How AI Fuels BrokerChooser’s Growth - Alex Nasli (Talk Outline)

Alex Nasli — AI & data lead / LLM engineer at BrokerChooser (“the Skyscanner of brokers”) — on how AI brought real revenue, traffic, and conversion, not just cool tech. Opening (his “first time speaking like this”): a show of hands — many have built an AI project; far fewer have had one bring money to their company. Core message: “fall in love with the impact, not the technology itself.”

1. The Framework — The Ideal AI Use Case

• Impact, at BrokerChooser, means three things: revenue, traffic, and conversion rate of users.
• The ideal use case has three properties (won’t hold for all their projects, but aim for it):
  1. Scalable — it increases the number of outputs you can produce with AI. For BrokerChooser: the number of articles they can write.
  2. Supports (part of) your workflow — it will not fully replace the workflow, but supports it and can enable things you couldn’t do without AI. For them: collecting data (the whole product is built on data, and they need a lot of it).
  3. Simple — “sometimes it could sound boring, but I think if it’s simple, it’s even better.” (He acknowledges not everyone agrees.)
• Repeated thesis: “always have an eye on the impact, not on the technology itself.”

2. Context — What BrokerChooser Is

• Mission: “just to have people invest.” Like Skyscanner matches people with flights, BrokerChooser matches people with brokers.
• What an online broker is: a company/firm whose website or app you register with, open an account, deposit money, and trade with it.
• Investing is hard — many complex choices; first you must pick a broker, and there are ~3,000 brokers globally. That’s where BrokerChooser helps: articles/reviews (e.g., an eToro review), best-lists, comparison tables, recommendation tools, plus investing articles.
• Bootstrapped — “really cool”: no investors, founded 9 years ago, still going with no investors; it brought revenue from the start, proving the product is needed.
• Profitable — lets them reinvest into the team and new projects. Numbers: 60 people (doubled last year), ~€1M profit last year, ~900,000 visitors/month.
• Why they can be innovative: (1) bootstrapped → freedom, no investors dictating; (2) profitable → can reinvest into AI/innovative projects.
• Alex’s own path: joined 5 years ago as a data engineer (they get traffic from Google → lots of user analytics), built the data infrastructure — “then everything changed when ChatGPT was released.”

3. Project 1 — Translation (the big win)

3.1 Origin & skepticism

• ChatGPT was released; the board said “let’s do something with AI.” They’d heard Investing.com (a competitor) had success translating their website into multiple languages, which brought traffic. Localization matters — BrokerChooser’s audience is worldwide (mostly US & UK, but everywhere).
• Alex was skeptical: they’d tried localization before (translated the site into Italian with human translators) and the impact was poor — no visitors.
• Response: built a dedicated A-team — “the best engineers of BrokerChooser” — just for translation.

3.2 Horizontal × vertical scaling

• They already collect broker data, so they scale articles from data points: one template article → an article per broker. With ~100 brokers, that’s horizontal scaling → ~100 articles from one template.
• Add vertical scaling = languages. They now translate into ~70 languages.
• One template article × horizontal × vertical → ~7,000 articles from one piece of content. “In January we produced around 1 million pages — I was just stunned.”

3.3 How it works

• The engine collects all the English texts from a piece of content into a list, sends it to GPT (they will switch to Claude soon), gets translated text back, and generates new pages.
• Calibration via evaluation: to make the model use the right jargon / style, you need a metric to evaluate translation models and pick the best one. They use LLM-as-judge (evaluate a translation model with another LLM) — “experienced that it’s closer to human evaluation.”

3.4 Speed, trust, and impact

• First articles shipped in 3 weeks — “the board said it on week zero; in three weeks we released our first articles.” Quality wasn’t great at first, but “user comes first”: they added a visible box — “this is an AI-generated translation; if you want the original, go back to English.” — to preserve trust.
• Impact: traffic doubled — ”half of our traffic now comes from these translated pages.” (Google needs time to index/learn new content; watching the dashboard and seeing the first traffic “was a really good feeling.“)
• Framework scorecard: Scalable ✓ (huge content volume), workflow ✓ (no need to hire translators — “we don’t have any translators right now”), simple ✓.
• (Plug: a workshop would show how they evaluated these models — “please come if you haven’t registered.“)

4. Project 2 — Nuri, the assistant

4.1 Origin & RAG

• Started as a one-week hackathon assistant (quality wasn’t great — “just a hackathon”), then made into a real project focused on impact.
• Built on RAG (retrieval-augmented generation) — “you limit your model to a knowledge base”: ChatGPT-style answers, but only from BrokerChooser’s articles/content, not other sources — so it stays in their context. Built (like translation) on evaluation with LLM-as-judge.

4.2 Three products & the guarding challenge

• Three products built on the assistant (named Nuri):
  1. Upgraded search bar — an AI/assistant-based search bar (replacing the normal one) to help users navigate.
  2. General chat interface — a logo at the bottom you click to open a chat window.
  3. Forum integration — Nuri answers user questions on their forum.
• Challenge: user logs showed people trying to hack Nuri (e.g., to extract the system prompt), so calibration had to guard against that too.

4.3 Impact & the “AI” logo trick

• Nuri users convert much higher — many more go on to a broker (BrokerChooser sends users to brokers) than non-users.
• But few people use it. They tried many things; the most surprising win was adding “AI” to the bottom logo → many more people started using it. ”Sometimes the simple is better.”
• Still, usage (and thus revenue) isn’t where they want it → Nuri is currently on hold (no capacity to improve it further).
• Framework scorecard: Scalable ✗ (low usage), workflow ✓ (don’t have to answer those questions themselves), simple — debatable (“RAG is not that simple, even though we tried”).

