Merge Duplicate Feature Requests (AI Suggest, HITL Approve)

How do you merge duplicate feature requests without losing votes or trust? Let AI suggest near-duplicates, then require a human to approve so vote totals combine into one honest canonical post. Autopilot-only merges feel fast until they glue unrelated ideas. Feedjolt’s model is HITL: you set a confidence bar — above it merges can proceed, below it waits for you — and you can split a bad merge with a keystroke. Start a trial at https://www.feedjolt.com/en/register.

Duplicates are the quiet killer of feedback boards. Ten posts about “SSO” look like ten weak signals instead of one strong theme. Sales links the wrong URL. Eng builds the wrong slice. Changelog notify only reaches voters on the post you closed. This guide covers why duplicates poison prioritization, suggest vs auto-merge failure modes, a human checklist before merge, Feedjolt’s HITL and MCP path, what happens to voters and the roadmap after a merge, governance, metrics, and an FAQ.

Why duplicates poison prioritization

Every duplicate splits attention across the system:

• Vote power dilutes across near-identical titles, so “top voted” lies
• Roadmap columns show noise instead of themes leadership can debate
• Support answers the same question on five threads
• Ship emails miss sibling voters who filed the same pain in different words
• Agents summarizing “top themes” double-count or under-count depending on luck

Boards without a merge policy become museums of slightly different phrasings. The fix is not banning new posts. The fix is a disciplined dedupe loop that protects vote honesty while still welcoming intake from Slack, email, widgets, and API.

If you are still choosing tooling, pair this playbook with how to choose a feedback board and the philosophy in human-in-the-loop AI feedback triage.

Suggest vs auto-merge (and the failure modes)

AI duplicate detection usually compares embeddings or semantic similarity as a user types or after intake. Useful. The product decision is what happens next.

Mode	Behavior	Helps when	Fails when
Suggest only	UI/agent proposes candidates	High-stakes themes, early boards	Humans ignore the queue; duplicates pile up
HITL thresholds	High confidence may merge; low waits	You need scale with guardrails	Bar set too low → wrong merges
Autopilot-only	Model merges silently	Almost never for customer-facing boards	Wrong SSO glued to password reset; trust collapse

Prefer HITL. Feedjolt’s homepage story matches that pattern: confidence bar you control; wait below the bar; split merges apart. That is the opposite of “AI closed tickets while you slept.”

Classic merge failure modes

• Theme collision — “Okta SSO” merged into “Google login” because both say auth
• Bug vs feature — crash reports merged into a wishlist item with different SLAs
• Account-specific vs product-wide — one customer’s workflow glued into a public theme
• Lost quotes — merge keeps titles but drops the best customer language
• Orphan voters — notify still points at a discarded post ID
• Status demotion — an In-progress survivor accidentally reset to Under review

Human checklist before you approve a merge

1. Read both titles and the top comments — not just the similarity score
2. Ask: would a customer who wrote post B feel heard under post A’s title?
3. Confirm same job-to-be-done (outcome), not just shared keywords
4. Prefer the clearer customer-facing title as the survivor
5. Keep the richest quotes; archive redundant fluff
6. Check tags and statuses so you do not demote active work
7. If ARR or renewal risk is attached, verify accounts still map to the survivor
8. When unsure, leave them separate and cross-link with a comment

If you cannot explain the merge in one sentence to a customer, do not merge.

Feedjolt HITL and MCP path

Feedjolt combines semantic dedupe with a human confidence bar and keyboard-first triage. MCP extends triage into Claude or Cursor: ask for unmerged auth clusters, review candidates, then apply only merges you approve. Safer prompt pattern:

• “Cluster last 7 days tagged auth; list duplicate candidates with confidence — do not apply.”
• “Which pairs share the same job vs only keywords?”
• “Flag candidates that mix bug severity with feature wishlist.”

Only after proposals look sane should you enable write tools. Deep dive: MCP feedback triage for Claude/Cursor. Plan note: REST & MCP API sit on the premium set (Startup by application and Scale — not Growth). Re-verify the homepage matrix before you assume Growth includes MCP.

Feedjolt homepage facts (Sep 2026, re-verify before buy): flat per-workspace pricing — Startup $9/mo by application, Growth $15/mo billed annually, Scale $39/mo billed annually; unlimited boards, posts, and contributors/voters; premium set (Slack, webhooks, REST & MCP API, SSO, audit logs, advanced AI, etc.) on Startup and Scale, not Growth; Linear two-way sync with voter email on ship; public roadmap + changelog; semantic merge with confidence bar; MCP for Claude/Cursor triage; 14-day trial at https://www.feedjolt.com/en/register (no card).

After merge: voters, roadmap, and honesty

A good merge should combine votes into one coherent count, preserve voter identity so ship notify still reaches everyone who cared, update the public roadmap so one card represents the theme, and leave an audit trail for disputes. Then keep the loop moving: status → Linear when ready → ship → changelog + voter email. Deduping without close-the-loop just makes a cleaner ignored backlog.

When posts were public, leave a breadcrumb: “Merged into [canonical] — votes combined.” Voters who bookmarked the old URL should land on the survivor. A 404 teaches customers that engagement is risky.

Weekly dedupe ritual (20–30 minutes)

1. Sort new posts; fold obvious typos into survivors
2. Run in-app or MCP duplicate review on the top three themes
3. Approve high-confidence merges; park ambiguous pairs
4. Reply on anything older than 14 days still stuck in “new”
5. Check that Planned duplicates are not competing on the roadmap
6. Spot-check that ship notify would hit the combined voter set

Governance: who may merge

Limit merge rights to PMs and a named triage owner. Support can flag candidates. Agents draft. Eng should not invent merges from Linear titles alone. Write the rule in the Slack topic for your intake channel so folklore does not appear under deadline pressure. Founders reopening merged themes as vanity demo posts is an anti-pattern — link the canonical post instead.

