8 Guild Blog

Your go-to source for the latest in web design, development templates, and creative digital solutions.

How to Use AI Assistants for Internal Linking: A Safe MCP Workflow for Large Sites

The safe way to use an AI assistant for internal linking on a large site: connect it to a dedicated linking tool through MCP, keep early sessions read-only, and approve links only after a dry run, in small batches you verify afterward.

That split matches how most SEO teams use AI. In Search Engine Journal's State of SEO 2026 survey (https://www.searchenginejournal.com/the-state-of-seo-2026-how-to-survive/555368/) of 371 SEO professionals, 58% were "Hybrid Strategists" who plan human-written content with AI support while "maintaining human oversight." Details come from vendor docs and the MCP specification as of November 2026.

What do you need before connecting an AI assistant?

You need three things: an AI app that supports remote MCP servers, a linking tool with an MCP server, and your site connected to that tool. The workflow fits teams that already use AI to update their site: the LinkBoss MCP server works on any website, so the assistant that edits your pages can also interlink them, fix past linking and optimize from Search Console data. Search Console access helps too, telling the assistant which pages deserve links first.

MCP, short for Model Context Protocol, is an open standard that lets an AI app call outside tools: the assistant asks, the server holds the facts. Remote servers connect over the web with OAuth rather than a pasted key.

How each app adds a remote server:

App How you add a remote server What to know about writes
Claude (web and desktop) Add a custom connector with the server URL Every plan; Free allows one connector
Claude Code claude mcp add --transport http <name> <url>, then /mcp .mcp.json servers get their own approval prompt
ChatGPT "Add custom MCP server" at chatgpt.com/plugins Business/Enterprise/Edu full; Pro read-only; writes need confirmation
Cursor .cursor/mcp.json per project or ~/.cursor/mcp.json globally Asks approval before using MCP tools by default
VS Code with GitHub Copilot An mcp.json file; HTTP first, then SSE fallback Org policy can limit availability

On ChatGPT Business, only admins and owners can create and publish custom servers.

Why use a linking MCP server instead of a WordPress connector?

A general WordPress connector can edit posts, but only a dedicated linking server knows your link graph, and that difference decides whether it can place links safely at scale.

The official WordPress MCP Adapter exposes WordPress abilities as MCP tools; what it can write depends on which abilities your site exposes. WordPress.com's MCP adds writing abilities for posts, pages and media on paid plans.

A general connector versus a linking-specific one:

Capability WordPress MCP Adapter LinkBoss MCP
Edit post content Yes, if the ability is exposed Adds and fixes links in content
Read every page's links and anchors Not by default Yes
Search Console data Not by default Clicks, rankings, queries, near-page-one and declining pages
Preview before writing Relies on your AI app's confirmation prompt Always shows a dry run with credit cost
Block conflicting anchors No Refuses an anchor already linked to a different page
Cost Free plugin Paid LinkBoss plan

Step 1: Connect LinkBoss's MCP server

Connecting takes a server URL and a sign-in, with no API key. LinkBoss is an AI-powered internal linking tool that builds semantically relevant links in bulk across WordPress, Shopify and other CMSs, used by over 8,000 SEOs. Its MCP server for internal linking (https://linkboss.io/internal-linking-mcp-server/) lives at https://mcp.linkboss.io/mcp.

Add the URL as a custom connector, log in to LinkBoss and click Approve to finish the OAuth sign-in; in Claude Code, run claude mcp add --transport http linkboss https://mcp.linkboss.io/mcp and then /mcp. The LinkBoss MCP setup guide (https://linkboss.io/knowledgebase/mcp-server/) walks through each app.

Check plans first: MCP is bundled with every paid LinkBoss subscription but not the free plan, a trial or a credit pack, and the cheapest tier, Launch, is $11 a month for one site and 200 credits as of November 2026. On ChatGPT, full access needs Business, Enterprise or Edu; Pro is read-only.

Confirm the connection with a harmless question:

"Which LinkBoss tools can you use? List them, and mark the ones that change site content."

Step 2: Audit before you add anything

Start with a read-only audit: it costs nothing, because reading, finding and fixing existing links doesn't use credits.