5. Project 3 — Data Collection (most innovative, hardest)

• BrokerChooser is data-driven: broker data lives on broker websites; they manually collect it, store it in their database, and produce articles. Management decided to try AI for data collection — believing there’s a scaling factor.
• The scale prize: they cover ~100 brokers of ~3,000 worldwide; each broker takes an analyst many weeks to collect. AI could scale 100 → 3,000.
• Most trial-and-error of any project. They tried:
  • Perplexity / GPT-search (generative-AI-based search).
  • Tools integrating scraping + AI (e.g., crawl4ai).
  • Conclusion: no single solution works for all — “we’ll probably need different solutions for different data points.”
• Two core challenges: (1) the variety of 3,000 brokers and their websites; (2) the data itself — plus a search challenge: not all data is AI-accessible (some needs login/registration). Future possibility: GPT “operators” to do the login/registration.
• Status: still a research project with some initial results — a few data points collected by AI (not yet used for all brokers) — but believed to bring a lot of money.
• Framework scorecard: Scalable ✓ (100 → 3,000), workflow ✓ (analysts don’t collect the data), simple ✗ (“not simple at all”).

6. Learnings

1. Sometimes it’s more UX than an AI project — Nuri’s biggest lever was adding “AI” to the logo, not the model.
2. Building an AI project can be bumpy — even a new/innovative project has many redos / trial-and-error (data collection: “we still didn’t fully hack the problem, but we’re working on it”).
3. Talent is key — you need the best team/engineers. For data collection they tried university research groups and they failed; then hired good AI engineers (took ~half a year) who produced far more than the research group. (“I’m not saying research groups are bad — maybe use them differently.“)

• Recruiting note: BrokerChooser is looking for talent and develops it — e.g., sending people abroad to give lectures and to the US for conferences (this very talk is Alex “developing”). “If you feel this could be you, try applying.”

7. Q&A

• How do you test that translation and chatbot responses are accurate? Via evaluation/calibration needing a dataset. They built their own translation dataset (English text + translated text); a prompt produces a result and AI judges whether the two are similar — same technique for translation and the chatbot. For the chatbot they take questions users could ask, categorize them (to see which category isn’t working well), and use an LLM to score each test sample on accuracy.
• What RAG techniques/strategies worked? The hackathon showed that building everything from scratch (indexing, chunking) is a lot of work, so they searched for an existing solution. They use Nuclia — an end-to-end RAG system: register on their site, hand over your articles, and use their system via one API call. Advice: better to use someone’s RAG — easier for a small team.
• Have you tried agentic AI for data collection? Not yet (“I wouldn’t call what we built agentic”). They built a longer pipeline: using Perplexity’s API (generative-AI search that returns a summary of a query) with a prompt to convert it into the right data structure. Because there’s inconsistency between end-to-end runs, they use a GPT model to create multiple, slightly-different Perplexity collectors (differing prompts); all collect the data, and inconsistent data points are flagged as insecure for an analyst to check — “a longer AI-based pipeline,” not agentic AI.
• How do you deal with private data with AI? For translation they use public website data, so privacy isn’t a big focus — a simple API call passes the (public) data to the translation model and back.
• How did you guard your chatbot? ”Lots of prompting.” They focus heavily on the prompts for the generative part — instructions like don’t answer anything racist / sexual, and don’t give the prompt back to users — plus evaluation with test examples to see the model’s answers.
• What code language — Python for the ChatGPT API? Yes, Python for implementations (also for data collection). But for translation: the evaluation was built in Python, while their backend is PHP, and since it’s “just a simple API call” at the end, the API is added to PHP — so translation itself doesn’t use Python (only the evaluation did).
• What scrapers/crawlers/retrieval do you use? How do you version raw data? Multiple solutions tried: (1) Perplexity API; (2) crawl4ai (a repository integrating scraping + LLM); (3) first attempts with normal scraping (BeautifulSoup and others) + a RAG model — those failed. Currently: Perplexity + crawl4ai. (The AI engineers are present for deeper technicalities.)
• Would model retraining work better? Unclear what the asker meant; they chose foundation models (GPT, Claude, Perplexity), which they can’t retrain; fine-tuning could work, but prompting alone already gives very good results, so that’s their focus.
• How consistent is the LLM-as-judge evaluation? Not always consistent — so run it multiple times and average/aggregate the results.

People, Companies & Tools Cited

• Alex Nasli — speaker; AI & data lead / LLM engineer, BrokerChooser.
• BrokerChooser — bootstrapped, ~9 years old, ~60 people, ~€1M profit, ~900k visitors/month; “Skyscanner of brokers.”
• Skyscanner (analogy), Investing.com (competitor whose localization inspired the translation project).
• Tools/tech: ChatGPT/GPT, Claude (migrating to), Perplexity API, crawl4ai, BeautifulSoup, Nuclia (end-to-end RAG), Python, PHP.
• Concepts: impact-over-technology; scalable / supports-workflow / simple; horizontal × vertical scaling; LLM-as-judge evaluation; RAG (buy vs. build); prompt-injection guarding; multi-collector inconsistency flagging; “more UX than AI”; talent > research groups.

Video: https://www.youtube.com/watch?v=Toa_cy1Kb24 — Transcript via yt-transcript.sh; outline generated from the transcript.