Project journal

My Big Data class covered how to evaluate classifiers, and it maps straight onto my project. Spotting a fake review is classification on imbalanced data: most reviews are honest, so a model that flags nothing still scores 99% accuracy. Precision, recall and a cost matrix tell the real story. Flagging an honest seller and missing a fake cost differently, and a transparency tool should say which way it leans.

What changed: I may check 2 things separately: how accurate the detector is (precision, recall, AUC), and how much the final list moves from the baseline (rank correlation).

I sorted the tactics into 3 buckets. Real time: fake reviews (train once, score new ones live), affiliate links and disclosures, badge lookups, and hidden text through an on-demand fetch. After a delay: inflated "was" prices and scarcity counters, which need price history. Prototype only, mostly because of access limits: bought followers, astroturfing, backlinks and review gating, now marked as future research. Next tasks: tell undisclosed affiliates apart from unclear disclosures, list the badge registries per product type, and work out how long a "was" price check takes, maybe through the Wayback Machine.

What changed: every tactic has a bucket, and 4 moved to future research. Next: chapter 1 of my dissertation proposal, due by the end of October.

The purpose took shape: understand whether and how marketing tactics relate to what regular consumers buy, using the patterns they leave in rankings as evidence. 3 objectives: a feasibility map of the tactics, a 2-step prototype for 1 platform and 1 category measured by rank correlation, and a survey after IRB. The open question is the category: everyday products have the open data, software subscriptions are what I know. A classmate leaned toward a narrow software category if the data exists, and later posted a feasibility map of his own, with fake or paid reviews, review hijacking and review bursts as the focus and a Bayesian average as the baseline.

What changed: the project has a purpose and objectives, the baseline has a formula, and the category is still open.

I keep a folder of posts where people catch marketing in the act. This week: a viral post compared the same food order with and without the app's delivery subscription and found the promised savings mostly on paper. Another called out an airline ad as false advertising. New York City's click-to-cancel rule took effect on Oct 1: if you can sign up with a click, you must be able to cancel with one. And the Wall Street Journal reported on restaurants using AI-made food pictures that do not look like the food. None of the 4 is a product ranking, and all 4 are the same gap between what marketing shows and what a buyer gets. 3 of the 4 are about subscriptions, which is also the business I know best, so I am asking myself whether my category should be subscriptions rather than toothpaste. Not decided.

What changed: nothing yet. The category is an open question again, and 2 sources joined the reading list.

The question holds: can you trust a product ranking? The tool is a tactics-aware layer that re-ranks what a shopper sees and shows why. 2 new thoughts: legitimate tactics and unethical ones may deserve different weights, and a ranking can lag the tactic that moved it. My homework is a map of the 10 tactics on my list: what each needs to be detected, whether a labeled dataset already exists in the open (planted Reddit posts, for example, would make astroturfing a classification problem), and which are easy, hard but doable, or beyond 1 research project. The main focus comes out of that map. My 2 data partners get 2 questions after that: what they can help with in general, and with these 10 in particular.

What changed: the next step is a feasibility map of the 10 tactics, and tactics may carry different weights.

The project has its problem now: shoppers have no way to see how much marketing tactics pushed a product up or down a list. Star ratings are the weak spot. Luca and Zervas saw about 16% of Yelp restaurant reviews filtered, and He et al. saw bought reviews lift ratings only for a while. A list I pull is not a ranking I can defend, so a baseline comes first. Still open: platform rankings or listicles ("best of" articles). The class also got the mock example to try.

