Help Centre Articles: The Gold Mine for AI Citations and New Customers

Your customers are asking ChatGPT about your product right now. “What are the fees for {product A}?” “How long does a withdrawal take on {product A}?” “how can i do this on {product A}?” The assistant answers with whatever it can find written down, and very often that is not anything you wrote.

It might be a Reddit thread from two years ago that’s no longer relevant, a competitor’s comparison page or a blogger’s half-finished tutorial. Meanwhile, the one place where your company states those facts officially, your help centre, sits on a vendor’s domain, covers half the questions customers ask, is set to noindex, or  evenlives in a Notion doc only your support staff can open. And it was written for the support team rather than the people searching.

In this guide I’ll show you why help centre articles are the most natural first-party source for AI citations, the hosting mistake that quietly gives that value away, and how to build a help centre plan from the complaints your customers already make.

Why AI assistants cite help centre articles

When someone asks an AI assistant a factual question about your product, the strongest source is the company itself. Help centre articles are where a company writes those facts down: branded, specific and in plain language, usually one question per page. That is close to the exact shape an AI answer needs.

I’ve seen this in the data. For example, a product i once worked on, a site of more than 3,000 pages, the help centre earned 105,000 of the 499,000 impressions the whole site got in Google’s generative AI features in September 2026. That is about 21% from a small corner of the site.

Wise shows what a good one looks like. Its help article, “How long will my transfer take?”, sits on wise.com, uses the customer’s own words as the title, answers in the first paragraph, then links straight to the common reasons a transfer runs late.

What Aleyda Solis’s prioritisation framework says about help content

In August 2026, Aleyda Solis published a content prioritisation framework for the AI search era, with a free worksheet. It asks which content is still worth creating when an AI answer can satisfy a question without a click, and scores each content type on six criteria: click resilience, citation potential, brand mention potential, business value, proprietary advantage and expected effort.

What struck me when I read it is how much of her “prioritise” list is help centre content. Five of the thirteen content types she recommends prioritising sit naturally in a help centre:

Content type (Aleyda’s framework)	The help centre version	Click resilience	Citation potential	Brand mention potential	Expected effort
Brand and entity pages	“What is [brand]?”, “Who owns [brand]?”, “Is [brand] regulated in [country]?”	Medium	High	High	Low
Official product documentation, specifications and policies	Fees, limits, eligibility, accepted documents, policies, release notes	Medium	High	High	Medium
Transaction and task-completion pages	Account tasks and support resolution: close an account, request a statement, dispute a payment	High	High	Medium	Medium
Live first-party databases and reference hubs	Service status, supported countries, current limits and rates	High	High	Medium	High
Product-specific implementation and troubleshooting guides	Fixes by device, app version, country or error	High	Medium	Medium	Medium

Scores from Aleyda Solis, “Content Prioritization in an AI Search Era: A Framework + Worksheet”, August 2026. All five types also score High on business value and proprietary advantage. Help centre examples are mine.

“Its role is to be the current source of record.”

Aleyda Solis, on official product documentation, specifications and policies

That line is the whole job of a help centre in AI search. Her scores also explain the order I recommend later in this guide: entity facts and official policies combine high citation and brand mention potential with low or medium effort, so they come first.

One caveat keeps this honest. Aleyda’s research also describes AI search as a third-party citation problem with an on-page corroboration base. Reviews, press and forums carry a lot of weight in AI answers, and your help centre won’t win every citation on its own. What it gives you is the official base those third-party sources get checked against.

What happens when you don’t answer

When the official answer is thin or hard to find, search engines and AI assistants fill the gap with whatever else they can read.

Search Google for how Paystack settles payments in Kenya and look at what comes back. The first organic result is not Paystack. It is a third-party developer tutorial from McTaba Labs, ranking above Paystack’s own support article on automated payouts. The AI Overview above both results cites McTaba Labs for three of its points, including where settlements land and how fees are deducted.

The details don’t agree, either. The AI Overview, citing Paystack, says settlement takes two working days. The McTaba snippet just below says one to two business days. A merchant reading that page gets two answers to one question, and only one of them comes from the company that holds the money. While researching this post I also found a McTaba tutorial on Kenyan settlements with the writer’s working notes still in it: a line asking someone to verify the default settlement timeline on the Paystack dashboard.

The healthy version looks like this. The business banking app Novo has a help article answering “How long does a transfer take with Wise?” for its own customers, and it ends by sending readers to Wise’s article on transfer times. Another brand answers a question about Wise, then points to Wise’s official page. That only works because Wise wrote the page.

The mistake that gives your citations away: a vendor’s domain

This is the rule I see broken most often, and it matters most.

Your help centre should live on your own website: help.yourbrand.com, support.yourbrand.com or yourbrand.com/help. It should not stay on the default address your help desk gives you on day one, such as intercom.help/yourbrand or yourbrand.zendesk.com.

