A customer searches your help center for "export", opens the one article search returns, and reads it to the end. The answer they actually need, scheduled exports, is finished, published, and one collection over. Nothing in the article points to it. So the page just ends, and the nearest clickable thing is New Conversation. You wrote the answer and got the ticket anyway. At 300 articles, with no view of what links to what, every one of those dead ends stays invisible until you happen to reread the right article.
The job is information architecture at grep scale: find every mention of every topic that has a canonical article, link the mention to it, and surface the articles nothing points to at all. AI is good at exactly that. A help center is also exactly where you do not let it publish unsupervised, because this content steers customers who are already frustrated. So the linking happens in copies, local files the agent can search whole, and nothing it writes reaches a customer until you have read the diff.
Your options
Curating links by hand
The editor makes a single link pleasant enough. Highlight text, click the link icon, and a search box finds the target article. You can even copy an anchor link to a specific heading, and the anchor survives if the heading is reworded. For a 20-article help center, an afternoon of this is the honest answer. The wall is direction. Intercom shows you what an article links to, with no view of what links to it, so "every article that mentions SSO" means opening articles one at a time. When a user asked for a bulk way to maintain link sections, Intercom confirmed there is none and pointed at the API.
Intercom's built-in Related Articles
One toggle, and up to 5 algorithmically chosen articles appear in a block under every article body, in the Help Center and web Messenger. Zero upkeep, and on a single-product help center the picks are often reasonable. The fine print is the problem. You cannot edit the picks; Intercom's docs say so directly. The algorithm scores keyword overlap in titles, descriptions, and bodies and ignores collections entirely, which is why multi-product help centers can see suggestions from the wrong product. The block can silently fail to render, skips audience-targeted articles, and is absent from the Mobile SDK Messenger. And it only ever sits at the bottom. It will not put a link in the sentence your reader is stuck on.
A help-center app
The strongest apps in this category watch links rather than write them: daily scans for broken incoming and external links, freshness reminders for stale content, fix tasks created automatically. If link rot is your problem, that is genuinely useful and cheap to keep running. Cross-linking is a different job. Nothing we found in the category reads your help center, decides that the mention of audit logs in the billing article should point at the audit log article, and writes the link. The category monitors and migrates. It does not do information architecture.
An AI agent on the Articles API
The first option that can author links at scale. The Articles API lists every article with its full HTML body in a couple of paginated calls, search takes a phrase and a state, and an update call writes title, description, body, and placement. Wire an agent to it and the job is technically possible. Two costs. The work runs one call at a time: a search per phrase, an update per article, metered by rate limits, so 40 feature names against 300 articles is hundreds of round trips before judgment even starts. And every update to a published article takes effect the moment it lands, because the API holds no pending revisions. The agent is editing the page your customer is reading. Review exists only if you build it.
Scratch
Scratch turns the help center into the one thing the Articles API never hands you: a corpus you can search whole. Your help center comes down as a folder, one file per article, drafts and published, with conversations alongside read-only if you want the context. Now grep answers in about a second what the API answers in an afternoon: every mention of those 40 features, across all 300 bodies, in one pass. The agent links each mention to the canonical article, builds the inbound-link graph with a few lines of script, and lists the orphans. Every inserted link comes back as a word-level diff for you to approve, and a published change reverts per article. The review is yours, every changed sentence of it. On pages customers hit mid-frustration, that reading is the point.
| Option | Links in the body | Whole center in one pass | Review before live | Undo after publish |
|---|---|---|---|---|
| Curating by hand | Yes | No, one article at a time | You are the review | Edit it back by hand |
| Built-in Related Articles | No, a block at the bottom | Yes, automatically | Exclude-only, no picking | Toggle off, or exclude an article |
| Help-center app | No, it monitors | For monitoring, yes | Not its job | Not its job |
| Agent on the Articles API | Yes | Yes | Only if you build it | No |
| Scratch | Yes | Yes | Every link, as a diff | Per article, even after publish |
How the loop works on your help center
- Scratch pulls your help center into files. Every article lands as its own file: title, description, body, parent collection, and state, drafts sitting next to published so the map covers the whole center, not just the public half. Conversations can come along read-only, so the agent can see which questions customers keep asking and which answers they failed to find. All of it is a local copy; your customers are still reading the same pages.
