Review Methodology: How We Analyse Google Reviews

Every hotel page on Luxury Bangladesh shows Google's own rating alongside our analysis of the reviews behind it. Software and AI read those reviews, summarise what guests say and flag reviews that show patterns linked with fake reviews. We then show you what the rating looks like without the flagged reviews. This page explains each step, what we can and cannot detect, and how hotels can ask for a re-check.

What you see on a hotel page

Step 1: Collecting the reviews

Google's Places API

We use Google's Places API (New) to fetch three things for each property: its overall Google rating (from 1.0 to 5.0), the total number of ratings, and up to five reviews. Google returns no more than five reviews per place, sorted by relevance. They are useful to show you, but they are not a random sample, and five reviews are far too few to reveal patterns.

When we display Google content, we follow Google's attribution rules. That means Google Maps branding, the reviewer's name and photo linked to their profile, and a link to the review on Google Maps. Google recommends telling readers that reviews aren't verified by Google, and that is worth knowing in itself. We refresh the Google rating and review count at least every 30 days.

Google also writes AI review summaries for some places. Where Google provides one, we show it labelled as Google's, with Google's own disclosure text. As of October 2026, Google's documentation does not list Bangladesh among the countries where these summaries are available. The summaries on our Bangladesh hotel pages are therefore our own.

Larger review sets

Five reviews cannot show a burst of suspicious activity. For most properties we therefore also import larger sets of public Google Maps reviews from review-data export services. These contain only public information: the star rating, the review text and date, the reviewer's public name, number of reviews and photos and Local Guide status, and any reply from the owner.

We use these reviews for analysis, not for republishing. On our pages you will only see short quotes, together with a link to Google Maps for the complete reviews. Each hotel page states the period the analysed reviews cover.

Step 2: Checking every review for warning signs

Before any AI is involved, our software checks each review against a fixed set of rules. The rules are the same for every property, and every flag records which signals fired, so a result can always be explained.

No single weak signal is enough. A one-line review from a first-time reviewer is normal. The same review arriving in a burst of near-identical 5-star reviews from new accounts is not. Each review receives a suspicion score between 0 and 1 that combines its signals, with strong signals counting for more. Reviews above our threshold are flagged as suspicious.

Replying to reviews is good practice, and we never treat a property's replies as a warning sign in themselves. Only the patterns described above count.

Step 3: What the AI does

Next, a large language model running on Cloudflare Workers AI reads the reviews. It has to return structured data in a fixed format, and our software checks that data before anything is saved.

For each review, the model records:

• the overall sentiment, on a scale from −1 (very negative) to +1 (very positive);
• what the review says about each aspect: location, rooms, cleanliness, service, food, value and facilities;
• warning signs that fixed rules can miss, such as an incentive mentioned in passing or praise that contradicts the star rating.

For each property, it produces:

• a short summary of what guests consistently say;
• pros and cons;
• who the property suits best, such as business travellers, couples or families;
• aspect scores, with the number of reviews that mention each aspect.

We set firm limits on what the AI can do:

• Summaries are built only from the analysed reviews, never from the hotel's marketing.
• Quotes are short and verbatim. We never put a paraphrase inside quotation marks.
• The AI never decides alone that a review is fake. Its findings feed into the same suspicion score as the rules.
• An editor reads each summary against the underlying reviews before it is published, and again whenever the figures change sharply.
• We record which model produced each analysis, and when.

Step 4: The figures

Suspicious share

This is the percentage of analysed reviews that were flagged as suspicious. For example, if we analysed 400 reviews and flagged 24, the suspicious share is 6%. It describes the reviews, not the hotel's intentions.

Adjusted rating

In words: we take all the reviews we analysed, set aside the ones flagged as suspicious, and average the star ratings of the rest.

• It is shown next to Google's rating, never instead of it.
• It is based only on the reviews we analysed, for the period stated. Google's rating also includes star ratings left without any text, which we cannot analyse, so the two figures are not directly comparable.
• If the two are close, the reviews look organic. If the adjusted rating is clearly lower or higher, part of the headline rating may rest on reviews you should discount.
• We show no adjusted rating when too few reviews have been analysed for it to mean anything.

