Scores trust, not popularity.
Deft reads public reviews and social posts about a brand, a talent or a public figure, rates how credible each source is, and flags coordinated activity, from bought praise to organised pile-ons. Every flag comes with the records behind it.
- Public data only
- Author handles pseudonymised
- A spend cap on every run
Coordinated praise suspected
Retail chain, Malaysia. Google reviews.
- Reviews in the cluster12
- Average sentiment+1.00
- Weighted source reliability0.07
- PatternNear-identical one-liners, one burst
The whole pipeline in 50 seconds.
From your brief to an evidence report a person reviews: plan, collect, clean, relevance gate, three separate scores, clusters, the coordination check, early warning and the report.
Read the transcript
One name in, evidence out. You brief and Deft plans: nothing is collected until you approve the exact plan and its spend cap. Deft collects public data only from Google Maps reviews, X, Threads and TikTok, honouring robots.txt. Deft cleans every post into one shape and turns author handles into salted pseudonyms. Deft judges relevance: posts about the subject are scored, namesakes are set aside, and unsure posts get a second check. Deft scores each post three ways, source reliability, output validity and sentiment, and never merges them. Deft groups similar posts into clusters. Deft checks whether a cluster looks organised, using a transparent rule on size, tone and concentration; in the pilot, one cluster scored 0.93 and was flagged for review. Deft watches for recurring topics and signals that rise before others, with confidence stated. You review with receipts: what was observed, what was inferred, and what is not known. Book a reputation read at deft-analytics.com.
The relevance cascade shown is a lab result; production uses a large language model today. Pilot figures come from an internal, anonymised study.
Watch it on its own page, with the full transcriptA star rating can't tell bought praise from the real thing.
Here are the same 12 reviews from the pilot, read two ways.
What a rating or listening dashboard shows
A great week for customer service.
- New reviews+12
- Positive share100%
- Service sentimentUp
Nothing to look at. Share it with the client.
What Deft shows for the same reviews
Twelve voices, but not twelve customers.
- Weighted source reliability0.07 of 1
- WordingNear-identical
- TimingArrived as one burst
Coordinated praise suspected, score 0.93. Records attached for a person to review.
It works the other way too. A pile-on from throwaway accounts can sink a rating overnight, and a coordinated attack looks like real anger until you check who is speaking. Deft checks the speaker before it counts the message.
What happens to every post.
Deft is a pipeline with checkpoints. A post has to clear each one before it can count toward a conclusion, and every checkpoint leaves a record you can inspect.
- Collect
Public posts and reviews from the sources in your approved plan.
- Relevance gate
Is it actually about your subject? Namesakes and look-alikes are set aside.
- Score
Reliability, validity and sentiment, each stored on its own.
- Cluster
Similar posts are grouped and checked for coordination.
- Evidence
Observed, inferred and not known, ready for a person to review.
The relevance gate
Every subject gets a card: its name, aliases, known namesakes and scope. Each post is labelled against that card in one of three ways. When a post doesn't carry enough context to tell, Deft says so instead of guessing.
Illustrative posts about a fictional subject. The gate is checked against human-labelled posts and reported per subject, not as one average.
Nova's mango smoothie at KLCC is way too sweet now.
About the subjectNova wallet down again, can't even pay for parking.
Namesake, set asidenova again?? seriously
Not enough context, set aside
Three scores, kept apart on purpose.
Most tools fold everything into one sentiment number. Deft stores three, side by side, so you can see why a conclusion was reached and challenge it if it's wrong.
-
Source reliability
0 to 1How much weight a source deserves, based on where it came from: the platform, the domain and how it was collected. A source that takes part in a flagged cluster loses weight.
Pilot cluster 0.070 · throwaway or unknown1 · established -
Output validity
0 to 1Whether each extracted record is well formed and grounded in the original text, so a summary can't claim something the source never said.
