Seo backlinks profile

Example output: examples/seo-backlinks-profile-stripe-com-20260514/PROFILE.md

Backlinks Profile

A complete backlink profile audit for a domain. Surfaces composition (where do links come from?), quality (what's the authority distribution?), diversity (concentrated in a few IPs/subnets, or spread out?), trajectory (growing or decaying?), and risk (which links look manipulative?). Output includes a health score and a reviewable disavow-candidate list — never an auto-disavow.

Single-source by design

This skill consults only the DataForSEO backlink index. We don't blend Ahrefs / Moz / Majestic / Common Crawl into the same report. That's a deliberate choice, not a limitation:

• Internally consistent metrics. Authority scores, anchor counts, and refdomain totals are computed against a single crawl. Multi-source blends produce numbers that look authoritative but actually average across crawls with different sampling, different freshness, and different definitions of "backlink" — the resulting ratios (e.g. dofollow %, anchor distribution) are noise.
• Reproducible health scores. The 100-point health score in this report can be re-run a quarter later against the same source and the deltas are meaningful. With multi-source blends, a score drift can mean anything: source A reweighted, source B refreshed, source C changed its toxic heuristic.
• No data-source independence to model. Any "do these sources agree?" question is unanswerable without a second backlink graph; we don't pretend to answer it. If you need cross-source confirmation (e.g. before legal disavow, before a high-stakes outreach campaign), pair this profile with a manual spot-check against Ahrefs/Majestic — that's a research task, not a skill output.

If your workflow specifically requires multi-source blending (large agencies, link-builders billing on link counts), this skill is the wrong tool — use a vendor that aggregates multiple indexes. For everyone else, single-source produces the more honest report.

Prerequisites

• DataForSEO MCP server connected.
• User provides: a target domain.
• Claude's WebFetch tool optional (for spot-checking flagged toxic candidates).
• mcp__firecrawl-mcp__firecrawl_scrape optional (for the new step 8b — link-source verification).

Process

1. Validate target & preflight. See skills/seo-firecrawl/references/preflight.md for the canonical 3-stage preflight (cost note, Firecrawl availability, Google APIs). Skill-specific notes:

   • Normalise domain before continuing.
   • Cost note: DataForSEO bills per call; this run issues ~15 calls. Use the documented limit/ceiling params to cap.
   • Firecrawl: optional. When --verify-sources is passed, step 8b (link-source verification) scrapes top-20 referring domains' linking pages to verify each link is still present and what rel it carries (dofollow / nofollow / sponsored / UGC), ~20 Firecrawl credits per run. Default off; pass --no-firecrawl to skip even if available.
   • Google APIs: not used.

2. Profile summary mcp__dataforseo__backlinks_summary

   • Total backlinks, total referring domains, dofollow/nofollow ratio, link-type distribution (text / image / form / frame), growth velocity over the last 30/90 days.

3. Referring domains mcp__dataforseo__backlinks_referring_domains

   • Top N referring domains by rank. Pull rank score, link count per domain, domain TLD, country.

4. Anchor distribution mcp__dataforseo__backlinks_anchors

   • Top anchor texts by frequency.
   • Classify each anchor: branded (contains brand name), exact-match commercial (the target's primary commercial keyword), partial-match, generic ("click here", "read more", "this page"), naked URL, image-alt-derived.

5. Rank distribution mcp__dataforseo__backlinks_bulk_ranks and mcp__dataforseo__backlinks_summary

   • Histogram of referring-domain rank (0–1000): how many rank 0-99, 100-199, etc. (Derive the histogram from the ranks of referring domains.)
   • A healthy profile has a long tail; an unhealthy profile is concentrated at low rank.

6. IP and subnet diversity mcp__dataforseo__backlinks_referring_networks

   • Total unique IPs hosting referring domains (set network_type to ip).
   • Total unique /24 subnets (set network_type to subnet).
   • Compute concentration ratio: referring_domains / unique_subnets. Healthy: ~3–10. Unhealthy: many domains share few subnets (PBN signal).

