Keyword vs semantic: the example that clears it up
A customer searches "dark circles" in a cosmetics store. Three engines react differently:
- Classic search engine (keyword). Returns 0 results if no product has the words "dark circles" in title or description.
- Search engine with manual synonyms. Returns "eye contour" products if the merchant previously set the rule "dark circles → eye contour." It works but requires constant manual upkeep.
- Search engine with semantic AI. Converts "dark circles" into a vector and compares with each product's vector. "Eye contour," "depigmenting serum," "decongesting patches" appear ordered by conceptual proximity. Without configuring anything.
Three scenarios where AI pays for itself
- High zero-result rate. If more than 10-15% of your searches return "no results," you're losing direct conversion. Semantic AI typically cuts that ratio in half because it understands variants and concepts.
- Mismatched vocabulary between customer and manufacturer. Your manufacturer writes "10% vitamin C serum"; your customer searches "anti-blemish for dull skin." Semantic search connects the two without you writing rules.
- Multi-language with uneven translations. Catalog translated into 5 languages but with thin descriptions in some. Cross-language embeddings let products be found even when the customer-language text is sparse.
Real cost of AI in search
The cost of inference (computing embeddings) has dropped sharply since 2023. Today a 10,000-product catalog is processed for under €1/month. The price difference between vendors doesn't reflect real cost — it reflects a commercial decision.
| Vendor | Cheapest plan with semantic AI | Catalog covered |
|---|---|---|
| Doofinder | Advanced ~€349/month (visual AI) | 400K searches/month |
| Algolia | NeuralSearch ~$500+/month | Custom |
| Klevu | From ~$449/month annual | Per contract |
| Findalo | Pro €49/month | Up to 6,000 products, 100K searches |
Frequently asked questions
What exactly is semantic search?
Search that understands what the customer's query means, not just what words they used. Behind it is a language model that converts both the query and each product into numerical vectors (embeddings), and measures semantic proximity. That's why "dark circles" finds "eye contour" — no word match, but close concepts in vector space.
Is keyword search still useful?
Yes, and in fact the best implementation is hybrid (what Findalo does): keyword + semantic at the same time. Keyword resolves exact searches ("Nike Air Max 90"), semantic resolves conceptual ones ("running shoes for long distances"). Removing one breaks the other.
Does AI make my search engine more expensive?
Depends. Some vendors charge AI as a premium add-on: Doofinder activates visual AI and personalization only on Advanced (~€349/month), Algolia NeuralSearch on ~$500+/month. Findalo includes it from Pro (€49/month) because your main fixed cost is the catalog, not the inference.
When is it worth adding AI to the search engine?
Three clear signals: (1) high zero-result rate (>10% of search traffic finds nothing); (2) catalog with non-normalized synonyms or technical vocabulary that differs from the customer's; (3) multi-language with uneven translations. If two of three apply, semantic AI pays for itself.
What about ChatGPT-style LLMs in search?
Two different things. Semantic embeddings (what Findalo does) are mature, cheap, and improve results without hallucinations. LLM generation (summaries, conversational answers) is less mature — higher cost, higher latency, risk of hallucination. We're evaluating for Phase D, but today's semantic search already covers most of the value.