You have 200 reviews, 500 support tickets, 1,000 social comments. Reading them all would take days — and you’d still miss the most important patterns. Sentiment Analysis and Topic Modeling solve exactly this: in ten minutes you get the emotional tone of every text, automatically grouped recurring themes, and a strategic summary ready for decision-making. Eighth installment of the Digital Stack 2026 series.
There’s a problem that scales very badly: manual text reading. With ten reviews it works fine. With fifty it starts to wear. With five hundred it’s impossible — and by the time you finish, the first ones you read have already faded from memory. The result is decisions based on the last ten reviews you happened to read, not on the entire corpus.
Look, here’s the thing: the problem isn’t a lack of data. It’s the human inability to process large volumes of text systematically and objectively. The brain isn’t designed for this — it tires, gets distracted, carries bias from previous readings. AI, on the other hand, analyzes text 500 with the same attention as text 1. And that’s exactly the operational advantage of Sentiment Analysis and Topic Modeling.
Sentiment Analysis: Assigning an Emotional Score to Every Text
Sentiment Analysis assigns each text a score representing its emotional tone. The standard range goes from -1 (extremely negative) to +1 (very positive), with 0 as neutral. It’s not a binary positive/negative classification — the score’s granularity is just as informative as the label: a text scoring -0.30 is very different from one at -0.85, even if both are labeled “negative.”
The practical value emerges when you analyze datasets, not individual texts. Seeing that 30% of your reviews score below -0.50 is a concrete warning signal. Seeing that the most negative-scoring texts all mention the same keyword — “refund”, “wait”, “courier” — transforms a perceived problem into an identified and localized one.
The Demo: 10 E-commerce Reviews Analyzed by Claude
To make this guide fully replicable, here are the exact texts used in the demo:
- Text 1: “Product arrived ahead of the expected date, careful packaging and item exactly as described. I’ll buy again without hesitation.”
- Text 2: “Waited three weeks for a response to my support email. The problem was never resolved and I had to open a PayPal dispute.”
- Text 3: “Within the norm. Delivery on time, product matching the photos. Nothing exceptional but no reason to complain.”
- Text 4: “Material quality disappointing for the price. I expected it to be more durable. I wouldn’t recommend it.”
- Text 5: “Excellent value for money. Customer support responded in under two hours and immediately resolved my size issue. Very positive experience.”
- Text 6: “The package arrived damaged. The item was fortunately intact, but the packaging was completely open. A rather unpleasant situation.”
- Text 7: “Professional service, clear communication at every stage of the order. The product exceeds expectations for the price range. Recommended.”
- Text 8: “Nothing particular to report. I received what I ordered within the stated timeframe. I didn’t need customer support.”
- Text 9: “Terrible experience. Wrong order, return process impossible to manage, and refund still not received after a month. Avoid this seller.”
- Text 10: “Beautiful product, very fast shipping. One downside: manual only in English. Everything else perfect.”
The prompt used: “Analyze the sentiment of these 10 texts. For each text: score from -1 (very negative) to +1 (very positive), brief explanation of the score (max 10 words). Output: markdown table with columns ID | Score | Sentiment | Reason.”
The table output immediately reveals the patterns. Text 9 (-1.00, Extremely negative) and Text 2 (-0.85, Very negative) are the critical cases. Texts 1 and 5 (+0.90, Very positive) show what’s working — timely delivery and responsive support. Text 6 (-0.30, Slightly negative) flags a packaging problem that would never surface in star ratings, because the item itself was undamaged and the customer likely left 3 stars out of 5. That kind of insight — the problem hiding in free text but not in the aggregate score — is exactly where Sentiment Analysis adds value beyond stars alone.
The Visual Approach: livechatai for Teams Who Don’t Want to Write Prompts
Writing a prompt and interpreting a markdown table takes some practice. If your team doesn’t habitually work with AI tools, a more immediate alternative exists: graphical interface tools that visualize sentiment without requiring any configuration.
The screenshot shows livechatai.com with the same texts in English. The pie chart at top left shows the percentage distribution of sentiment across multiple emotional dimensions — not just positive/negative/neutral, but Happy, Angry, Content, Frustrated, Joyful, Disappointed, Satisfied, Furious, Neutral, Excited. Below, each sentence is color-coded: green for positive (“This Product Exceeded My Expectations In Every Way”), red for negative (“Terrible Experience, I Will Never Order From Here Again”), yellow/orange for neutral (“Nothing Special But No Complaints Either”). The 1/9 navigation lets you step through results sentence by sentence.
This approach is ideal for teams sharing results with non-technical stakeholders — a colored pie chart is more readable in a presentation than a table of decimal scores. Claude and livechatai don’t compete: Claude for in-depth analysis and granular scoring, livechatai for communicating results externally.
