But agentic AI doesn\u2019t run on vibes. It runs on structured facts: consistent fields, reliable evidence, and clear confidence signals.
If you want AI to support real decisions (vendor ranking, expansion planning, rollout risk, investment priority), you need more than \u201ccategory tags.\u201d You need data you can trust and explain.
I wrote more about the broader retail context in my post \u201cFrom Raw Retail Data to Agentic AI\u201d<\/a>.<\/p>\n\n\n\n
My experiment: EuroShop\u2019s public exhibitor list<\/h1>\n\n\n\n
Instead of theorizing, I picked a dataset that\u2019s familiar to everyone in retail tech:
the public exhibitor list for EuroShop 2026.
Goal: turn it into something you can safely use for:<\/p>\n\n\n\n
- \n
- AI search (semantic + keyword)<\/li>\n\n\n\n
- Ranking & shortlists (explainable)<\/li>\n\n\n\n
- Predictions & planning (trend signals)<\/li>\n\n\n\n
- Chatbot \/ RAG training (low hallucination risk)<\/li>\n<\/ul>\n\n\n\n
The initial export looks \u201cokay\u201d\u2026 until you try to use it as an AI brain.<\/p>\n\n\n\n
The first surprise: the dataset is not the problem \u2014 the missing context is<\/h1>\n\n\n\n
A trade\u2011show directory is optimized for humans scrolling, not for machines reasoning.
In the raw-ish export (in this moment, maybe they will added it later):<\/p>\n\n\n\n- \n
- Only 43.5% exhibitors even have a description.<\/li>\n\n\n\n
- Only 35.6% have any social link.<\/li>\n\n\n\n
- PII is everywhere (emails in 90.5% of entries; phone numbers in nearly all).<\/li>\n\n\n\n
- And here\u2019s a fun one: 1327 exhibitors contained a placeholder image fragment inside category text (pure noise).<\/li>\n<\/ul>\n\n\n\n
When you ask an LLM to \u201canalyze\u201d this, it has two choices: be vague \u2014 or hallucinate.
So the real task became: create a dataset where the model can stay grounded.<\/p>\n\n\n\nThe second surprise: websites are 60\u201380% junk<\/h1>\n\n\n\n
If you crawl exhibitor sites, you quickly learn that most text is not product truth:
cookie banners, navigation menus, country dropdowns, footers, job pages, login screens\u2026
That\u2019s why the pipeline needed an aggressive cleaning layer before any extraction.
Otherwise you end up tagging a company as \u201cAI\u201d because their cookie policy says \u201cwe use analytics.\u201d<\/p>\n\n\n\nThe pipeline (plain English, no hype)<\/h1>\n\n\n\n
I ended up building a multi\u2011phase enrichment pipeline:<\/p>\n\n\n\n
- \n
- Crawl: start from the homepage, then prioritize the pages that usually contain truth (About, Products, Technology).<\/li>\n\n\n\n
- Clean: remove boilerplate, menus, cookie text; detect navigation patterns; truncate safely.<\/li>\n\n\n\n
- Extract: pull structured fields (what they do, products, capabilities, industries, tech signals) and capture evidence snippets.<\/li>\n\n\n\n
- Classify: assign business category, themes, and controlled tags (not free\u2011form chaos).<\/li>\n\n\n\n
- Score: separate truth confidence (how verified) from data richness (how complete) and flag quality issues.<\/li>\n<\/ol>\n\n\n\n
Key design choice: evidence-first. If a field has no evidence, it gets lower confidence \u2014 and your AI can be instructed to treat it as a hint, not a fact.<\/p>\n\n\n\n
![\"\"](\"https:\/\/darkopavic.xyz\/wp-content\/uploads\/2026\/01\/Bildschirmfoto-2026-01-24-um-12.12.30-1024x728.png\")
<\/figure>\n\n\n\nLinks to the technical deep dive:<\/p>\n\n\n\n
- \n
- App overview<\/a><\/li>\n\n\n\n
- Algo details<\/a><\/li>\n<\/ul>\n\n\n\n
Raw vs AI\u2011ready: the numbers that actually matter<\/h2>\n\n\n\nHere\u2019s what changed after enrichment (same 1,803 exhibitors):<\/p>\n\n\n\n
Metric<\/td> Original export<\/td> AI\u2011ready export<\/td><\/tr> Exhibitors<\/td> 1803<\/td> 1803<\/td><\/tr> Countries<\/td> 63<\/td> 63<\/td><\/tr> Approx. file size<\/td> 8.8 MB (JSON)<\/td> 33.4 MB (JSON, AI-ready)<\/td><\/tr> Websites present<\/td> 1717 (95.2%)<\/td> 1717 (95.2%)<\/td><\/tr> Descriptions present<\/td> 784 (43.5%)<\/td> 784 (43.5%)<\/td><\/tr> Value proposition generated<\/td> \u2014<\/td> 1587 (88.0%)<\/td><\/tr> Any social link found<\/td> 641 (35.6%)<\/td> 893 (49.5%)<\/td><\/tr> Avg. pages fetched per exhibitor<\/td> \u2014<\/td> 1.98 (max 3)<\/td><\/tr> PII included<\/td> Emails: 1632 (90.5%), Phones: 1803 (100.0%)<\/td> PII stripped: True (tracked via redaction counts)<\/td><\/tr> Avg. tags per exhibitor<\/td> \u2014<\/td> 2.61 (cap 12)<\/td><\/tr> Data Richness score (median)<\/td> \u2014<\/td> 47\/100<\/td><\/tr> Overall Confidence score (median)<\/td> \u2014<\/td> 53\/100<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n A couple of extra signals that became possible:<\/p>\n\n\n\n
- \n
- Business categories (distribution): services-other 593, industrial-manufacturing 562, retail-tech 428, storefit-fixtures 220.<\/li>\n\n\n\n