0 · malformed or ungrounded1 · grounded -
Sentiment
−1 to +1, plus intensityTone and how strongly it's expressed, scored in English and Bahasa Melayu. Extreme, uniform tone is one of the signals of coordination.
Pilot cluster +1.00−1 · hostile+1 · glowing
Nothing is collected until you approve the plan.
Every study follows the same bounded loop. You see the scope and the cost before a single page is read, and every run stops at its spend cap.
Brief
Name the subject and the question. Deft proposes the sources worth reading.
Plan and estimate
See the exact collection plan and what it will cost.
You approve
Runs start only against the plan you approved. Cancel any time.
Collect and analyse
Public sources are scored, grouped by similarity and checked for coordination.
Evidence report
Findings with scores, records, timestamps and the limits of each conclusion.
A flag is evidence for a person to review, not a verdict. Each report separates what was observed from what was inferred, and says how sure it is.
For the teams who answer for a reputation.
-
PR and crisis agencies
Is this backlash organic, or organised?
Bring the client evidence instead of a hunch: which accounts, how reliable they are, and how the activity was concentrated in time.
-
Talent agencies
Can we prove the support is real?
Show a brand partner that a talent's following holds up under scrutiny, and see a pile-on forming before it sets the narrative.
-
Brand and CX teams
Which of these reviews should we act on?
Separate real customer complaints from review farms and competitor sabotage before they skew a decision.
What Deft reads, and what it won't claim.
Sources today
- Google Maps reviews
- X
- Threads
- TikTok
- News and public web pages
- Languages
- English and Bahasa Melayu.
- Collection
- Public pages only. The crawler identifies itself and honours robots.txt.
- Privacy
- Author handles are salt-hashed before storage, so reports show stable pseudonyms, not identities.
Limits we state up front
- Public data is what was said, not proof of what is true. Reports say so.
- Early warnings are correlations with a stated confidence, not predictions.
- No private accounts, no logins and no attempts to re-identify authors.
- Reliability, validity and sentiment are never merged into one opaque score.
Questions people ask first.
What is Deft Analytics?
Deft Analytics is evidence-first reputation intelligence for Malaysia and Southeast Asia. It reads public reviews and social posts about a brand, talent or public figure, rates how credible each source is, and flags coordinated activity with the records behind every flag.
How does Deft detect fake reviews and coordinated activity?
Deft groups similar posts into clusters, then checks each cluster for signs of coordination: a tight group of near-identical posts, extreme tone, and activity concentrated in one source or one short window. The astroturf score is 1 minus the weighted reliability of the cluster's sources. Exact thresholds are stated in every client report rather than published, so they cannot be used to evade detection. Every flag goes to a person for review.
Is a Deft flag proof that reviews are fake?
No. A flag is evidence for a person to review, not a verdict. Each report separates what was observed from what was inferred and states what remains unknown, such as who organised the activity.
What data does Deft collect?
Public data only: Google Maps reviews, X, Threads, TikTok, and news and public web pages. Deft does not collect private accounts or log in anywhere, its crawler honours robots.txt, and author handles are salt-hashed before storage.
Does Deft use AI?
Where it helps. Today a large language model checks whether each post is really about the subject and helps score tone in English and Bahasa Melayu. Coordination detection is a transparent rule, not a black box, so every flag can be traced to its numbers.
Which languages does Deft support?
English and Bahasa Melayu. Heavy slang, sarcasm and mixed-language posts are harder to score, and Deft has not yet published an accuracy benchmark for Malay or mixed-language text.
How do I start with Deft?
Book a bounded reputation read. You pick one brand, talent or public figure and the platforms that matter, the scope and a budget cap are agreed before anything runs, and you receive an evidence report you can take to a client.
Start with one bounded reputation read.
Pick one subject and the platforms that matter. We agree the scope and a budget cap with you, run the study, and walk you through the evidence report.
Private beta. Onboarding is done with you, not by a sign-up flow.
- One brand, talent or public figure
- Scope and cost agreed before anything runs
- An evidence report you can take to a client