"Stay read-only. Give me the 20 pages with the fewest internal links pointing to them, every anchor text that points to more than one URL, and any internal links still using http://. Don't change anything yet."

Clear anchor cannibalization, http:// links and repeats before adding anything new, so later steps work on a clean graph; ask for the fixes as a dry run.

Step 3: Find the pages worth linking to first

Use Search Console data to decide where new links go: pages just off page one and pages losing clicks, not whatever was published last.

Under-linked pages are where new links pay off. In a SearchPilot split test on about 8,000 regional pages, linking each page to its six closest regional pages produced a 7% organic traffic uplift; SearchPilot concluded the pages "were starved of link equity."

"From Search Console, list the 10 pages closest to page one with the most impressions, plus any page whose clicks are declining. For the top five, find link opportunities from related pages. Show the source page, the exact sentence and the anchor score for each."

Drop any opportunity where the source page is off-topic, the sentence is weak, or the anchor is too generic for the target.

Step 4: Ask for a dry run

LinkBoss shows a dry run before any write, with the linked words in [brackets] and the credit cost; request one anyway, so the limits are yours.

"Prepare a dry run for the first three target pages: no more than 12 links in total and no more than two new links on any one source page. Prefer links inside existing sentences. Show the total credit cost."

What to check in the preview:

Step 5: Approve a small batch

Approve part of the dry run by number, and state what the assistant must not do; the cap is 25 links per call, but a first session is easier to check at five to ten.

"Apply links 1 to 7 from the dry run. Skip 8, 10 and 11. Don't add any link that wasn't in the preview."

Two safety nets sit behind you: LinkBoss refuses an anchor that already links elsewhere, and ChatGPT requires confirmation before write actions. Neither replaces reading the preview.

Step 6: Verify on the live site

Check the tool's log, then the page itself. LinkBoss records every change in Link History, and new links go live in about a minute.

"Show today's Link History entries. Then re-read the links on each source page and confirm every new link uses the approved anchor and target."

Open two or three source pages in a browser, confirm each link sits in the sentence you approved, and recheck the target pages' Search Console numbers in a few weeks before another batch.

Which guardrails keep an AI linking session safe?

Five rules: approve every write, distrust labels, grant minimal access, keep crawlers away from write tools, one anchor one target.

Keep a human in the loop. The MCP specification says "there SHOULD always be a human in the loop with the ability to deny tool invocations." Don't auto-approve write tools, and disable them before using Research mode, where Anthropic notes Claude may call connector tools without asking.

Don't trust labels alone. The spec tells clients to treat tool annotations, such as a read-only hint, as untrusted, and OpenAI's docs add that write actions can occur even on actions tagged read-only. Judge a tool by what it does.

Grant the least access that works. MCP security best practices recommend a minimal starting scope of low-risk reads, with broader access only when a privileged action is first needed. Connect only the servers a session needs; switch off write tools during audits.

Keep crawled content away from write tools. A crawler reads pages you don't control, the route OWASP names for indirect prompt injection. Keep write tools behind confirmation, or run them in a separate session.

One anchor, one target. A phrase linking to two pages muddies both; reject any proposal that reuses an anchor for a new target.

How reliable is the AI in this workflow?

The model is the least reliable part of the chain, so this workflow keeps its job small. Anthropic's Claude documentation (https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) states: "Even the most advanced language models, such as Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context." In Thinking Machines Lab's test, 1,000 runs of one prompt at temperature 0 gave 80 unique completions. Treat any URL the tool didn't return as suspect.

How to choose your setup

Choose by how much the assistant may change on your site.

FAQ

How many links should I approve in one session? Start with five to ten and read each one, then raise the number once the previews hold up. LinkBoss caps a call at 25 links, a sensible ceiling for one review pass.

Will an MCP session slow down my hosting? LinkBoss runs the heavy work on its own servers rather than your hosting, and changes go live in about a minute.

What happens to the links if I stop using the tool? With LinkBoss, links are written into the content as normal HTML, so they stay if you cancel or remove the plugin. Ask any other tool the same question first.

For more info visit our website 8guild.com

Find us on: