Can Reviews and Support Conversations Stand In for Surveys When Response Rates Are Low?
Reviews and support conversations reach far more customers than a low-response survey. Score them the same way, measure each channel's tilt, and keep a small survey to calibrate.
Yes, for most of what a low-response survey is used for, provided you score every review and support conversation the same way and calibrate against the few survey answers you still get. In Dimension Labs research on about 10,000 survey responses, a language model reading only what a customer wrote predicted their own 0 to 10 rating within one point 67% of the time, and the predictions ran about one point lower, which a small calibration sample corrects. In the worked example below, reviews and support conversations heard from 5.7 times as many customers as a survey answered by under 1% of them, and each channel's tilt can be measured. For a team whose customers speak mostly in reviews, chats and tickets, Dimension Labs reads all of it with one set of scores and Dimensions, and keeps the survey as the check on them.
Below: whether reviews are a vocal minority, what the research says about estimating scores from text, and a worked example on synthetic sample data, a fictional aquarium with 95,000 visits in a quarter and a post-visit survey that 0.56% of visitors answer.
Are online reviews just a vocal minority compared to random-sample surveys?
Reviews are self-selected, and so are surveys. A study of over 280 million reviews on 25 platforms by Schoenmueller, Netzer and Stahl found that reviews skew to the extremes, mostly positive, and traced it to "polarity self-selection": people with strong experiences are the ones who post (Columbia Business School, 2020).
A survey with a low response rate is not a random sample either. Response has been falling even for the government's own surveys: the U.S. Current Population Survey's response rate fell from 87% to 69% between 2015 and 2025, and the establishment survey behind the monthly jobs numbers from 62% to 42% (GAO, June 2026). A post-visit email survey answered by a fraction of a percent of customers is answered by whoever chose to, which is the same problem reviews have.
So the useful question is how each channel tilts and whether you can correct for it. In the sample, the survey answerers were the happiest group, the reviewers were spread to both ends, and the people who contacted support were the least happy.
Sample data, labelled
Every channel tilts. In the sample, the survey's few answerers were the happiest group
NPS computed from the true 0 to 10 experience score of the visitors heard from in each channel, against all 95,000 visitors.
The survey's NPS of +21 is 18 points above the +3 you would get from every visitor. That gap was built into the sample to show the method; in your data the direction could be the other way. The point stands either way: a single channel answers for the customers who use it, and reading several, each with its tilt measured, is closer to the whole customer base than any one of them. Did our NPS drop, or did different customers answer? shows how a change in who answers moves the score on its own.
Can I estimate NPS from reviews and support chats instead of running a survey?
You can estimate the score a customer would have given from what they wrote, and our published research measured how well. On about 10,000 post-game survey responses from five Major League Baseball teams, GPT-4.1 read each open-ended answer and predicted the fan's 0 to 10 rating with an unoptimized prompt. Two thirds of predictions fell within one point, 36% matched exactly, and three separate scoring runs agreed exactly 87% of the time (LLM Predictive Scoring and Validation).
Dimension Labs research
From one open-ended answer, a language model predicted the fan's own rating within one point two times in three
GPT-4.1 read only the text of each response and predicted the 0 to 10 overall experience rating the fan gave, across about 10,000 post-game survey responses from five MLB teams.
The predictions were about one point lower than what fans reported. The paper reads that as a difference in what is measured: the fan's rating is an overall verdict, while the text dwells on the moments that stood out. In practice it means you keep a small survey and use the customers who both answered it and wrote something to set the offset, then apply it to everyone you heard from in other channels. That turns the gap into a correction you measure once, and the scored text then covers every customer who wrote.
Most of the remaining error sits in the text itself. A follow-up paper found prompt customization added about two points of within-one agreement, from 67% to 69%, while accuracy varied far more with how much the text said than with any choice of prompt or model (The signal is the ceiling). A one-line review gives a model little to go on; a long chat gives it a lot.
Support conversations work too, if you score satisfaction rather than tone. On 70,450 support conversations, an estimate of whether the customer was satisfied tracked their own ratings at a correlation of 0.47, against 0.36 for sentiment, and tone and satisfaction disagreed in 44% of conversations (What sentiment analysis can't see). A neutral-sounding customer may be quietly satisfied or may have quietly given up; a satisfaction score separates the two, which matters most in the channel where the problems are.
Turning scores into an NPS-style number is then arithmetic: count predicted 9s and 10s and predicted 0 to 6s per channel, apply the calibration offset, and report each channel next to the others rather than pooled. The predictive scoring course covers the setup and validation step by step.