- Your AI maps the center and writes the links. Point Claude, Codex, Cursor, or Copilot at the folder with the brief. For every topic that has its own article, find every other article that mentions it and link the first mention to the canonical article. Then list every published article nothing links to. The agent greps the bodies, builds the link graph, and applies judgment about which mention deserves which link. It does 99% of the work, and it holds no Intercom credential while it works.
- You review every diff and publish. Each inserted link shows as a word-level diff, the changed sentence next to the original. Validators run first if you set them, length caps and required fields, so an article never publishes broken. Scratch writes the approved articles back through the Intercom API and touches nothing else: drafts stay drafts unless a diff you approved says otherwise, and a published change reverts per article. The 1% you keep is the judgment a help center cannot delegate, where a stuck customer gets sent next.
Run the first pass on one collection. Once those links read right, the rest of the center, orphan list included, is the same brief.
What people use it for
- Link the first mention of every feature to that feature's article, in one reviewed pass instead of 300 editor sessions.
- Find the orphans: published articles no other article points to, reachable only if search happens to surface them.
- End every dead-end article with a real next step, so the bottom of the page is not the New Conversation button.
- Point every troubleshooting article at the setup guide it silently assumes the reader has finished.
- Carry the same link pass through translated articles, each language reviewed as its own diff.
Questions people ask
Can I choose which related articles Intercom shows?
No. The picks cannot be edited; the closest control is the on/off toggle plus excluding an individual article from suggestions in its Details panel. You still cannot put the right article in. The links inside your article bodies are the part you do control, which is why this page is about putting the right link in the right sentence rather than pruning a block you cannot steer.
Is there a bulk way to add links inside Intercom itself?
No. The editor adds links one article at a time, and when a user asked for a faster way to maintain link sections across articles, Intercom confirmed no bulk option exists and recommended the Articles API. Bulk is possible; Intercom just expects you to bring the tooling.
How does the agent find every article that mentions a feature?
Grep. With the help center as local files, a search for a feature name across 300 bodies returns every mention in about a second, including the off-hand ones buried in step 6 of an unrelated guide. The agent's judgment is the second pass: deciding which mentions deserve a link and which canonical article each one should point to.
How do I find the articles nothing links to?
From the link graph. Article bodies are HTML and links are hrefs, so the agent can chart which articles point to which and list every published article with zero inbound links. Intercom's articles report tracks views, reactions, and failed searches; it does not map links between articles, so this graph is one you have likely never seen.
Will the links break when an article title changes?
Rarely. Intercom logs URL changes when a title changes and redirects old links to the current article, and anchor links to a heading keep working when the heading is reworded. There are scattered reports of links breaking after edits anyway, so a periodic broken-link pass is still worth running. That pass is the same loop with a different brief, and it has its own guide.
Can I review every link before customers see it?
Yes. The order matters: the Articles API holds no staged revision for a published article, so anything written through it goes live the moment it lands. The review cannot come after the write. Scratch puts it before: the agent edits files, you read each changed sentence as a diff, and only approved articles are written back.
Will the AI read my customer conversations?
It can. Read-only, and only if you pull them. Conversations are useful context here: what customers ask in chat tells the agent which answers people fail to find, which is where a link earns its place. Nothing writes back to the inbox, ever.
Does this work for translated articles?
Yes, and a link pass adds a judgment call that plain text edits never face: a link inside a translated article should point at the translated version of its target where one exists, not bounce the reader back to the default language. The agent makes that call per language as it works, and you check it the way you check everything else here, each language version as its own diff before it ships.
Can it create a new article when a topic has no home?
No. What the link graph hands you is the gap list: topics that keep getting mentioned with no canonical article to point at. Writing those articles stays your job. Creating new articles is not supported; the loop covers editing what the help center already has, which includes wiring the inbound links in once the new article exists. If article creation is the half of this you need, tell Curtis.
See it on your own help center
Bring your real articles, dead ends and all. See it run on your help center →, or download Scratch free, pull your help center today, and start with the orphan list. The articles nothing points to usually justify the pull on their own.