TL;DR
Export 50-200 real support tickets, paste them into a long-context model (GPT-5.5 Pro, Claude Opus 4.7, or Gemini 3.1 Pro — all 1M-token as of June 2026), and ask it to cluster by intent, count tickets per cluster, and draft a 2-3 sentence answer per cluster. Keep the human in the loop for two things: contradictory agent replies and any answer the model invents. Ship the top 12-15 entries; route disagreements and unanswered questions to your support lead. A dedicated platform (Zendesk Knowledge Builder, Intercom Fin) automates more but costs per seat or per resolution; a raw chat model costs almost nothing and gives you full editorial control.
The task
You have a stack of support tickets covering the same dozen questions on repeat. Customers ask, agents reply, the answers vary slightly each time, and the help center has not been touched in nine months. The goal is a help center that handles the top 12-15 questions consistently: clustered from real tickets, written in a direct tone, and short enough that a reader can scan and act in under 30 seconds per entry.
Two things have changed the calculus in 2026. First, the big chat models now hold a full ticket export in one window — Claude Opus 4.7 and Sonnet 4.6, Gemini 3.1 Pro, and GPT-5.5 (full 1M context only on the $200 Pro tier) all run 1M-token context. You can paste 150+ tickets and cluster them in a single pass instead of stitching batches together. Second, Google retired FAQ rich results on May 7, 2026, so the old reason to add FAQ schema for SERP dropdowns is gone. The reason to build a clean FAQ now is reader self-service and AI-search citation, not a star-rating snippet.
When AI helps — and when it does not
AI is excellent at clustering tickets by intent, deduping near-identical questions (embedding-based tools flag duplicates at cosine similarity above ~0.85), and drafting answers in a non-marketing voice. It is poor at deciding which answer is correct when ticket replies disagree — agents sometimes contradict each other on edge cases, refund windows, or plan limits. Surface those disagreements explicitly; never let the model pick a “best” answer silently. It will pick the most fluent one, not the correct one.
Pick your path: chat model vs. support platform
You have two realistic routes as of June 2026. The DIY chat-model route gives you control and near-zero cost. A support platform automates ingestion and deflection but charges ongoing.
| Route | What you get | Cost (June 2026) | Best for |
|---|---|---|---|
| Chat model (paste tickets) | One-pass clustering + drafts, full editorial control | ChatGPT Plus $20/mo, Claude Pro $20/mo, Gemini in Google AI Pro $19.99/mo; API ~$3-30 per 1M tokens | Small/mid teams, one-time or quarterly rebuilds |
| Zendesk Knowledge Builder | Auto-drafts up to 40 articles from the last 30 days of tickets | Add-on to Zendesk Suite seats (per-agent) | Teams already on Zendesk |
| Intercom Fin (live agent) | Answers customers directly, not just drafts | $0.99 per resolution, ~67% avg resolution rate; $49.50/mo minimum (50 resolutions) standalone | Teams wanting live deflection, not just an FAQ |
For most teams writing a static FAQ once a quarter, the chat-model route wins on cost and control. Reach for a platform when you want the FAQ to answer live tickets, not just sit on a page.
What to feed the AI
- 50-200 past tickets (both the customer message and the agent reply — the reply is where the real answer lives)
- Audience (consumer / business / developer)
- Brand voice in a sentence (“direct, friendly, no exclamation marks”)
- Known disagreements or edge-case answers from your team
- Existing help docs to link out to (so it deep-links instead of duplicating)
- Compliance constraints (regulated language to quote verbatim)
Strip personally identifiable information (names, emails, order numbers) before pasting tickets into any third-party model. A quick find-and-replace to [CUSTOMER] / [ORDER_ID] keeps you out of a privacy problem.
Copy-ready prompt
Below are [N] support tickets. Build a help center FAQ.
Audience: [consumer / business / developer]
Voice: [one sentence]
Existing docs to link out to: [list]
Known disagreements in past replies: [list]
Compliance constraints: [list]
Tickets:
"""
[paste — customer message + agent reply per ticket]
"""
Return:
1. Top 12-15 question clusters with ticket count per cluster, sorted by count
2. For each cluster: canonical question (as a user would ask), a 2-3
sentence direct answer, suggested link to deeper docs, ticket IDs in
the cluster
3. A "disagreement watch" — clusters where past agent replies contradicted
each other
4. A "missing answer" list — questions in the tickets with no consistent
answer yet
5. A short tone audit — flag any draft that drifted into marketing voice
No "Great question!" No emoji bullets. Do not invent product behaviour to
fill gaps; if an answer is not in the tickets, list it under "missing answer".For developer audiences, append: Add a 1-line code snippet per answer where applicable; mark which need API verification.
If your export is too large for one window (rare at 1M tokens, but possible with verbose tickets), cluster in batches of ~50, then run a second pass that merges the cluster lists and re-counts.
Recommended output structure
A list of 12-15 FAQ entries, each with question / answer / deep link / ticket count. Then three control artifacts: a “disagreement watch” callout, a “missing answer” list, and a tone audit. The FAQ entries go live; the disagreements and missing-answer items go to your support lead as a backlog, not to the page.
How to check the output is usable
- Every entry traces to real ticket IDs (no invented FAQs)
- Answers are 2-3 sentences, not paragraphs
- No marketing language slipped through
- Disagreement entries are flagged for human resolution, not silently resolved
- Each answer suggests one deeper link, not three
- Ticket counts add up to roughly your sample size (a cluster of “1” is not “common”)
Common mistakes
- Marketing-language answers (“We’re so glad you asked!”). Kill on sight.
- Fake FAQs that do not reflect real tickets. Readers can tell, and so can support agents who get the same question the next day.
- No link to deeper docs. The FAQ becomes the documentation, then goes stale.
- Letting AI pick among contradictory agent replies. Surface the contradiction instead.
- Skipping the tone audit. Small drifts compound across 15 entries.
- One-off questions promoted to “common.” Anchor every entry in ticket count.
- Leaving customer PII in the pasted tickets. Redact before you paste.
A note on FAQ schema in 2026
You can still add FAQPage JSON-LD markup — it is a valid Schema.org type and AI search engines still parse it. But Google removed the FAQ rich result (the expandable dropdowns under your listing) for all sites on May 7, 2026, and is winding down the related Search Console reporting through August 2026. So do not build the FAQ for a SERP snippet that no longer exists. Build it for humans scanning your help center and for the AI assistants that cite clean, well-structured answers.
FAQ
- How many tickets do I actually need? Aim for 50 at minimum so clusters are statistically meaningful; 150-200 gives stable counts. Below ~30 you cannot tell a real pattern from noise.
- How often should I rebuild this? Quarterly, or after any major release. Tickets pile up faster than you expect, and a release shifts which questions dominate.
- Should AI write the deeper docs too? Different job. The FAQ is short and scannable; deeper docs need structured technical-writing work and a real review cycle.
- What about multilingual FAQs? Cluster once in your primary language, finalize the canonical answer, then translate that. Translating raw agent replies one by one re-introduces the inconsistency you just removed.
- Is the FAQ still worth it now that Google dropped FAQ rich results? Yes — the rich-result snippet is gone, but the page still deflects tickets, and AI search tools (and your own on-site search) reward clean, direct answers. The schema markup remains valid; it just no longer changes how Google displays the listing.
Related
- FAQ writing prompts — alternative FAQ prompt variants
- Customer service reply — agent reply style
- User feedback clustering AI — same clustering pattern for feedback
- App review reply AI — same tone discipline for app store reviews
- Bug report AI — when tickets are bug reports
- AI knowledge base cleanup tutorial — broader KB cleanup