An AI Agent That Actually Reads Your Site Then Links It Together For You

For every post, Linkilo’s AI extracts and stores: keywords in four tiers (primary, secondary, long-tail, semantic), user intent (informational / commercial investigation / transactional / navigational with confidence), funnel stage (TOFU / MOFU / BOFU with confidence), entity type (Concept, Product, Process, Person, Place, etc.), content format (How-to Guide, What is, Comparison, Listicle, etc.), topical authority signals, and both post-level and sentence-level OpenAI embeddings. All of this is stored in WordPress so subsequent suggestion runs are pure database math — no repeated API calls.

Every suggestion gets a composite score combining three weighted signals: post-to-post semantic similarity (75% weight, computed as cosine distance between OpenAI embeddings), keyword embedding overlap (15%, the share of keywords with cosine similarity above 0.75), and topic-cluster signal (10%, where same cluster = 1.0 and parent-child related cluster = 0.5). On top of that, a hard gate requires a usable anchor to literally exist in the source post — if no anchor fits, the candidate is dropped. The composite is exposed on every suggestion as a confidence percentage, and the reason text breaks down which signals contributed.

Linkilo never invents anchor text — it picks anchors that already exist verbatim in your source post, in four phases of decreasing precision. Phase 1: full target keyword appears verbatim. Phase 2: longest 2+ word sub-phrase of the target keyword. Phase 3: longest 2+ word sub-phrase from the target’s title (must include at least one content word). Phase 4: a single salient head noun, after a 150-term stoplist removes weak words like “ultimate”, “guide”, “perfect”. The justification on each suggestion tells you which phase won. English-aware morphology lets humidor match humidors while strict left-boundaries stop humid matching dehumidifier.

The metabox shows two engines side by side. AI Link Suggestions uses semantic similarity, keyword embeddings, and topic-cluster scoring — the highest-quality matches, requires an OpenAI key, costs OpenAI credits. Link Suggestions (the free engine) does literal keyword matching against your post titles and target keywords — no AI, no API key, no cost. Both work. Pick whichever fits the post. Many users run AI for cornerstone content and the free engine for everyday posts.

Two engines run automatically. The semantic detector compares OpenAI keyword embeddings via cosine similarity at a 0.85 threshold — this catches synonym variations and rewordings that look identical to Google but different to a literal SQL match (e.g., “best cigar accessories” ≈ “top cigar accessories” ≈ “essential cigar accessories”). It uses union-find clustering with transitive grouping and runs in seconds on up to 1,000 keywords. A SQL exact-match fallback runs on installs without an OpenAI key. Results are cached for 24 hours. The Keyword Conflicts page exposes everything for review with manual or auto-resolve actions.

For each conflict, Linkilo builds a rich per-post context including post title, URL slug, word count, whether the keyword appears in the title (with relevance score), keyword frequency & density, whether it’s a primary or secondary keyword for that post, AI analysis (intent, funnel stage, entity type, content format, topic cluster), and Google Search Console clicks & impressions when GSC is connected. This is sent to gpt-4.1-mini with the explicit instruction: “Prefer the post that is the most comprehensive and authoritative on the topic. Prefer pillar/guide content over specific examples.” The AI returns a single winning post ID. If you don’t want to use AI, a heuristic mode falls back to scoring on content length, keyword in title, frequency, URL match, and recency.

The shared keyword is removed from each losing post’s metadata (your blog content is never touched — only the keyword assignment). Then the Keyword Specialist kicks in: it sends all the colliding posts to the AI in a single call and asks for 2–3 specific long-tail keyword replacements per post that differentiate them from each other and reduce internal competition. You see each suggestion with reasoning attached, and can apply individually or in bulk. The result: the losing posts get more targeted keywords that they’re actually a better fit for.

Yes. The metabox works natively in the Classic Editor and the Block Editor (Gutenberg) via WordPress’ standard meta box API. We also ship dedicated integrations for Elementor, Beaver Builder, Oxygen, Kadence, Thrive Architect, Themify, Enfold, Cornerstone, Muffin Builder, Origin Builder, and WPRecipe. If you’re using a builder we don’t have a dedicated integration for, the standard Gutenberg metabox path usually still works.

Yes. Linkilo is built to complement, not replace, your SEO plugin. We read the focus keywords you’ve set in Yoast or Rank Math and treat them as user-curated — meaning AI re-extraction will never overwrite them. We also import keywords from Rank Math’s focus keyword fields. The two tools coexist cleanly: your SEO plugin handles meta tags, schema, and sitemaps; Linkilo handles content understanding, internal linking, and keyword strategy.

About a fifth of a cent per post analysed end-to-end (~$0.001-0.005 per post). The default chat model is gpt-4.1-mini at $0.40/1M input tokens, and the embedding model is text-embedding-3-small at $0.02/1M tokens. A 500-post site costs roughly $1-2 to process completely. After that, content-hash skipping means you only pay to re-process posts that actually changed. The conflict scan, link-suggestion engine, and per-post re-extraction all use cached embeddings for subsequent runs — pure database math, zero API cost. You can swap to GPT-4.1-nano, GPT-4o-mini, or even GPT-3.5-turbo from a dropdown if you want to drive cost down further.

No. AI processing happens in the background via WP-Cron — never on a visitor request. Bulk processing uses fastcgi_finish_request to send the response immediately and continue work in the background. The admin metabox lazy-loads suggestions on demand. Results are cached aggressively with content-hash skipping, so subsequent dashboard loads are pure database reads. The plugin has dedicated rate limiting, error handling, and a retry queue with stack traces for transparency. Your visitors don’t feel a thing.

When you turn on autopilot, Linkilo schedules a background sweep that processes new posts, refreshes suggestions, and detects new conflicts — without you opening the dashboard. The sweep frequency is computed automatically based on your publish rate and pending queue: more than 2 posts/day or 100+ unanalyzed posts triggers a twice-daily sweep; 0.5+ posts/day or 20+ unanalyzed triggers daily; quieter sites get weekly. You don’t configure cron schedules; the engine adapts.

Yes. Every time you click Apply Link or Reject on a suggestion, that decision is recorded. Rejections also open a feedback modal where you can flag the reason (wrong category, keyword mismatch, poor context, low-quality target, already linked, anchor text issue, content relevance, other). Once a week, a background job recomputes learned preferences from this feedback and uses it to bias future suggestion ranking — so the more you use Linkilo, the better the suggestions match your editorial taste.

Every suggestion includes a plain-English justification, e.g. “Post similarity 0.72, keyword overlap 33%, same topic cluster. Anchor ‘cigar humidor cabinet’ matches target keyword exactly.” A confidence percentage sits next to the title. For power users, enable Debug Mode in Advanced settings to see the per-source-post diagnostic funnel: candidates → passed similarity threshold → with anchor → returned, plus anchor method distribution (exact / partial / title / salient / none) and composite score min/median/max. Black-box suggestions are a non-starter here.

No. Use the Re-extract keywords for this post (1 OpenAI call) button in the metabox to refresh just that post in 3-6 seconds for a fraction of a cent. Manually-curated keywords are preserved. The autopilot sweep also auto-detects content changes via content-hash and re-processes anything that actually changed. The full Re-analyze post button under the Advanced disclosure rebuilds everything for one post if you want to be thorough.