How AuthoritySource.ai Matches Buyers to Publishers

Most link marketplaces ask you to shop by domain — you scroll a spreadsheet of URLs, sort by DR, and hope the audience matches. We built AuthoritySource.ai the opposite way. Buyers never see the domain until an order is placed and accepted. Instead, every publisher site is classified, scored, and matched to your target URL by an AI pipeline that reads the actual content on both sides.

Here's how it works, end to end.

Why relevance beats DR shopping

Google's guidance on link spam and the way modern AI systems like ChatGPT, Perplexity, and Gemini choose citations both point in the same direction: topical fit matters more than raw authority. A DR 80 site that has never written about your niche is a worse placement than a DR 40 site whose readers are already looking for what you sell. Hiding the domain until match time forces the decision back onto fit — which is what search engines and LLMs actually reward.

The three things we classify

Every match involves three inputs, and all three get run through the same classifier:

1. The publisher's site — what they cover, who reads them, how commercial their audience is.
2. The buyer's website — the brand asking to be placed, so we understand context.
3. The buyer's target URL — the specific page being linked to or mentioned, which is often narrower than the brand itself.

Treating the target URL as its own subject is the part most marketplaces skip. A SaaS company might sell to marketers and finance teams; the target URL tells us which audience this particular placement is for.

Step 1 — Crawling with Firecrawl

When a site or URL is submitted, we fetch it through Firecrawl. We pull the homepage plus up to ~25 internal links, strip navigation and boilerplate, and keep the main article content. We also capture the page title, meta description, and detected language. That gives the classifier a representative sample of what the site actually publishes, not just a marketing homepage.

Step 2 — AI classification

The scraped content is sent to Google Gemini 2.5 Flash through the Lovable AI Gateway. The model is bound to a strict JSON schema, so every classification returns the same structured fields:

• Primary niche and up to five secondary niches
• Topic tags and audience tags
• Site type — blog, news, magazine, ecommerce, SaaS, agency, portfolio, forum, directory, other
• Language and up to six top countries
• Commercial intent — low, medium, high
• Restricted-category flag — adult, gambling, crypto, pharma, weapons, and similar categories that require admin review
• Risk flags — anything the model wants to warn a human about
• A confidence score from 0 to 100
• A public summary that describes the site without ever quoting distinctive sentences or naming the domain

That last field is important. The public summary is what buyers read on the marketplace. It's deliberately generic — "a UK-focused personal finance publication for early-career professionals" — so a buyer can judge fit without being able to reverse-engineer the domain.

Step 3 — Layering in DataForSEO metrics

Content classification tells you what a site is about. DataForSEO tells you how it performs. We layer in:

• Domain Rank — DataForSEO's authority rank, scaled to 0–100
• Organic traffic and ranking keywords
• Top 3 countries by organic traffic share

Those metrics are what you see in the Marketplace table. They aren't the whole story — a well-matched rank 200 site often outperforms a poorly-matched rank 500 site — but they help you sanity-check that a site has real, indexed readership.

Step 4 — The semantic matching engine

This is where the buyer's target URL meets the publisher's classification. Our matching engine (src/lib/matching.server.ts) does four things:

1. Tokenizes both classifications into normalized topic tokens.
2. Walks a niche graph we maintain — niches have parent/child relationships and "sibling" edges, so content marketing is close to SEO, and SEO is close to digital marketing.
3. Rewards semantic bridges. Direct overlap is best, but adjacency counts too. A "small-business advertising" target URL still fits a "marketing for SMBs" publisher even when no single tag matches exactly.
4. Combines that with the Site score — a quality signal derived from DataForSEO metrics, publisher history, and past order outcomes — to produce a Match score for that specific buyer/publisher pair.

The Site score is stable per publisher. The Match score is different for every buyer target URL — that's why the marketplace looks different depending on which target URL you're browsing from.

What buyers see vs. what stays hidden

Every publisher site shows up in the marketplace as an anonymized alias — {Niche} — {Identifier} — along with:

• Site score and Match score
• DR, organic traffic, ranking keywords, top 3 countries
• The AI-generated public summary
• Starting price for link insertions and/or brand mentions

The domain, the URL of the article we'd propose, and the publisher's identity stay hidden until you place an order and the publisher accepts it. That protects publishers from being scraped and pitched off-platform, and it forces buying decisions to be about fit rather than about a domain a buyer already had a bias toward.

Two placement types, one classifier

Publishers can choose to sell one or both:

• Link insertions — a contextual backlink placed inside a natural-fit article. The traditional SEO play.
• Brand mentions — an unlinked mention of your brand in a relevant article. This is the GEO (Generative Engine Optimization) play: LLMs like ChatGPT and Perplexity cite brands they see mentioned in trusted editorial contexts, whether or not the mention includes a link.

Both types use the same classification and matching pipeline. The only thing that changes is the deliverable and the price the publisher sets.

Human review and guardrails

AI does the heavy lifting, but humans stay in the loop at the moments that matter:

• Every new publisher site is reviewed by an admin before it becomes visible in the marketplace.
• Restricted categories cannot self-approve. If the classifier flags adult, gambling, pharma, or similar, an admin decides.
• Low-confidence classifications route to manual review instead of going live automatically.
• Every order ends with the publisher approving the AI-proposed article and writing the surrounding sentence themselves. The AI suggests; the publisher decides.

Continuous re-analysis

Sites change. When a publisher updates their site or when we improve the classifier, we can re-run analysis without demoting the site back to "awaiting review." Aliases refresh once a niche is known, and new classification versions are versioned in the database so we can always see how a site's understanding has evolved.

The takeaway

The marketplace you're looking at isn't a directory. It's the output of a pipeline that reads every site, understands every target URL, and scores every possible pairing before you ever see it. When a site shows up for your target URL, it's because the content actually fits — not because someone paid to be at the top of a list.

Ready to see it? Buyers can browse the marketplace, and publishers can list a site in about five minutes.