The Word Cloud: the First Visual Snapshot of Your Corpus
Before even asking Claude to run structured topic modeling, there’s an immediate visual tool that helps you orient yourself in a text corpus without reading a single line: the word cloud. The largest words are the most frequent — in seconds it catches the eye on the thematic priorities of the dataset.
The word cloud above was generated from the 30 texts in Sets 2 and 4 used for the demo. The largest words — prezzo (price) and prodotto (product) — immediately confirm the two central themes of the corpus. Following in size: assistenza (support), spedizione (shipping), concorrenza (competition), internazionale (international), imballaggio (packaging) — exactly the clusters that structured topic modeling then identified systematically. A word cloud isn’t a substitute for topic modeling: it’s the visual starting point that helps form a hypothesis before asking AI for quantitative confirmation. Generating one takes under a minute on wordclouds.com — paste the texts and click generate.
Topic Modeling: Finding Hidden Themes in Hundreds of Texts
Sentiment Analysis answers “how does the customer feel?”. Topic Modeling answers “what are they talking about?”. The difference is crucial: you can have an average neutral sentiment score (+0.05) across 200 reviews, but if Topic Modeling reveals that 40% of texts mention shipping delays and 30% mention refund problems, you have two precise strategic priorities — not an average that says nothing.
Topic Modeling analyzes all texts simultaneously and automatically identifies natural thematic clusters: words that co-occur frequently, concepts appearing in blocks of similar texts, the percentage presence of each theme across the full corpus. You don’t define the themes in advance — they emerge from the data.
The Demo: 20 Texts, 5 Emerging Themes
The 20 texts used for topic modeling covered four thematic areas deliberately undeclared: shipping and logistics, product quality, value for money, and customer service. The prompt: “Identify the 4-5 main themes in these 20 texts. For each theme: name, main keywords (5-7 words), estimated percentage presence, brief description (1 sentence), text examples (sentence numbers). Output: structured list with bold headings.”
The detailed per-theme output is the richest part: each theme comes with automatically identified keywords, estimated percentage presence in the corpus, a one-sentence description, and — very useful — numbered examples with ✅ or ❌ to indicate whether the experience in that text was positive or negative. Theme 1 (Shipping and Logistics, 25%) shows texts 1 ✅ (next-day delivery), 5 ✅ (real-time tracking), 9 ❌ (4-day delay without communication) and 13 ❌ (package left in the rain). In a single view you see that shipping is polarized — not uniformly good or bad, but split.
The Thematic Summary: the Table That Drives Decisions
The summary table is the operational document of the entire process: five themes with percentage presence and prevailing tone for each. The most interesting data point is Value for Money at 25% with tone “Negative (2 pos / 2 neg / 1 neutral)”: it’s the theme with the most critical tone in proportion. This doesn’t necessarily mean the product is overpriced — it means that when customers talk about price, the perception is predominantly negative, and that’s a signal to address in positioning or value communication. Shipping, Customer Service, and Product Quality all show mixed tone — both positive and negative experiences, and the next task is figuring out what differentiates them.
Combining Sentiment and Topics: The Complete Picture
The real value emerges when you cross the two outputs. From Sentiment Analysis you know 20% of texts score below -0.50. From Topic Modeling you know the dominant themes are Shipping and Customer Service with mixed tone. Crossing them: which negative texts are about shipping? Which about customer service? The answer — obtainable with a third prompt: “Which negative texts (score < -0.50) belong to the Customer Service theme?” — localizes exactly where to concentrate interventions.
This is the shift from impressions to data-driven decisions: not “I have the feeling customers are complaining about shipping” but “25% of feedback concerns shipping, 40% of that has negative tone, the three recurring problems are: delays without communication, damaged packaging, unreliable courier.”
Three Limits Not to Ignore
Sentiment Analysis struggles with irony and sarcasm. “Fantastic, the third damaged package in a month” contains positive words but clearly negative sentiment — generic models often misclassify it. The practical fix: add to your prompt “Pay attention to irony and sarcasm: interpret contextual meaning, not literal words.”
Topic Modeling on Italian texts is less precise than models trained primarily on English-language corpora. Claude and GPT-4 handle Italian well, but theme presence percentages should be considered indicative, not exact — especially on small corpora (under 50 texts). Manually verify a 15-20% sample of results before using them for important decisions.
Finally: never upload texts containing personally identifiable data to cloud services without prior anonymization. Replace names, email addresses, and specific references with codes before uploading — this is a GDPR requirement, not an option.
Up Next: Supabase and the Backend for Vibe-Coded Apps
With Sentiment Analysis and Topic Modeling we close the advanced text analysis block. Next week we shift gears: from analyzing data to storing it in real applications. Supabase is the open-source backend that completes apps built with Lovable and Bolt — managed PostgreSQL, free authentication and storage, Row Level Security explained simply. Everything you need to take a vibe-coded app from demo to production. Sentiment analysis and topic modeling tell you what your customers think. The next step is building the systems that collect that data in a structured way.
