omitted from the negativity ranking.)

The spread between facility types is enormous - and it is the first real insight of the labeled corpus:

• Psychiatric hospitals (72% negative) top the list by a wide margin. Public reviews of inpatient psychiatric care are dominated by complaint - a finding consistent with what patient advocates have long argued about that sector, and a reason to treat it as its own analysis later.
• Mental-health clinics and university hospitals (both 66%) share second place - for university hospitals a surprise: prestige does not protect against negative public feedback; scale, complexity and ER load may explain part of it.
• Emergency care (~53%) sits well above the average, matching the dominance of waiting-time complaints seen in the next section.
• Urgent care centers (28%), surgical centers (26%) and specialized hospitals (24%) are the calm end. Planned, focused settings collect far friendlier words - urgent care in particular seems to deliver the fast, simple visits that ERs are blamed for not providing.

In short: the more acute, involuntary, and system-constrained the setting, the harsher its public record.

What that reads like in practice:

9. The geography of experience

The same statements, cut by state (extracted from each facility's address). The differences are far larger than one might expect.

The spread between states is enormous - from about 36% negative in Florida, Texas and Nevada to 62% in New York, with Iowa, Pennsylvania, Massachusetts and North Carolina all above 56%. Broadly, the Northeast reads harsh and the Sun Belt reads calm.

A word of caution before reading this as "hospitals are worse in New York": state comparisons mix facility composition (a state with many urgent-care centers will look calmer than one dominated by big-city ERs), reviewer expectations, and genuine experience differences. Separating those requires the per-facility-type controls planned for the rating-driver analysis - this map is the motivation for it, not its conclusion.

The big-city view

The same lens, zoomed to the ten largest hospital-review markets in the sample (city proper, extracted from facility addresses; New York's boroughs are counted as one city).

Among the ten biggest markets the spread is just as dramatic as between states: Chicago (63%) is the harshest big-city record, followed by Los Angeles (56%), Indianapolis (54%) and Atlanta (52%) - while Las Vegas (36%), Austin, Miami and Houston (38–39%) anchor the calm end. The Texas metros are uniformly friendlier than the national average, echoing the state map.

Two caveats specific to cities: the comparison is city proper, not metro area (suburban facilities count toward their own towns), and the sampling design caps very large facilities at 200 reviews, which mutes the mega-hospitals that dominate the densest cities - one reason New York City does not appear among the top-10 markets despite New York state's prominence in the state map.

10. What people talk about - and what drives the ratings

Every statement is tied to one of 29 topics, each belonging to one of seven broader arenas of hospital experience. Three questions, three views: where does the conversation live (the arenas), how angry is each topic when it comes up (tone), and which topics travel with the worst ratings (rating gap).

Read the four charts together and the central finding of this first pass appears:

Where the conversation lives: in the human encounter. Half of all statements fall into the Staff, Communication & Nursing arena, and clinical staff bedside manner alone appears in 42% of labeled reviews - by far the most common theme, and mostly positively. Wait times (26%) and front-desk conduct (25%) follow. The everyday hospital review is about people and time, not about medicine.

The quiet topics are the angriest. The scatter makes the asymmetry visible in one view: the most-mentioned topics sit in the calm lower half, while patient safety, privacy, billing and discrimination - each mentioned in only 1–5% of reviews - are more than 90% negative when they come up at all. Nobody writes a five-star review about billing.

The same topics anchor the worst ratings. Reviews mentioning safety errors, discrimination, privacy, or billing average about 1.5 stars - two full stars below the corpus mean. And the positive pole is just as clear: rehabilitation (+1.0★), surgical care (+0.9★) and maternity and pediatric care (+0.6★) mark the experiences people write thank-you notes about - planned care with visible outcomes.

The practical implication: a hospital that wants a better public record by fixing what is most mentioned would work on warmth and waiting. One that wants to eliminate its worst reviews would look instead at billing surprises, safety incidents, and how complaints of unequal treatment are handled - the quieter topics that anchor the deepest dissatisfaction.

11. How much should you trust the labels?

Every finding above rests on model-generated labels, so the honest last section is about their quality. Some of it can be verified mechanically; the rest needs human eyes.

The mechanical checks are strong: 98.8% of everything the model returned passes all gates at once, and the strictest gate - the extracted phrase must quote the review verbatim - passes at over 99%. This matters because it means the labels are grounded in actual review text, not paraphrases the model may have invented.

What machines cannot verify is whether the topic and sentiment are right. For that, a 300-statement human-validation sample accompanies this notebook (qa/label_validation_sample.csv), deliberately oversampling the four rare, serious topics - safety, discrimination, privacy, billing - on which the strongest claims rest. Until that annotation is done, the per-topic findings above should be read as model-labeled, and the serious-topic findings in particular as provisional. This is the single remaining step between this analysis and something that can be defended externally.

What comes next

This was the descriptive pass - the shape of the record. The immediate next step is small: annotate the 300-statement validation sample from section 11, which converts every claim above from model-labeled to validated. Beyond that, three deeper analyses are planned on this foundation:

1. Reviewer composition. About 7% of reviewer profiles have reviewed more than one hospital, and they write measurably more negative reviews. Do hospital rankings change if we account for who is doing the reviewing?
2. The Google–HCAHPS comparison. Matching these facilities to the official CMS patient-experience survey to ask what public reviews capture that the survey misses - and where the two agree.
3. Failure journeys. Which problems co-occur in the same review more often than chance - does a billing complaint travel with an ER wait, does a safety complaint follow a staffing one?

The full sample is now labeled (two passes: 100,000 + 117,343 reviews), and the complete 2.09-million-review eligible corpus stays reachable behind it with the same frozen taxonomy, should a census-level analysis ever be needed.