What Hotels Say Back: 49,433 Owner Replies to Sri Lanka Guest Reviews, Analysed

Every study on this site so far has measured the guest. This one measures the property. Out of 363,148 verified reviews in the byc.lk corpus, 49,433 — 13.6% — carry an owner response. That is the only place in the whole dataset where a hotel, guesthouse or villa says something in its own words, and it turns out to be one of the most honest signals in the corpus: not because properties lie, but because the shape of what they write back tells you how they run the place.

We counted three things: whether a property replies, how long the reply is relative to the review, and what the reply actually contains. All three point in the same direction, and one of them contradicts the advice every review-reading guide gives.

1. The short answer

2. How often do Sri Lankan properties reply at all?

The base rate first, because almost everything else is relative to it. Of 363,148 scored reviews, 49,433 (13.6%) carry a non-empty owner response. Split by the score the guest gave:

• 1–4 (“bad”): 27.9% of these reviews get a reply.
• 5–6: 10.3%.
• 7: 9.2% — the lowest rate in the corpus.
• 8: 9.7%.
• 9–10 (“rave”): 13.6%.

Two things stand out. First, complaints get replied to at roughly three times the rate of a middling review — rational. Second, a perfect score is more likely to get a reply than a 7, 8 or a 5–6. That is not rational. It is the tell of a template: properties respond to the reviews that are easiest to answer, and “thank you so much” is cheapest under a rave.

There is a selection effect worth naming before anything else: reply rate and review score are entangled by construction. A property whose reviews are mostly complaints will show a high reply rate on the complaint band and a low overall score. That is exactly why every comparison below is either within a single band or within a single region.

3. The finding that matters: the reply that does not read the review

We took the 43,577 reviews that have both real text (40+ characters) and an owner reply, and measured the overlap between the words of the review and the words of the reply — the share of the guest's distinct content words that reappear in the response. Median overlap across the whole set is 0.06. That is, the median reply shares about one word in sixteen with the review it is answering.

Quarter the reviews scoring 8–10 by how much their reply overlaps them, and watch what happens to the group least engaged with:

• Least overlapping quarter: mean score 9.62, 89.7% raves, median overlap 0.00.
• Second quarter: 9.56, 88.4%, overlap 0.00.
• Third quarter: 9.58, 87.5%, overlap 0.12.
• Most overlapping quarter: 9.50, 85.3%, overlap 0.44.

Read that the right way round, because the direction is counter-intuitive. Within already-happy reviews, the ones whose reply engages with their words score slightly lower — and the reason is visible in the raw numbers: 40.0% of rave reviews get a reply with zero overlap, against 24.7% of 1–4 reviews. Properties do not engage with happy guests; they paste a thank-you. Overlap is a property of the complaint, not of the compliment, which is why it cannot be used as a quality score.

Then length, which is the cleaner signal. We bucketed replies by length relative to the review they answer:

• Reply under 20% of the review's length: the property averages 8.77/10 and 79.5% of its reviews are raves (n=4,414).
• 20–50%: 8.54, 76.6% raves (n=4,896).
• 50–100%: 8.15, 69.1% raves (n=6,391).
• 1–2× the review's length: 8.16, 66.6% (n=9,607).
• More than 2× the review's length: 8.22, 64.6% raves (n=18,269).

A property whose replies are longer than the reviews they answer is, on this data, not a property with a communications problem — it is a property whose guests arrive already unhappy. The correlation is entirely explained by the composition of the reviews: the mean reply to a complaint is 1,437 characters against 559 for a rave, because there is more to say when something went wrong. The useful reading is the one you can do on a single page: if every reply on a property's page is a paragraph five times the length of a glowing review, the property is replying defensively.

4. What the replies actually say

We classified all 8,522 replies to reviews scoring under 7/10. A reply can contain more than one signal, so these overlap:

• Thanks the guest: 84.7%.
• Apologises: 65.2%.
• Disputes the guest's account (“however”, “unfortunately”, “does not reflect”, “our policy”, “documented”): 39.9%.
• Escalates to a manager (names the role or invites contact with one): 34.6%.
• Invites the guest back: 21.2%.
• Names a specific fix — the reply says something was repaired, replaced, renewed or changed: 9.1% (774 replies).
• Offers money back or a discount: 9.5%.
• Gives a direct contact channel (email address, WhatsApp, phone): 4.8%.