- Top themes detected: Sustainability 207, Digital Signage 200, Other 175, Logistics & Supply Chain 137, Store Design & Fixtures 131, Point of Sale 122, Lighting & Energy 119, Payments & Checkout 103.<\/li>\n\n\n\n
- Top exhibitor countries: Germany 502, China 236, Italy 154, Netherlands 76, Poland 66.<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/darkopavic.xyz\/wp-content\/uploads\/2026\/01\/Bildschirmfoto-2026-01-24-um-12.14.14-1024x767.png\")
<\/figure>\n\n\n\nWhat you can do with this dataset (real prompts, real value)<\/h1>\n\n\n\n
Below are examples that go way beyond \u201cshow me exhibitors in category X\u201d. They\u2019re designed to save time and reduce bad vendor meetings.<\/p>\n\n\n\n
1) The 10\u2011minute meeting brief<\/strong><\/p>\n\n\n\n
I\u2019m meeting {company_name} in 10 minutes. Create a 1\u2011page brief: what they do, 3 likely strengths, 3 risks\/unknowns, 5 sharp questions, and a recommended next step (demo\/pilot\/workshop). Use only evidence-backed info from the dataset, and cite the evidence fields.<\/p>\n\n\n\n
2) \u201cEdge cases only\u201d vendor shortlist (returns\/offline\/promos)<\/strong><\/p>\n\n\n\n
My biggest issues: offline transactions, cross\u2011channel returns, coupon stacking, and fiscal compliance changes. Find exhibitors whose capabilities suggest they can handle these edge cases. For each exhibitor: which edge case they likely handle, what evidence supports it, and what you\u2019d validate live.<\/p>\n\n\n\n
3) Build my EuroShop visit plan (who + where + when)<\/strong><\/p>\n\n\n\n
I have 1 day at EuroShop. I care about {themes}. Build an itinerary with 12 exhibitors: morning\/afternoon blocks, halls\/stands, and 2 goals per meeting. Prioritize high confidence + high richness profiles, but include 2 \u2018wildcards\u2019 with strong innovation signals.<\/p>\n\n\n\n
4) Create a retailer\u2011specific top\u201120 list<\/strong><\/p>\n\n\n\n
I run 180 fashion stores in 5 countries. Recommend 20 exhibitors for (checkout, loss prevention, ESL, store ops automation). Group them by problem. For each, give a 2\u2011line fit summary, key capabilities, and 3 questions that expose integration and operational risks.<\/p>\n\n\n\n
5) Vendor risk radar (what the dataset cannot prove)<\/strong><\/p>\n\n\n\n
For the top 30 exhibitors by innovation_score, list the top 5 unknowns you would validate before a pilot (integration, support model, certifications, regional readiness, references). Use data_quality_flags and confidence scores to justify each unknown.<\/p>\n\n\n\n
6) Competitive landscape map<\/strong><\/p>\n\n\n\n
Cluster exhibitors into 8\u201312 clusters based on themes + tech signals. For each cluster, describe the \u2018battlefield\u2019 (what they compete on), typical buyer persona, and what\u2019s changing right now.<\/p>\n\n\n\n
7) Find partners, not vendors (for POS \/ platform providers)<\/strong><\/p>\n\n\n\n
I\u2019m a POS platform vendor. Identify exhibitors that look like strong partnership targets (complementary tech, integration-friendly, clear value proposition). Give 10 targets, the partnership angle, and the evidence behind it.<\/p>\n\n\n\n
8) \u201cExplainable ranking\u201d for a decision meeting<\/strong><\/p>\n\n\n\n
Rank exhibitors for {use_case} using a weighted approach: 40% capability fit, 25% confidence, 20% evidence quality, 15% innovation signals. Output the score breakdown and the evidence lines used. No guesses.<\/p>\n\n\n\n
How to use it safely (so your AI doesn\u2019t overpromise)<\/h1>\n\n\n\n
A few rules I recommend when you plug this into an LLM, search engine, or agent:<\/p>\n\n\n\n
- \n
- Treat low\u2011confidence fields as hypotheses \u2014 not truth.<\/li>\n\n\n\n
- Prefer evidence-backed fields (snippets\/sources) when generating recommendations.<\/li>\n\n\n\n
- Use data_quality_flags to avoid ranking errors (e.g., no_pages_fetched, listing_based_extraction).<\/li>\n\n\n\n
- Keep the OPS dataset separate from AI-ready exports (PII vs non\u2011PII).<\/li>\n\n\n\n
- For prediction: never train on one signal; combine themes + capabilities + evidence + market context.<\/li>\n<\/ul>\n\n\n\n
What I learned (the part nobody tells you)<\/h1>\n\n\n\n
- \n
- Enrichment is less about \u2018more data\u2019 and more about better structure.<\/li>\n\n\n\n
- Controlled vocabularies beat free-form tags. Otherwise your embeddings become a junkyard.<\/li>\n\n\n\n
- Separating product technology from website stack prevents embarrassing false positives.<\/li>\n\n\n\n
- Two scores matter: truth confidence (trust) and data richness (coverage). Don\u2019t mix them.<\/li>\n\n\n\n
- Quality flags are not negative \u2014 they\u2019re what makes the dataset usable for serious AI.<\/li>\n<\/ul>\n\n\n\n
Want the dataset?<\/h1>\n\n\n\n
- Algo details<\/a><\/li>\n<\/ul>\n\n\n\n
- App overview<\/a><\/li>\n\n\n\n