Our post-visit survey response rate is under 1%. Can online reviews replace it?
They can replace most of it, and add what it misses. In the sample, 528 visitors answered the survey, 1,412 left a public review and 1,645 contacted support by chat or email: 3,028 unsolicited voices, 5.7 times as many as the survey.
They also hear about different problems. Survey comments ranked crowding and ticket value first and almost never mentioned timed-entry booking or refunds, because people with a booking problem write to support, not to a survey about the visit. Reviews and support together ranked booking first, at 6.9% of people heard from, against 0.8% in the survey.
Sample data, labelled
Reviews and support conversations put booking and refund problems at the top; the survey barely sees them
Six problem topics ranked by the share of people heard from who mentioned each one: survey answerers against public reviewers and support contacts.
What reviews and conversations cannot do is answer a question you never asked. If you need a recommendation score from a random set of visitors, or a rating of a new exhibit by name, keep a short survey for that and for calibration. Everything else, the ranking of problems, the trend by location and month, the share of unhappy customers, comes from the larger pool. For venues that see visitors every day, how attractions and venues analyze visitor reviews covers what a 1% response rate supports by day, week and month, and how many responses you need for driver analysis covers the sample size question.
How do I combine reviews, conversations and a small survey into one measure?
Score every record the same way. One predicted 0 to 10 satisfaction score and one set of topic labels for every review, chat, ticket and survey comment.
Calibrate on the overlap. Use survey answerers who also wrote a comment to measure the gap between predicted and reported scores, and apply it.
Keep the channel as a column. Report each channel's score and topics side by side; pool only after you know each one's tilt. How to combine reviews, surveys, tickets and calls covers the layout.
Count people, not records. A customer who chats three times about one refund counts once.
Check it against an outcome. Where you can tie customers to repeat visits or renewals, test whether the text-based score predicts them as well as the survey score did. That is the test that matters most.
Once reviews and conversations are scored and joined to what customers did next, they are the "what customers say" layer of a customer world model: a record you can use to see how customers reacted to past changes, not only how they felt last month.
Why read every review and conversation with Dimension Labs rather than wait for more survey answers?
Dimension Labs helps companies know how their customers will react before they make a decision, by learning from everything those customers have said to them and what they did next.
When response rates are low, that starts with hearing from everyone who already wrote:
It is fully omni-channel: reviews from every site and location, chats, tickets, emails, calls and the survey you keep, read with the same scoring and the same Dimensions.
It connects to the data that checks the score: visits, bookings, memberships and renewals from your systems or a file. Joining feedback to CRM and revenue outcomes, with effects and confidence intervals, is on Prove ($2,999 a month) and up.
Every interaction is enriched with a predicted satisfaction score and topics written for your business, from the pre-built library on Explore ($299 a month) or custom Dimensions from Discover ($999 a month).
The result goes out as dashboards, decks and PDF reports shared by link, so the monthly satisfaction report no longer waits on survey volume.
The NBA reports 100% Voice of Fan coverage across structured and open-ended feedback, and analyzes 100,000+ open-ended responses within hours instead of months (NBA story).
Your survey platform is the better choice if your survey still gets a healthy response from the customers you care about and its comments are all you need to read, its own text analytics is enough. Dimension Labs is built for the case where the survey has thinned out and most of what customers say arrives in reviews and conversations.
FAQ
Can AI predict NPS or CSAT from customer comments?
It can predict the rating a customer would give from what they wrote, directionally and consistently: within one point 67% of the time in our research, about one point low on average. Calibrate against a small survey sample before reporting it as a score.
Are online reviews a vocal minority?
Yes, reviewers self-select toward strong experiences. So do low-response survey answerers. Measure each channel's tilt and report channels side by side rather than trusting any one as the whole picture.
Should we stop running our survey?
Keep a short one. It sets the calibration offset and answers questions customers would not raise on their own. Let reviews and conversations carry the volume.
What is the best way to measure satisfaction when survey response rates are low?
For a business whose customers speak mostly in reviews, chats and tickets, Dimension Labs, which scores and tags every one of them and, on Prove, joins them to visits and renewals. If you only need to read a few hundred survey comments, your survey tool or a spreadsheet will do.
See what your reviews and conversations say that your survey does not: export a quarter of reviews or support chats, add your survey responses if you have them, and start with one file. Each plan from Explore to Scale can be tried free for 14 days.