Look at three companies on Intercom’s default domain today:

• Caresmartz, a home care software company, has 587 articles across 20 collections on intercom.help/caresmartz, while its product site sits at caresmartz360.com. Years of documentation are building up on someone else’s domain.
• Bitnob, a bitcoin and payments app, keeps 78 articles on intercom.help, including its transaction fees article and its withdrawals and deposits FAQ. Those are exactly the facts people ask AI assistants about.
• Alia, a website pop-up tool, explains its free trial, plan upgrades and user permissions on intercom.help. These are the pages a buyer reads just before paying. (Its footer LinkedIn link also has the LinkedIn address typed twice, which tells you how often anyone opens that page.)

Here is what that costs:

• Citations show the vendor’s domain. When an AI answer cites your fees article, the link it shows reads intercom.help. Any authority those links and citations build sits on a domain you will never control.
• Your answers look less official. Search engines and language models connect facts to a brand partly through the domain. The same answer reads as more authoritative on yourbrand.com than on a shared vendor domain.
• You can’t measure it. A subdomain you own sits inside a Google Search Console domain property. A vendor’s domain can’t be verified, so you never see the impressions, clicks and queries your help centre earns. The report that showed me the 21% above only exists because that help centre was on the company’s own subdomain.
• You lose technical control. Canonicals, hreflang, sitemaps, robots rules and schema range from limited to impossible on many default vendor domains.
• Switching vendors breaks everything. Move to another help desk and every URL on the old vendor’s domain dies with your account. On your own subdomain, you redirect.

The fix is small. Intercom, Zendesk, Freshdesk and Help Scout all support custom domains, and setup is usually a CNAME record plus an SSL certificate. In Intercom, you enter your custom domain in the Help Center settings and pick a security protocol. Intercom’s own guide says old intercom.help links then redirect to your new domain, so you lose nothing by moving.

After the switch, check that articles are indexable (no stray noindex), the help centre sitemap is submitted, and the subdomain is covered by a domain-level Search Console property. My technical SEO page covers the wider checklist.

How to build a help centre AI will cite, from customer complaints

Once the help centre is on your domain, the next question is what to write. Your customers have already told you. They told you in complaints. These nine steps turn those complaints into a plan.

Step 1: Audit what is already live

Start with an inventory before you plan anything new. Export every live article with its title, URL, collection, language versions and last-updated date.

Then run three checks:

1. Duplicates. Two articles answering the same question split authority and confuse AI. Merge them into one canonical article.
2. Stale facts. Fees, limits and timelines change. An outdated fee in your help centre gets repeated by AI as fact.
3. Thin clusters. Count articles per collection. In one audit I ran, a whole business product line had no help content at all. That is the biggest gap you can find.

Article counts per collection tell you a lot in seconds. On Caresmartz’s help centre, release notes (99 articles) outnumber getting-started articles (51) almost two to one. Release notes describe what the company shipped. Getting-started articles describe what the customer needs to do. Your plan should lean towards the second.

Step 2: Mine complaints from six sources

Collect questions and frustrations from every place your customers talk.

1. Trustpilot and review sites. Filter for one to three-star reviews from the last 12 months. Copy the specific problem. “Deposit returned after five days with no explanation” is a content idea. If your team keeps pasting the same explanation into review replies, that explanation should be a help article.

Review sites matter for a second reason: AI assistants read them too. How a review site rates your product, and what reviewers say about it, feeds straight into what AI answers say about you. So treat your review profile as part of the help centre plan, and make your lifecycle and support teams responsible for it. Ask happy customers for reviews at the right moment, reply to the unhappy ones with a link to the article that solves their problem, and don’t let a handful of bad reviews drown out the good ones.

Paystack shows what happens when nobody owns this. On Trustpilot it has a TrustScore of 1.5 from 250 reviews. The page notes that Paystack hasn’t invited customers to review recently and has replied to 23% of negative reviews, typically within a month. Trustpilot’s own AI-written summary of recent reviews leads with delayed payouts and funds held during account reviews. That summary is exactly the kind of text an AI assistant can pick up when someone asks whether Paystack is reliable.

2. App store reviews. Apple App Store and Google Play reviews surface login, verification and card problems that rarely reach your inbox. Sort by most recent and most helpful. The same rule applies here: your rating is public evidence, so keep it healthy.

3. Your support tool. This is the richest source. In Intercom, Zendesk or similar, export conversation tags and topics for the last quarter. Look for high-volume topics with no matching article. Then ask your support team one question: “What do you answer by copy and paste every day?”

4. Social listening. Search X, Facebook groups, LinkedIn and TikTok comments for your brand name plus words like “why”, “how”, “can’t”, “stuck” and “reversed”. Listening tools speed this up, but a manual search for a week works too.

5. Forums and communities. Reddit, Quora and local forums (Nairaland, if you serve Nigeria) are where people ask the questions they think are too basic or too awkward for support. They are also where AI assistants often find their answers.

