Written one- and two-star Google reviews of at least five words, up to 60 per business, from high-review-count towing companies in eight US metros. A language model, with no human-coded check, classified the main reason in 5,916 of them: 86.9 percent how the business is run, 11.6 percent the tow itself, 1.5 percent unclear.
The 5,916 reviews are written one- and two-star Google reviews of at least five words, drawn from the up-to-60 lowest-rated reviews retrieved per business, from 285 high-review-count towing companies in Phoenix, Dallas, Atlanta, Charlotte, Columbus, Denver, Tampa and Kansas City. In each metro, up to 200 listings were screened and, after filters including a cap of two locations per brand, the 40 businesses with the most Google reviews were sampled. Exclusions then left 311 businesses, of which 285 have reviews in the corpus. Review dates run from 2010-03-09 to 2026-09-30. A separate multi-label model pass assigned at least one business-process label to 90.8 percent of the 5,916 reviews.
This study is written for towing company owners. Every figure below states its denominator. None of them is a share of customers.
The model-classified main reason
The main-reason pass was run by a language model, Claude Haiku 4.5, which chose between how the business is run, the towing work itself, and unclear. It classified all 5,916 reviews.
| Model-classified main reason | Share of the 5,916 reviews labelled by the main-reason pass |
|---|---|
| How the business is run | 86.9% |
| The tow itself | 11.6% |
| Unclear | 1.5% |
A second model, Claude Sonnet 4.5, classified 200 of the same reviews. On the binary question of how the business is run against every other main reason, the two models agreed on 96.5 percent of the 200, with a Cohen’s kappa of 0.801 (95 percent confidence interval 0.64 to 0.92). It is not validation against human coding, and no sample was checked against human coding.
Process and the tow, counted label by label
A separate pass assigned theme labels, and a review can carry more than one. Some labels describe the business process, such as conduct, pricing or communication. Others describe the towing work itself. A single review can carry both kinds.
| What the labels cover | Share of the 5,916 reviews |
|---|---|
| A business-process label only | 72.4% |
| Both a process label and a core-service label | 18.4% |
| A core-service label only | 4.0% |
| Neither a process nor a core label | 5.2% |
The model assigned at least one core-service label to 22.4 percent of the 5,916 reviews. The multi-label pass and the main-reason pass answer different questions. One asks what a review mentions, the other asks what the model classified as the main reason, and the two are not reconciled review by review.
The model-assigned themes
The table lists every label in the taxonomy. These are themes a model assigned to what reviewers wrote, not verified events. A review can carry several labels. The shares overlap and are not summed.
| Model-assigned theme | Share of the 5,916 reviews |
|---|---|
| Staff conduct | 55.0% |
| Pricing and fees | 35.5% |
| Communication failure | 34.2% |
| Response time | 22.4% |
| Private-property or impound tow | 18.4% |
| Vehicle release | 17.1% |
| Vehicle damage or loss | 16.7% |
| Refused or abandoned service | 12.9% |
| Other | 9.9% |
| Unsafe operation | 7.1% |
Staff conduct
The model assigned the staff conduct label to the largest share, 55.0 percent of the 5,916 reviews. It is also one of the two catch-all process labels in the taxonomy. The two excerpts quoted here describe a driver and an employee.
“the driver yelled at me, told me he just pulled up the map and he would be there, then he hung up the phone.”
“The rude employee speaking to me refused to tell me why, just shrugging and saying it’s ‘policy'”
Pricing and fees
The model assigned the pricing and fees label to 35.5 percent of the 5,916 reviews. In the two examples here, the reviewers describe an added charge and a named fee.
“I was quoted a set price and needed towed .2 miles literally 2 blocks. When I got the bill it had an extra $40 tacked onto it.”
“a substantial portion of the bill was a $250 administrative lien-release fee”
The study did not check any bill or whether any fee was lawful.
Response time
The model assigned the response time label to 22.4 percent of the 5,916 reviews.
“I was quoted 1 hr arrival, they took nearly 2.”
“they said they would pick it up by noon, around 5:30 they finally got it”
In both quoted excerpts, the reviewer describes a time they were given and a later time they report.
Vehicle release
The model assigned the vehicle release label to 17.1 percent of the 5,916 reviews. Both quoted excerpts describe delays in getting a vehicle back from the company.
“Next day they tell me I cannot pick up my car until 4 days.”
“had to call someone to let me get my car out 8 times before anybody showed up”
Private-property and impound tows
The model assigned the private-property or impound tow label to 1,089 reviews, 18.4 percent of the 5,916. They are reported separately. In private-property or impound situations, the reviewer may not have chosen or hired the company.
Within the 1,089, the model assigned the pricing and fees label to 52.1 percent and the vehicle release label to 24.1 percent. When the 1,089 are set aside, 91.4 percent of the remaining reviews carry at least one business-process label, against 90.8 percent of all 5,916. The study did not measure local rules on private-property tows or impound fees.
Disclosure: this study was produced by Sales Roadmaps, which sells operations consulting services. About the operations consulting service
The number that keeps this honest
The 311 businesses sampled after exclusions hold 163,832 lifetime Google reviews between them. Of those, 19,747, or 12.1 percent, are one or two stars, and 84.6 percent are five stars. Of the 311, 285 have at least one review in the corpus studied here.