Trust score

The trust score is a number from 0 to 100 that sums up how much confidence the review pattern deserves. It combines:

• the suspicious share (the lower, the better);
• the gap between Google's rating and our adjusted rating (the smaller, the better);
• the number of reviews analysed (larger samples are more reliable);
• how evenly the reviews are spread over time (a steady flow is more natural than sudden waves);
• how recent the reviews are.

The trust score is about the reviews, not the hotel. An excellent property can still be hit by a fake-review attack, for example from a competitor.

Why we don't mark up Google's rating as our own

Search engines can show star ratings in their results when a site marks up ratings in its code. Google's guidelines say these ratings must come directly from the site's own users and that sites should not aggregate ratings from other websites. Google's ratings are not ours, so we display them with credit but do not mark them up as our rating.

Limits and error rates

Fake-review detection is probabilistic. A flag means a review shows patterns associated with fake reviews; it is not proof. We flag; we don't accuse.

• False positives happen. A genuine guest who writes one short review, a tour group whose members all post on the same day, or translated reviews that read alike can all trip our signals.
• False negatives happen too. Well-written fakes from established-looking accounts can pass undetected.
• We only see public data. Google can use account history, device and location signals that we cannot see, and it removes fake engagement under its own policies. Reviews that Google removes are invisible to us.
• Small samples mislead. Where we only have Google's five reviews, we show a summary but no adjusted rating or trust score.
• Language is hard. Reviews that mix Bangla and English, sarcasm and machine translations can be misread by both the rules and the AI.

We don't yet have a measured error rate, and we won't invent one. We will hand-label a random sample of analysed reviews, with the people labelling them unaware of our verdicts, and publish the results on this page: how often a flagged review is judged genuinely suspicious (precision), what share of suspicious reviews we caught (recall), the sample size and the date. We will repeat the exercise whenever we change the method. Until then, treat our flags as a guide, not a verdict.

Hotels: how to request a re-check

If you represent a property and believe our analysis is wrong, contact us with the subject "Review re-check" and include:

1. the address of the property's page on our site;
2. the figure or flag you believe is wrong, and why;
3. any evidence that helps, for example that a cluster of reviews came from a real group stay or event, or that Google has removed reviews since our last analysis. Please do not send us guests' personal data.

We will re-run the analysis with the latest reviews, have an editor examine the flagged reviews and explain the outcome to you. If we got it wrong, we correct the figures and note the date. We aim to acknowledge requests within five working days and to complete most re-checks within 15 working days.

Re-checks are free and follow the same rules as everything else. We do not drop genuine negative reviews from our analysis, accept payment, or let a property approve its summary before publication.

The rules behind this method

• United States. The Federal Trade Commission's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials (16 CFR Part 465) was announced on 14 August 2024 and took effect on 21 October 2024. It bans fake reviews and testimonials, including AI-generated ones. It also bans incentives conditioned on a review's sentiment, undisclosed insider reviews, company-controlled review sites presented as independent, review suppression through threats or intimidation, and fake social media indicators.
• United Kingdom. Since 6 April 2025, the Digital Markets, Competition and Consumers Act 2024 has banned fake reviews, concealed incentivised reviews and false or misleading consumer review information. Businesses that publish reviews must take reasonable and proportionate steps to prevent and remove them, as the Competition and Markets Authority's guidance on fake reviews (CMA208, published on 4 April 2025) explains.
• Google Maps. Google's policy says contributions should reflect a genuine experience, and that fake engagement is not allowed and will be removed. That includes paid reviews, content posted from multiple accounts by one person, and incentives for posting, changing or removing a review. Reviews based on a conflict of interest, such as those by current or former employees or by competitors, are not allowed either.

We are independent of every property we cover, and the same signals decide which reviews we set aside, whether those reviews are positive or negative. To see how this analysis fits into our overall scores, read how we rate hotels. Our editorial policy covers how we use AI more broadly.

This page was last updated on 22 October 2026.