6. Comparison articles and press. Read the “Brand A vs Brand B” articles that mention you. Note the objections and the facts they get wrong. Every wrong fact needs an official article that corrects it.

Step 3: Cluster complaints into themes

Put every complaint in one sheet with five columns: source, verbatim quote, product area, theme and frequency. Then group the rows into themes.

For a payments product, the biggest unanswered themes usually look like this:

1. Account restrictions and deposit reversals
2. First-party vs third-party deposits
3. Fee and exchange rate transparency
4. Transfer and deposit timelines
5. Tax and reporting questions
6. What the card works for, and why it gets declined
7. Business account questions
8. Fraud and account recovery

Your themes will differ by industry. A home care software company will see scheduling, payroll and compliance. An e-commerce tool will see billing, integrations and cancellations. The method stays the same.

Step 4: Turn themes into titles in the customer’s words

Write each title as the question a customer would type, or ask an AI assistant. Use their words, not your internal product names.

• Weak: “Deposit Reconciliation Policy”
• Strong: “Why was my deposit reversed or returned to the sender?”

Wise’s “How long will my transfer take?” works because a worried customer would type almost exactly that. So does “How do I get my referral code?”, the help article that came second in AI impressions in the data above.

Then check every proposed title against your live articles and your blog. A plan that duplicates existing pages creates cannibalisation, not coverage.

Step 5: Tier the plan by citation value

Not every article is worth the same. I rank help centre articles in four tiers and write them in that order. The order follows Aleyda’s scores: the content types with the highest citation and brand mention potential for the least effort go first.

Tier	What it covers	Why it matters	Example
1. Specs, policies and entity facts	Fees, limits, eligibility, ownership, regulation, accepted documents	Highest citation and brand mention value. AI repeats these as fact	“Who owns [brand]? Founders and company structure”
2. Reference hubs and task completion	Supported countries, fee tables, “how do I” guides	High citation value, and AI agents use them to complete tasks	“Supported countries and the services live in each”
3. Troubleshooting	Errors, delays, declines, restrictions	Deflects tickets. Medium citation value, so keep each one focused	“Why is my card declined online?”
4. High-demand how-tos	Account, login, card and payment fixes people search for	Wins organic clicks and grows help centre impressions	“How to reset your [brand] password”

Write Tier 1 first. These are the articles that decide what AI says about your brand. Troubleshooting matters, but it is not where your citations come from.

Keep about 20% of the plan empty and hand it to your support team. They fill it from live signals: the top unresolved tickets this month and whatever product launched last week.

Step 6: Brief each article for people and machines

Every help article brief should ask for:

• An answer-first opening. The first sentence answers the title question. Context comes after.
• Specific numbers. Exact fees, limits and timelines, with units. Vague answers don’t get cited.
• Tables for reference facts. Fees by country or limits by account type belong in a table, not a paragraph.
• One canonical article per question. No near-duplicates.
• A visible last-updated date. It signals freshness to readers and crawlers.
• Links to related articles. Every article links to its nearest neighbours, so readers and crawlers find the whole cluster.

If you brief blog content the same way, my content strategy page shows how I run briefs across a whole site.

Step 7: Create collections for missing clusters

If a product area has no home in your help centre, give it one. A clean collection such as “Business accounts” or “Security and compliance” groups related answers under one clear label, which helps readers and helps search engines and AI tools tie that topic to your brand.

Bitnob’s help centre shows the shape: separate collections for sending and receiving money, virtual cards, account security and compliance, and Bitnob Business. A reader can see at a glance what the product does.

Step 8: Localise your most important articles

If you serve more than one market, translate your Tier 1 and Tier 2 articles first. A few technical rules:

• Use your tool’s built-in multilingual articles, so one article holds every language version and authority stays on one canonical page.
• Show a language switcher and add hreflang on every localised article.
• Set the correct lang attribute for each version.
• Translate titles and meta descriptions, not only the body. Those are what get cited and shown in results.
• Test right-to-left layouts properly for Arabic and Hebrew.

For the full approach to multi-market sites, see my notes on international SEO.

Step 9: Measure what changes

Take a baseline before you publish, then track these every month:

1. Impressions for your help centre pages in Search Console’s generative AI features report, compared with the whole site
2. Mentions and citations of your brand in ChatGPT, Perplexity, Gemini and Claude
3. The share of brand questions where your own help article is the cited source
4. Your rating and review volume on the review sites customers use
5. Support ticket volume for each theme you covered

The third number is the one I watch most.

Start with the complaints you already have

You don’t need a research budget to fix this. You need last quarter’s tickets, a week of review reading and a custom domain setting you can change this afternoon.

Your help centre is the one place where your company states its own facts, so put it on your own domain and fill it with the answers your customers keep asking for.

If you want a second pair of eyes on your help centre or your AI visibility, get in touch.