Four in ten replies to a complaint open by contesting the guest's version of events. Two in five name a manager. Fewer than one in ten name a fix. That proportion is the single most useful number here: 78% of the effort in a Sri Lankan complaint reply goes into either thanking or defending, and roughly 9% into saying what changed.

There is an inverse relationship between the two: complaints whose reply disputes the guest average 2.59/10 (n=3,399); complaints whose reply apologises average 2.83 (n=5,560). Complaints that got a reply naming a fix average 2.39 against 2.89 for those that did not. None of that means fixing things makes guests angrier — it means the most damage produces both the strongest complaint and the most fix-shaped reply. The reply is a response to severity, not a cause of it.

5. The template problem: 31.4% of replies are copy-paste

We grouped every reply in the corpus by its exact lowercased text. 4,364 distinct reply strings account for 15,541 of the 49,433 replies — 31.4%. The largest single string appears 231 times, under 231 different guests at the same property. Others:

• “thank you very much” — 226 copies.
• “thank you so much” — 220.
• “thank you” — 157.
• “thanks for your review” + heart emoji — 148.
• “Property response: thank you” — 120. The literal field label is inside the text.

Nearly a third of the reply field in the largest Sri Lanka review dataset we can assemble is not writing. It is a default value. That matters most under raves, where a template is harmless, and least under complaints, where a template is the whole answer the guest gets.

6. Reply rate by region (and why it is not a quality ranking)

Share of reviews carrying an owner reply, by region, minimum 3,000 reviews:

• Pasikuda 25.8% (13 properties, mean 8.65) · Bentota 21.2% (8.81) · Arugam Bay 21.0% (8.41)
• Colombo 19.2% (8.22) · Sigiriya 19.2% (9.07)
• Nilaveli 16.3% · Negombo 14.5% · Yala 14.4%
• Unawatuna 13.8% · Kandy 12.4% · Trincomalee 12.2% · Anuradhapura 12.2%
• Nuwara Eliya 11.1% · Hikkaduwa 11.3% · Matara 10.6% · Galle 9.7% · Ella 9.4%
• Mirissa 8.0% · Weligama 7.7% · Rathnapura 6.2% · Jaffna 8.4%

Pasikuda replies four times as often as Rathnapura, and Bentota replies more than twice as often as Ella. But the rank ordering of reply rate is not the rank ordering of quality: Sigiriya (19.2%) has the second-highest reply rate and the highest property-weighted score of any major region at 9.07, while Bentota (21.2%) sits at 8.81 and Pasikuda (25.8%) at 8.65. Colombo has one of the higher reply rates and the second-lowest score. Reply rate tracks how corporate a region's accommodation stock is — chains and large hotels reply, family guesthouses do not — not how good it is. 101 of 1,193 properties with 100+ reviews have never posted a single reply.

7. The test that kills the “reply and you get better reviews” advice

The marketing claim is straightforward: responding to reviews lifts your rating. The corpus can actually be asked this, because a property's reply behaviour is observable independently of any individual guest. Here is the property-level comparison, 1,193 properties with 100+ reviews, grouped by the share of their reviews that carry a reply:

• Replies to 50–90% of reviews: 47 properties, mean score 8.64, 69.1% raves.
• 20–50%: 234 properties, 8.53, 67.4%.
• 5–20%: 389 properties, 8.63, 68.8%.
• Under 5%: 523 properties, 8.72, 70.9% raves.

Flat, and in the direction nobody expects: the least responsive properties score highest. That is confounded by property type and region, so we ran it again inside each region — high-reply properties (15%+) against low-reply properties (under 5%) in the same place, minimum three properties on each side. The high-reply group comes out lower in 19 of 27 regions, higher in 8, with a mean difference of −0.11. Negombo −0.34, Mirissa −0.40, Kandy −0.39, Bentota −1.25; Jaffna +0.74, Sinharaja +0.97, Nilaveli +0.61.

Our reading: replying does not lift a score, and the correlation that exists runs the wrong way because replying is what properties do when they need to. That is a real limitation of this data — we have no before/after window, no stay dates, and no way to see a property's rating trajectory around the moment it started replying. Anyone claiming a causal lift from a corpus like this is overreaching. What the corpus can support is narrower and still useful: a reply tells you what management thinks a problem is.

