Chrome extension

Is this review worth trusting?

Highlight one review, one comment, one paragraph. SiftReviews quotes the phrases that make it read formulaic and the ones that read like a real buyer, then leaves the call to you.

Add to Chrome See how the reading works

Free. No account. Nothing you scan is kept beyond 24 hours.

Point at the text. Read the evidence.

Whether a machine wrote a review is the wrong question. A generated review can describe a real purchase, and a person can write three sentences that say nothing. The useful question is which phrases are doing the work, and whether the passage says anything specific enough to act on.

🖍️

You choose the passage

Highlight one review and scan only that. Nothing else on the page is sent. Selections under 200 characters are declined, because a confident score on two sentences would be worse than no score.

💬

The reading comes with quotes

Gemini 3.6 Flash returns the phrases it read as formulaic and the ones that read as first-hand, quoted from your passage. You can disagree with the reading by looking at the same lines it used.

🌐

The site is checked separately

The URL goes to Google Safe Browsing for malware, phishing, and unwanted software. That result is verified against an external list. The passage reading is an estimate, and the popup labels which is which.

What the reading means

Three plain-language readings, driven by how likely the passage is to have been generated. None of them is a verdict about the person who wrote it.

Reads like a real user
Under 30% chance generated

Uneven phrasing, specific detail, the kind of thing people write when they are not selling anything.

Mixed signals
30 to 60%

The band where evaluation showed scores are least reliable, so it gets a hedge rather than a call.

Reads formulaic
Over 60%

Even enthusiasm, tidy structure, nothing specific to one buyer. Worth a second look, not an accusation.

Formulaic writing is not proof of anything. A person can write this way, and a generated review can describe a real purchase. Scan a whole page instead of a selection and the popup still shows a combined 0 to 100 score: score = (domain_reputation × 0.4) + ((100 − ai_probability) × 0.6). That number is not shown on a selection, because averaging a site safety flag with a passage reading produces a figure that describes neither. The 40/60 weighting is a starting judgment, not a fitted parameter.

Built in the open

Most products don't publish what they cost to run or how fast they actually are. These are measured from the live service, and the cost figure is reconciled against the billing statement rather than the app's own instrument.

$0.0033
Inference cost per scan
~6s
Typical scan time
~90%
Of cost is model output
24h
Before your data is deleted
Cost and latency are read from the model API per request and shown on a live admin dashboard. The full write-up covers the code review that found three governance controls not enforcing what their documentation claimed, and the evaluation that found three lines of page furniture could move a score 50 points. It is in the case study.

Privacy by construction

The product needs to identify repeat users for rate limiting and nothing else. Everything follows from that.

✓
No accountsInstall and scan. There is no sign-up, because there is nothing to sign into.
✓
Your IP is never storedIt is hashed with a secret salt before it touches the database, and the hash cannot be reversed.
✓
Only what you highlight is sentIn selection mode the passage is all that leaves your browser. In page mode it is the first 2,000 characters, trimmed before the request goes out.
✓
Names are stripped in your browserAuthor names, reviewer handles, and bylines are removed from the text before it is sent. They never reach the server.
✓
The sub-processor is namedText goes to Google Gemini for the reading. Per Google's API terms it is not used to train their models.
✓
Deleted after 24 hoursRaw scan records are purged daily. Only anonymous totals survive.

What it can't do yet

A tool about trust should be honest about its own. These are open, not hidden.

!
It catches writing that sounds like an assistantAcross the evaluation sets, human-written text never scored high enough to be flagged. It also missed machine text that reads nothing like a chatbot. Trust a flag more than you trust a clean result.
!
There is no single accuracy number yetThe public benchmarks available separate genres about as much as they separate authorship. The curated set of real pages that would give an honest figure is being collected by hand.
!
Short selections are declinedUnder 200 characters the extension will not score. Plenty of real reviews are shorter than that, and those are the ones it cannot help with.
!
Page mode can pick the wrong blockOn a listing with dozens of short reviews, automatic extraction can return boilerplate instead of the text you meant. Highlighting the passage removes that guess.
!
The page weighting is a judgment call40% domain, 60% content was chosen by reasoning, not fitted to data. It stays a judgment until the curated set exists to fit it against.
!
One scan per dayA deliberate cost guardrail while the service runs on a small budget, not a paywall.

Install it

Available now on the Chrome Web Store. One click to install, no account required.

Add to Chrome
Click Add to Chrome above and confirm the installation.
Pin SiftReviews to your toolbar for easy access.
Highlight a review or a paragraph, open SiftReviews, and click Scan Selection.
Or open it with nothing highlighted and click Scan Page for a page-level read.
Open source. Want to see how it works or run it locally? Check out the GitHub repository.