Relevance matching: the new link building moat
Last Tuesday we shipped topic-embedding matching across the Bazsy marketplace. It's the biggest change to how the platform surfaces publishers since we launched, and the result is that buyers now see publisher recommendations that look nothing like them on paper but rank on the same keywords.
This post explains what changed, why it matters, and how to use it.
What was wrong with category matching
Until last week, our matching engine worked the way every other marketplace works: buyers select a category (SaaS, e-commerce, finance, gambling, etc.), and we show publishers tagged with that category. Simple, intuitive, and structurally limited.
The limit is that categories are coarse. A buyer in "B2B SaaS" sees publishers tagged "B2B SaaS", which means thousands of buyers and hundreds of publishers all converge on the same matches. The placements get bought repeatedly. The publishers get saturated. Over time, the placements look less editorial and more like a category placement farm, even when the underlying publisher is high-quality.
Worse, category tagging misses publishers whose category label doesn't match the buyer's but whose actual content overlaps significantly. A finance publication writing about workflow automation for finance teams is, in keyword terms, a perfect placement target for a B2B SaaS workflow tool. The categories don't match. The keyword universes do.
What changed
We replaced category-first matching with topic-embedding matching. Every publisher's recent content (their last 50 published articles, on a rolling basis) is encoded into a vector representation of what they actually write about. Every buyer's target page is similarly encoded. The matching engine surfaces publishers whose content vectors are closest to the buyer's target page vector.
The mechanism is similar to how semantic search works in Google itself. Publishers and buyer pages live in the same topical space, and proximity is measured by content similarity rather than tag overlap.
What buyers see now
The first surprise is geographic and category diversity in the recommendations. A SaaS buyer targeting accounting workflow keywords now sees:
- Traditional matches: accounting publications, B2B SaaS reviewers
- New matches: regional business journals with strong accounting coverage, fintech newsrooms that write about payments-and-accounting overlap, productivity publications whose recent content has skewed toward financial operations
The new matches are often stronger than the traditional ones because they're less saturated. A regional business journal that has written six pieces on accounting workflows in the past quarter ranks for those queries, has audience overlap, and isn't on every other SaaS buyer's shortlist.
Why this matters for ranking impact
Google's algorithm has been a semantic algorithm for nearly a decade. The 2013 Hummingbird update introduced semantic understanding. The 2019 BERT update strengthened it. The 2022–2024 wave of generative-search features deepened it further. What Google evaluates when assessing a link's relevance isn't the linking site's category tag. It's the topical proximity of the linking content to the linked content.
Topic-embedding matching aligns our recommendations with the signal Google actually uses. A publisher whose recent content sits two embedding units away from your target page provides stronger relevance signal than a publisher in the same category whose recent content sits ten embedding units away.
The placement looks more editorial because it is more editorial. The content fit on the publisher's side reads natural to both readers and to the algorithm.
What changes in your workflow
Most workflows don't need to change. The platform surfaces the matches; you select from them. The change is that you'll see more recommendations that surprise you, and we'd encourage you to evaluate them on the same quality criteria as your traditional category matches rather than dismissing them because the category label is unfamiliar.
For agencies managing multiple clients, the new matching engine produces more diversified shortlists per client. Two SaaS clients on your roster will no longer converge on the same 40 publisher recommendations. The diversification reduces the kind of placement-pattern overlap that algorithmic enforcement increasingly flags.
What's next
The embedding model retrains weekly on new publisher content. Publishers whose editorial direction shifts will surface in different match sets over time, which is the right behavior — a publisher that pivots from B2B SaaS coverage to consumer fintech coverage shouldn't keep ranking as a B2B SaaS match six months after the shift.
We're also experimenting with embedding-based anchor recommendation. The same vector representation that identifies relevant publishers can identify natural anchor candidates within an article. Early results suggest meaningful improvements in anchor naturalness over rule-based anchor generation. We'll share more once that work matures.
Relevance matching is live in your dashboard now. The recommended publisher list looks different. The bottom line is that the new list is closer to the publishers your content actually deserves.