How Alternatives Are Computed

Overview

Every product on Liners has an alternatives section showing similar products you might want to consider. We compute these automatically using a scoring formula based on Jaccard similarity — measuring what proportion of tags two products share, not just how many they have in common. I've got QA Quinn overseeing the quality of the results, and DD Dave runs the actual computation.

How the Scoring Works

When we compute alternatives for a product, we look at every other published product that passes the hard gates (see below) and calculate a relevance score. The higher the score, the more relevant the alternative.

Jaccard Similarity (Tag Scoring)

Instead of awarding fixed points per matching tag, we use Jaccard similarity for each tag group:

Jaccard = shared tags / union of tags

This measures the proportion of combined features two products share. A product with 20 tags that shares 3 with another product scores much lower than a product with 5 tags that shares 3 — because proportional overlap is a stronger signal of true similarity.

Each tag group has a different scaling factor:

Product Features carry the most weight because shared features are the strongest signal that two products are genuine alternatives. Technology overlap is a useful secondary signal, and other groups (business model, audience, etc.) contribute a smaller amount.

Country Scoring

Country overlap is a binary signal — either the products operate in the same market or they don't:

Pan-African products are treated as operating in every country, so they always get the same-market bonus. Products with no countries assigned yet are treated neutrally — no bonus or penalty.

Penalties

• B2B/B2C mismatch: -5 points. If one product is strictly B2B (business-to-business) and the other is strictly B2C (business-to-consumer), they're probably not great alternatives for each other, so we apply a heavy penalty. Products tagged as B2B2C or both B2B and B2C don't get penalized.
• Different market: -5 points. If both products have countries assigned but share none, they're likely not competing in the same market.

Note: Verified status does not affect the alternatives score. Alternatives are based purely on product similarity — what the products do, where they operate, and who they serve.

Hard Requirements

A product must pass two gates before it can be considered as an alternative:

1. Shared category — must share at least 1 category with the source product. This is enforced as a database filter when fetching candidates. Products in completely different categories are never scored.
2. Shared Product Features tag — must share at least 1 tag in the "Product Features" group. This prevents products from qualifying as alternatives based solely on country, business model, or category match without any actual feature overlap.

Minimum Score Threshold

Any candidate with a score below 7 is dropped. This keeps our alternatives list genuinely relevant — no noise, no random matches.

When Alternatives Are Computed

We compute alternatives at several points:

• When a product is published — we automatically calculate alternatives.
• When products are published in bulk — alternatives are computed for each one.
• On request — there's an option to regenerate alternatives when needed.
• Cascade recomputation — when a product's categories or tags change, we recompute alternatives for products that previously listed it as an alternative, and for products sharing the same categories or tags. Cascade fan-out is capped at 200 products per update to prevent runaway recomputations.

Cascade recomputations are processed via QStash (Upstash's serverless message queue) when configured. Each affected product is published as an individual job, processed with a concurrency limit of 5 to avoid overloading the database. When QStash is not configured (e.g. local development), recomputations fall back to sequential in-process execution.

How Alternatives Are Stored and Displayed

All alternatives are pre-computed and stored in the ToolAlternative table with a relevance score. Every surface — product cards, the product detail page, and the full alternatives page — reads from this same pre-computed data. This ensures consistent, scored results everywhere.

Example

Say Paystack is a B2B payments platform with Product Features tags: api, payment-gateway, payment-processing, payment-links, online-store, ussd. It's in the "Fintech" category, operating in Nigeria, Ghana, Kenya, and South Africa.

Teamapt shares the "Fintech" category (passes gate 1). Its Product Features are: api, bulk-payments, direct-debit, payment-gateway, payment-processing, pos. Shared PF tags: api, payment-gateway, payment-processing (passes gate 2). Union: 9 total unique PF tags. Jaccard = 3/9 = 0.33, scaled: 0.33 × 15 = 5.0. Both are B2B (+3 from Other group Jaccard). Same country (+3 bonus). Total score: 11.0 — well above the threshold.

Cowrywise shares the "Fintech" category (passes gate 1). Its Product Features are: investments, savings, wealth-management. Zero shared PF tags with Paystack — fails gate 2. Not considered as an alternative, which makes sense: a savings platform and a payment gateway serve fundamentally different needs.