1,200 pages relinked over lunch

Internal linking is the most procrastinated task in SEO. Everyone knows orphan pages bleed authority, and nobody wants to read 1,200 pages to fix it. So teams either pay a frontier model to read the site one page at a time — slow, expensive, prose you can't act on — or they never do it at all.

Here is the reframe that makes it cheap: linking is not writing. For every passage, the question is "does page X have an honest reason to link to page Y, and is the anchor already sitting in the copy?" That is a yes/no judgement over a pair of texts. Ask it 8,790 times in batches and you have a link map by lunch.

The architecture

Notice what the model never touches: crawling, status codes, canonicals, existing-link detection. Those are facts, and facts belong in code. The model gets only eligible pairs and answers one thing — does this link deserve to exist?

Build it

One state per source-candidate pair, batched 128 at a time:

curl $WF_URL/predict/batch \
  -H "authorization: Bearer $WF_KEY" -H 'content-type: application/json' -d '{
  "states": [{
    "source_url": "/technical-seo-checklist",
    "passage": "A crawl reveals useful pages with few contextual links.",
    "candidate_title": "How to find orphan pages",
    "candidate_purpose": "Diagnose underlinked pages and reconnect them"
  }],
  "policy": "seo_internal_link"
}'

rows = run("internal_link", pairs)  # examples/seo_bulk.py
links = [r for r in rows if r["decision"] == "link"]

Or run the ready-made script: python examples/seo_bulk.py internal_link --out links.csv.

Reading the verdicts

Each pair returns a relevance score (0–3), two yes/no probabilities, and a confidence. The decision rule is deliberately conservative:

Expect a chunk of honest "no link" verdicts. A page the model refuses to link is a finding, not a failure: it tells you where the content has no natural next step.

What it costs

A 586-page site rebuilds its link map in under a minute for about twenty cents at metered rates — roughly 190x cheaper per page than a single-pass frontier run that got through 21 pages on the same clock. Self-hosted, it is just compute.

Pitfalls

• Never ask for URLs. The model picks among candidate IDs you already know. Invented slugs are how you ship 404s.
• Bound the candidate set. Three to ten eligible targets per passage. Beyond that you are paying for noise.
• Keep a holdout sample. Have an editor label fifty pairs first, then check agreement before you auto-queue anything.