Percentages in this study use the denominator stated with each figure. Main-reason shares are of the 5,916 reviews the main-reason pass labelled. Any-label and theme shares for the full corpus are of the same 5,916 reviews unless a subset is stated, and the private-property and impound figures use the 1,089 reviews the model labelled that way. The base rate and the five-star share are of all 163,832 lifetime reviews. None of them is a share of customers.
Operational questions these reviews raise
These are questions for an owner reviewing their own operation. They are not findings. The study reports what reviewers wrote and how models labelled those reviews. Related reading on Sales Roadmaps covers dispatch boards, pricing discipline and route optimization.
How are staff expected to explain a decision?
The model assigned the staff conduct label to 55.0 percent of the 5,916 reviews and the communication failure label to 34.2 percent. One quoted example describes a reason that was not given, a situation the study did not count separately. An owner can examine what staff are expected to say when someone asks for a reason.
Does the final bill match the first number given?
The model assigned the pricing and fees label to 35.5 percent of the 5,916 reviews, and to 52.1 percent of the reviews it labelled as private-property or impound tows. The quoted examples describe an added charge and a named fee. An owner can review how added fees are explained, and treat private-property or impound cases separately, where the reviewer may not have chosen or hired the company.
How is an arrival time set, and who updates it?
The model assigned the response time label to 22.4 percent of the 5,916 reviews. The quoted examples describe arrivals later than the time given. An owner can ask how arrival estimates are produced and what happens when a truck will miss one.
How the study was done
Google reviews were retrieved through the DataForSEO business data API, sorted by lowest rating, up to 60 per business. In each of the eight metros, up to 200 listings were screened. After filters, the 40 businesses with the most reviews per metro were sampled, with at most two locations per brand. Nine businesses were excluded after their names were read: seven junk-car buyers, one national roadside-membership program and one auto reconditioning shop.
The corpus consists of the retrieved reviews rated one or two stars with at least five words, after exclusions and deduplication. No identical review texts were found under two listings. Candidate review-theme codes were proposed by a language model, Google Gemini 3.1 Flash-Lite, on a seeded sample of 200 reviews, and the research team consolidated them into a fixed label set.
Labels were then assigned by the same model. A second model, DeepSeek V3.2, relabelled 200 reviews and agreed on process against not process in 96.5 percent of them, with a kappa of 0.74. Each label was asked to carry a verbatim span from the review, and 95.8 percent of labels carry one. The rest are counted without a span.
As a label-based sensitivity test, which is not a main-reason result, the two catch-all process labels, staff conduct and communication failure, were ignored. On that basis, 66.5 percent of the 5,916 reviews still carry at least one process label. Reweighted by each business’s lifetime one- and two-star count, the any-label process share is 91.2 percent. With equal weight per business it is 89.2 percent.
The stored review data carry a business key, the star rating, the date, the text and an owner-reply flag only. No reviewer names were stored. No reviewer or reviewed business is named in this article, and the quotes carry no reviewer names, review dates or ratings.
Limitations
Google reviews are self-selected. The sample covers high-review-count businesses in eight metros: after filters including a cap of two locations per brand, the 40 businesses per metro with the most Google reviews among up to 200 listings screened, before exclusions. Review dates run from 2010-03-09 to 2026-09-30, and the sample is sorted by rating, not by date, so the study makes no claim about trends over time.
All labels and main reasons were assigned by language models, and no sample was checked against human coding.
Contact: Get in touch with Sales Roadmaps
Frequently Asked Questions
In this study, which model-assigned themes are most common?
Staff conduct, which the model assigned to 55.0 percent of the 5,916 reviews studied. Pricing and fees follow at 35.5 percent and communication failure at 34.2 percent. A review can carry several labels, so these shares overlap.
In this study, what share of reviews was classified as mainly about the tow itself?
In the main-reason classification, 11.6 percent of the 5,916 reviews were assigned to the tow itself and 86.9 percent to how the business is run. At least one core-service label was assigned to 22.4 percent.
How many reviews did the model label as private-property or impound tows?
The model assigned that label to 1,089 reviews, or 18.4 percent of the 5,916. The model assigned the pricing and fees label to 52.1 percent of them and the vehicle release label to 24.1 percent.
Among the businesses sampled, what share of all reviews are one or two stars?
Among the 163,832 lifetime reviews held by the 311 businesses sampled, 12.1 percent are one or two stars and 84.6 percent are five stars. That is a share of reviews, not of customers.
How closely did the models agree?
On the binary main-reason question, how the business is run against every other main reason, a second model agreed on 96.5 percent of 200 reviews, kappa 0.801. For the theme labels, process against not process, a separate second model agreed on 96.5 percent of 200 reviews, kappa 0.74. No sample was checked against human coding.
Does the study verify whether what reviewers wrote is true?
No. It reports what reviewers wrote and how often the model assigned each theme. No bill, arrival time or tow was checked against company records.
Related studies
- The negative dental reviews study
- The negative plumbing reviews study
- The negative electrician reviews study
- The negative law firm reviews study
- The negative auto repair reviews study
- The HVAC profit benchmark review
- The negative veterinary reviews study
- The negative pest control reviews study
- The negative moving company reviews study
- The negative property management reviews study
- The negative gym reviews study