7. Growth / decay trend mcp__dataforseo__backlinks_timeseries_new_lost_summary, mcp__dataforseo__backlinks_bulk_new_lost_referring_domains

   • Net new backlinks per month (last 6 months).
   • Net new referring domains per month.
   • Velocity changes — sharp spikes or sharp losses both deserve flags.

8. Lost links list mcp__dataforseo__backlinks_bulk_new_lost_backlinks, mcp__dataforseo__backlinks_bulk_new_lost_referring_domains

   • Sample recent losses. Are any high-rank losses?

8b. Optional: live link-source verification mcp__firecrawl-mcp__firecrawl_scrape

• Triggered only when --verify-sources is passed (default off — credit-conscious).
• For the top 20 referring domains by authority (from step 3), pick the highest-authority linking page per domain. Scrape each (20 Firecrawl credits typical).
• For each scrape, parse the returned html for <a href> matching the target domain. Capture: link still present (true/false/page-404), rel attribute (dofollow if absent or empty, else the literal value: nofollow, ugc, sponsored, or combinations), surrounding context (anchor text + 50 chars before/after).
• Surface mismatches against the DataForSEO-reported state in evidence/08b-source-verification.md:
  • Link gone — DataForSEO still reports it as live (lag/error).
  • rel attribute differs from what DataForSEO flagged.
  • Source page returns non-200.
• Feeds into step 9: a verified-gone link or rel=nofollow discovered post-hoc upgrades the toxic-candidate signal for that referring domain.
• If Firecrawl unavailable (or flag not passed): skip entirely. DataForSEO's flagged state remains the source of truth — the skill's "Single-source by design" framing already explains why that's a deliberate trade-off.

9. Toxic candidate detection mcp__dataforseo__backlinks_bulk_spam_score (plus the heuristic — see Tips for the rules)

   • Pull spam_score for the referring domains — this is the toxic-signal proxy. Then apply the toxic heuristic to the referring-domain list.
   • Flag candidates. Each row gets a risk_score and triggers (which heuristic rules fired, including high spam_score).
   • Never auto-disavow. Output is a reviewable list, not an action.

10. Synthesise PROFILE.md

Output format

Create a folder seo-backlinks-profile-{target-slug}-{YYYYMMDD}/ with:

seo-backlinks-profile-{target-slug}-{YYYYMMDD}/
├── PROFILE.md                       (synthesised report — primary deliverable; inlines summary, authority distribution, diversity, trend)
├── 02-referring-domains.md          (top N with rank — load-bearing reference for outreach/audit)
├── 03-anchors.md                    (anchor distribution + classification — load-bearing reference)
├── disavow-candidates.csv           (toxic-flagged rows for review — load-bearing CSV)
└── evidence/
    ├── 01-summary.md                (backlinks_summary top-line — raw step output)
    ├── 04-rank-distribution.md      (histogram — raw step output)
    ├── 05-diversity.md              (IPs + subnets + concentration — raw step output)
    ├── 06-trend.md                  (last 6 months new/lost — raw step output)
    ├── 07-losses-sample.md          (recent lost backlinks)
    └── 08b-source-verification.md   (only if --verify-sources ran: live link + rel attribute checks for top-20 sources)

Step files 01, 04, 05, 06 are inlined as sections in PROFILE.md; the copies in evidence/ preserve raw step output for reproducibility. 02-referring-domains.md, 03-anchors.md, and disavow-candidates.csv stay at top level — outreach/audit teams consult them directly.

Note: 04-rank-distribution.md holds the referring-domain rank histogram described in step 5.