8. How to read a hotel's replies when you book

• Find the worst review and read the reply under it. That one reply is worth more than a hundred thank-yous. Does it name the thing that broke? Does it say what was done? Or does it say “we regret the misunderstanding” and describe policy?
• Look for the words of the complaint in the reply. If the guest wrote “the air conditioning dripped all night” and the reply never says air conditioning, you are reading a template. Median overlap in this corpus is one word in sixteen — the bar is low, and most replies clear it only by accident.
• Check whether the same paragraph appears twice. 31.4% of replies are exact duplicates. Two identical multi-line responses under different complaints is a property that has decided not to read.
• Do not treat a long reply as a good reply. Replies to complaints average 1,437 characters and replies to raves 559 — length tracks how angry the guest was, not how well the property handled it. A 200-character reply that names the fix beats a 2,000-character reply that restates policy.
• Discount the reply rate entirely as a ranking signal. Colombo and Pasikuda reply far more than Ella and Rathnapura and score worse. High reply rate is a proxy for corporate ownership, not for care.
• Value silence on a small property differently from silence on a big one. A 12-room guesthouse with no replies may simply be a family that does not use the platform. A 300-review hotel with zero replies (101 properties in this corpus have exactly that) has made a choice.

9. Method, limits, and what this does not show

The sample. 363,148 verified Booking.com guest reviews of Sri Lankan properties on a 10-point scale. 49,433 (13.6%) carry a non-empty owner response; 8,522 of those are replies to reviews scoring under 7/10. Phrase classification is regex over lowercased reply text and a reply can match several patterns, so the shares do not sum to 100. Duplicate detection is exact match on the lowercased, stripped reply string — it undercounts near-duplicates and cannot see edits of a single word.

A different cut of the same words. The five-language study applies the same counting method to the reviews rather than the replies, and finds that German, French, Spanish, Italian and English guests gave almost identical scores while writing about almost entirely different things.

The biggest limitation, stated plainly. The corpus has no reply date and no stay date. We can see that a property replies and what it wrote, but not when, so this study cannot measure response speed, cannot order a reply against the review that followed it, and cannot observe a property before and after it started replying. Every statement above is cross-sectional. The regional comparison inherits the same problem: high-reply properties differ from low-reply ones in size, ownership and price, and controlling for region only removes the geography part of that.

What we deliberately did not claim. We did not attempt to score reply quality with a language model. 29% of the corpus is German, French, Spanish and Japanese, and a model scoring English politeness would be measuring the wrong thing with great confidence. We also did not use the review’s score to infer whether the reply “worked” — the reply is written after the score, so the score cannot reflect it.

What would settle the open question. Reply timestamps plus stay dates would turn this from a snapshot into a before/after test: pick every property that started replying mid-corpus and compare its next 50 reviews against its previous 50. That is the study this data wants to be. Our word study showed the same structural lesson from the guest side — that the words people use are a better instrument than the score they give — and the original reply-rate analysis covered the 291,887-review slice this one extends.

Four later cuts of this same corpus. The fix-the-artefact index counts how often a complaint reply names something actually repaired, region by region. The reply-language audit separates the replies that name a fix from the far smaller share that say in the past tense it was already done. And the review-consistency study turns the same lens on the scores themselves: 69.0% of 363,148 reviews are a 9 or a 10, which is why an average describes almost nobody. A fourth cut, reviews read by language, asks whether the words inside them agree across five languages: the scores do, to within 0.07 of a point, and the vocabulary does not at all.

Frequently asked questions

10. The bottom line

Sri Lankan properties reply to one review in seven, and three times more often when the review is bad. That is the correct instinct, and it stops there: only 9.1% of complaint replies name a fix, 39.9% begin by disputing the guest, and 31.4% of all replies in the corpus are exact duplicates of another reply. The properties that reply the most do not score better — inside the same region they score 0.11 of a point lower — because replying is a response to trouble rather than a cure for it.

So treat the reply as the most specific piece of evidence on a booking page and the reply rate as almost no evidence at all. Read the one under the worst review. If it names what broke and what was changed, you have found a property that pays attention. If it says the guest misunderstood, you have found out what will happen to you too.

Method: 363,148 verified guest reviews of Sri Lankan properties (10-point scale) from the byc.lk corpus; 49,433 owner replies (13.6%). Phrase shares from regex over 8,522 replies to reviews scoring under 7. Duplicate share from exact lowercased string matching. Property comparisons over 1,193 properties with 100+ reviews, 0.5% two-sided trimmed means. Regional comparison restricted to regions with 3+ properties in each reply-rate group. No reply or stay timestamps exist in the data, so all results are cross-sectional. Last updated: 25 October 2026.

Plan your Sri Lanka trip with BYC at byc.lk — Free and honest. Some links are affiliate links and never change our advice.