Playbook examples (anonymized)

Good merge: “SSO with Okta,” “Need SAML before security review,” and “Okta login for admins” share one job. Survivor: “SAML/Okta SSO for workspace admins.” Votes combine; best security-review quote stays pinned.

Bad merge: “Dark mode” into “Theme customization API.” Same visual keywords; different buyers and scope. Keep separate; cross-link.

Hold for human: “CSV export broken” vs “Scheduled CSV exports.” Bug vs feature. Different SLAs. Never let similarity override severity.

Metrics for dedupe health

• Duplicate rate on top 10 themes (should fall)
• Percent of merges reversed within 7 days (should stay low)
• Time-to-first-triage for new posts
• Share of ship emails that notify the combined voter set
• Support tickets that reinvent an existing board post (should fall as search + suggest improve)

If reverse rate climbs, raise the confidence bar and return to suggest-only for a sprint.

Migration note

Imports from Canny, Frill, or CSV dumps usually create a duplicate spike. Do not auto-merge the whole import on day one. Import → human pass on top themes → then enable HITL thresholds. Flat pricing helps you leave the portal open while you clean house — see flat-price feedback boards.

Peers (qualitative, re-verify)

Duplicate detection appears across Canny-class boards, Triagly-class triage tools, SignalBoard-class indie boards, and open tools like Quackback. Compare confidence controls, undo/split, voter preservation, and whether AI can silently rewrite the board. Do not invent scores. Click through the week you buy.

FAQ

Should every near-duplicate be merged?

No. Link related posts when jobs differ. Merge when it is the same customer outcome.

Will AI always be right?

No. That is why HITL exists. Treat scores as hints.

What happens to votes on the discarded post?

They should roll into the survivor. Verify in product — table stakes for honest prioritization.

Can agents merge without me?

You can configure thresholds. Prefer conservative bars at first. Keep customer-visible merges reviewable.

Does Feedjolt meter voters?

Homepage positioning is flat workspace pricing with unlimited contributors/voters — not a tracked-user tax. Re-verify plans.

Start merging without losing trust

Import or recreate your top themes, set a conservative confidence bar, and run one week of HITL merges. Start the 14-day Feedjolt trial — no card, cancel in one click. Pair with HITL triage and MCP triage so agents propose and humans decide.

Intake hygiene that prevents duplicates upstream

Merging is downstream. Prevention starts at intake. When customers open the “new post” form, show near-matches before they submit. When Slack messages become posts, require a quick search. When sales pastes a deal note, paraphrase into a customer-safe title first. Widgets that suggest “this already exists — vote instead” prevent half the merge queue.

Train CS to search before filing. Put the portal search URL in the support macro. Celebrate “I found the existing post” as a win in standup. Culture beats another model parameter.

Confidence bars in practice

Start conservative. A practical rollout:

1. Week 1: suggest-only; humans merge manually while watching false positives
2. Week 2: enable auto-merge above a high bar (only near-identical titles/bodies)
3. Week 3: lower slightly if reverse rate stays near zero
4. Anytime reverse rate spikes: raise the bar and audit the last twenty merges

Document the bar in your internal runbook so a new PM does not “turn it up” for a demo week.

What good canonical titles look like

Canonical titles should be outcome-shaped and scannable: “SAML/Okta SSO for workspace admins,” not “SSO stuff???” or “as discussed on the call.” Include the actor when it matters (admins vs end users). Avoid vendor lock-in in the title if the job is broader (“SSO via SAML” beats “Okta only” unless Okta is truly the only ask).

After merge, skim the first two comments. If they only make sense under the old title, edit a clarifying note at the top of the survivor.

Connecting dedupe to prioritization math

Once votes combine, re-rank. A theme that looked like three mid-tier posts may jump the queue. That is the point. Feedjolt’s prioritization story also includes ARR and renewal risk signals alongside votes — the loudest request is not always the most valuable. Dedupe makes both signals readable instead of fragmented.

Do not let a newly merged mega-post auto-create a Linear issue. Promote deliberately after a human looks at scope.

Security and privacy during merge review

Agents summarizing duplicates may pull quotes that include emails, account IDs, or deal amounts. Prefer tools and prompts that keep PII out of chat logs. Start MCP access read-only. Rotate API keys. Never paste keys into the chat. For Feedjolt, store secrets in the agent’s secret store and confirm plan gates for MCP on Startup/Scale.

When open-source or peer boards are enough

If you must self-host and will staff ops, Quackback/Fider-class tools can work. If you already pay for a support suite with boards, switching only for dedupe is rarely worth it. Choose Feedjolt when you want HITL merge plus roadmap, changelog, Linear notify, and flat workspace pricing in one loop — not as a generic “AI feedback” buzzword buy.

Operational ownership

Name a DRI for triage. Put the portal URL in onboarding. Review Planned monthly. Treat changelog + voter notify as definition of done for user-visible work. Tools amplify ownership; they do not replace it. Re-verify vendor docs the week you implement — product surfaces move.

Trial mindset

Prefer a fourteen-day trial with a real ship over a feature matrix debate. Connect intake, merge one duplicate carefully, promote one item, ship, and confirm a voter heard from you. That loop teaches more than another comparison table. Feedjolt’s trial is at https://www.feedjolt.com/en/register.