PROFILE.md follows this shape:

# Backlinks Profile: {domain}

> Snapshot dated {YYYY-MM-DD}

## Health score: **{n}/100**

| Dimension | Score | Notes |
|---|---|---|
| Rank distribution | {n}/20 | {comment} |
| Anchor diversity | {n}/20 | {comment} |
| IP/subnet diversity | {n}/20 | {comment} |
| Growth trajectory | {n}/20 | {comment} |
| Toxic candidate ratio | {n}/20 | {comment} |

## Top-line numbers

| Metric | Value |
|---|---|
| Backlinks | {n} |
| Referring domains | {n} |
| Dofollow / nofollow | {n}% / {n}% |
| Unique IPs | {n} |
| Unique subnets | {n} |
| Domain : subnet ratio | {ratio} |
| New ref-domains last 30d | {n} |
| Lost ref-domains last 30d | {n} |
| Toxic candidates flagged | {n} ({% of total}) |

## Rank distribution

| Rank bucket | Domains | % |
|---|---|---|
| 700+ | {n} | {%} |
| 500–699 | {n} | {%} |
| 300–499 | {n} | {%} |
| 100–299 | {n} | {%} |
| 0–99 | {n} | {%} |

## Anchor distribution

| Class | Count | % | Healthy range | Status |
|---|---|---|---|---|
| Branded | {n} | {%} | 30–60% | {✓/⚠} |
| Generic | {n} | {%} | 15–30% | {✓/⚠} |
| Naked URL | {n} | {%} | 10–25% | {✓/⚠} |
| Partial-match | {n} | {%} | 10–20% | {✓/⚠} |
| Exact-match commercial | {n} | {%} | <5% | {✓/⚠ over-optimised} |
| Image-alt-derived | {n} | {%} | <10% | {✓/⚠} |

## Trend (last 6 months)

| Month | New backlinks | Lost backlinks | Net |
|---|---|---|---|
| {M-5} | {n} | {n} | {n} |
| {M-4} | {n} | {n} | {n} |
| ... |

## Toxic candidates ({n} flagged)

See `disavow-candidates.csv`. Top 10 by risk_score:

| Domain | Rank | Triggers | Risk |
|---|---|---|---|
| {domain} | {rank} | {rank<100, sitewide>5, exact-match-anchor} | High |
| ... |

**⚠ NEVER AUTO-DISAVOW.** Hand this list to a human for review. Disavow a domain only after confirming the link is manipulative AND the domain is not delivering referral traffic AND removal requests have failed.

## Recommended next steps

1. {Action}
2. {Action}
3. {Action}

disavow-candidates.csv columns: domain,rank,spam_score,backlinks_count,sitewide_links,top_anchor,anchor_class,risk_score,triggers,sample_url

Tips

• DataForSEO allows up to 2,000 calls/min, 30 concurrent. Skills pace sequentially. The endpoints in steps 2–8 are ~15 calls.
• Cost: DataForSEO bills per call; a full profile run issues ~15 calls. Use limit/filters to cap large referring-domain/backlink lists. Optional step 8b adds 20 Firecrawl credits when --verify-sources is passed (one scrape per top-20 source domain).
• Toxic heuristic rules (any 2+ triggers = candidate):
  • DataForSEO rank < 100 or high spam_score (low-trust source).
  • Sitewide link count > 5 (footer/sidebar links across many pages — manipulation signal).
  • Exact-match commercial anchor on >50% of links from this domain.
  • Hosted in known link-farm subnet (when unique IPs / unique subnets ratio is heavily concentrated).
  • Domain name is a non-pronounceable string of characters (very strong PBN signal).
  • TLD is in the high-spam list (.xyz, .click, .work historically; verify against current spam-domain reports).
• Healthy anchor distribution: branded should be the largest class (30–60%); exact-match commercial should be small (<5%) — over-optimised commercial anchors trigger Penguin-era penalties.
• Healthy growth: steady 10–20% YoY referring-domain growth is the goal. Sharp spikes (>50% in a month) often indicate paid links and trigger algorithmic suspicion.
• Disavow conservatively. Removing links via outreach is preferred. Disavow only as a last resort; never disavow domains that send referral traffic.
• Pair with seo-backlink-gap for prospecting (gap analysis vs competitors).
• Pair with seo-drift to track profile composition over time.