On the class worksheet the whole project fit in 1 sentence: whether, and how, marketing tactics relate to what regular consumers buy, with evidence from the patterns those tactics leave behind. That makes it a relational question, not an evaluative one. 1 row of my data is 1 product, and I already know 1 limit: I will not be able to say the tool works for every category. A classmate added a useful line: the re-ranking can show how positions change under rules I state, but not how many places a tactic caused a product to move, since nobody knows where it would have sat without the tactic. I agree, and the demo now says only that much. My professor's note went the same way and further: pick 1 platform, 1 product category, 1 definition of manipulation and 1 ranking outcome; weight the baseline by how many reviews support a rating (a Bayesian average or a Wilson lower bound beats a plain star average); and keep detection apart from cause. The feasible question offered: how does adjusting for signs of suspicious review activity change the order of products, against the unadjusted ratings.

What changed: the idea became a problem statement, the tool became 2 steps (build a baseline, then re-rank it), the claim shrank to what the tool can show, and the baseline rule is up for a stronger one.

I brought my mentor 10 marketing tricks. The tool is now a re-ranker: take public ratings, account for tricks, give shoppers a transparent ranking. Fakespot is the case study. Yang et al. (2027) survey nearly 2 decades of detection research. A classmate suggested Ott et al.'s (2011) labeled corpus. Classmates asked how I confirm fakes and measure false positives.

What changed: the tool became a re-ranker for marketing tricks, and transparency replaced deception.

Roughly 30% of reviews are fake, costing $770.4 billion in 2025. Luca and Zervas: Yelp filtered about 16% of restaurant reviews. He et al. found fake Amazon reviews sold on Facebook. The FTC's Consumer Review Rule allows $53,088 per violation. 10 companies got warning letters. My project: a consumer tool flagging products ranking on tricks, not quality. Likely 1 category.

What changed: the idea has numbers, plus a warning: Fakespot did something similar and shut down in 2025.

A professor and I discussed an Amazon browser extension counting fake-looking reviews, telling human from AI writing, and flagging marketing tricks. My mentor said to limit scope to 1 category and pointed to cold-start and content-based recommendation. Surveys need IRB training. Working definition: a fake review is fabricated by the company itself. Homework: 10 tricks so far, with detection methods.

What changed: the tool became a consumer-facing browser extension, and the scope shrank to 1 product category.

My focus is clearer: how 1 variable influences LLM recommendations, LLM citations, and search rankings. Candidates are reviews, backlinks, or social media footprint. Very likely reviews, testimonials, and case studies. Audience: online shoppers and SEO or GEO marketers. Zhang et al. (2023) found fake reviewers measurably hurt platform trust, which favors reviews. 2 professors helped me narrow it down.

What changed: a field became 1 variable. Classmates pushed for 1 input, 1 outcome, 1 system, 1 LLM.

People ask AI instead of visiting websites. Cited sites get seen. AI-answer ranking factors: recency, reputable backlinks, reviews on Google, Trustpilot, G2, or Capterra, and social media presence. 2 ideas: a legitimacy checker for websites AI cites, and a consumer tool summarizing product reviews with AI. Homework: more ideas, plus a look at Perplexity, NotebookLM, and OpenAI Deep Research.

What changed: the problem moved from marketing in general to how AI answers pick which websites to cite.

My topic is marketing and technology. Marketing is now about human behavior. I have about 5 years in corporate marketing, years running my own businesses, and am Director of Marketing at a tech company. Affected: anyone buying goods or services. My only source is Chen et al. (2025) on generative engine optimization: marketers influencing purchasing decisions at their core.

What changed: a field and 1 source, no problem yet, but GEO gave me a thread to pull.

Kickoff with my mentor for the applied project. I came in wide: marketing and technology. We discussed marketing as a science: understanding human behavior, then putting content, links, and offers in front of people, matched to their needs. AI moves fast and belongs in the picture. Homework: a couple of ideas connecting marketing and data science, to brainstorm together.

What changed: a whole field became a homework task: a couple of ideas due next week.

The first 2 semesters of the doctorate were coursework: methods, systems, and a lot of reading that had nothing to do with product rankings yet. No research question, no mentor meetings about it, no journal. The thinking got serious in the fall, and that is where the entries above begin.

What changed: the ground got laid. The research direction did not exist yet.