Feedback from colleagues and partners

Classmates suggested testing on reviews already labeled real or fake, starting with a 2011 deceptive review corpus, but I have not committed to that. I need to work through the detection research, and have no answer yet on false positives.

Reviews are the likeliest input, the outcome is shaping up as a re-ranked product list, and I have not locked in a system or category. LLM platforms may drop out of the first build, so that has no answer yet.

Fakespot's shutdown is my case study. I think scraping cost at scale was a big factor. A classmate suggested labeled datasets instead of scraping. I have not settled the cost side or how the tool keeps up with GenAI reviews.

I frame the tool around transparency and ranking credibility, not product quality, with the FTC Consumer Review Rule as my legal line for reviews. I have not decided where sponsored listings fall or how far into product quality to go.

No hard answer yet, but an estimated $770.4 billion in unwanted purchases in 2025 suggests shoppers miss fakes. Verified-buyer filters help only partly, since sellers pay people to buy and review. I have not decided how to use verified status.

I have a working list of 10 tactics with detection signals for each, shown under Research. Reviews are the first input, and I have not decided which other tactics make the first build.

Yes, and that is the claim I can stand behind. The demo shows the move from my baseline to my re-ranking with the rule behind it, and says nothing about where a product would have sat without the tactic. Platform rankings or listicles is still open, and I agree 1 of the 2 is enough for the first study.

That is my homework now: a map that sorts the 10 into easy, hard but doable, and beyond 1 project, with the datasets I can verify at their source. The main focus comes out of that map, and legitimate and unethical tactics may get different weights.

Taken. All 4 are open as I write this. The category is between a consumer product and subscriptions, and the outcome is between a platform's order and a listicle. I will pick 1 of each before the proposal, and the feasibility map decides most of it.

It is not, and the mock example only uses it because it is easy to read. A Bayesian average or a Wilson score lower bound, both of which weight a rating by the reviews behind it, is on the table for the real baseline, and the demo will follow once I choose.

By wording the claim as what the tool shows: how the order changes under rules I state, measured as rank changes and rank correlation, never as what caused the platform's order. The suggested question, how adjusting for signs of suspicious review activity changes the ordering against unadjusted ratings, is the shape I am working toward.