Guru

What it is

Guru is a knowledge platform built on human-verified cards: short, owned, expiry-dated answers with a named subject-matter expert attached, surfaced where the work already happens instead of in a wiki someone has to remember to open. Answers reach people through the Chrome, Edge and Opera browser extension, Slack, Microsoft Teams, the web app, the API, and Guru’s own remote MCP server.

The 2026 product is agent-shaped. Knowledge Agents connect Guru Cards alongside Google Drive, SharePoint, Confluence, Notion and Slack, retrieve fragments across those sources, and synthesize a cited answer with role-based permissions enforced against the person asking. Every interaction lands in the AI Agent Center, which logs the question, the sources cited, and the gap when no answer existed.

The pole Guru occupies is curation before retrieval. Glean and its class index everything and rank it. Guru’s premise is that a smaller corpus somebody owns and re-verifies beats a larger one nobody does. Which premise fits you is a question about your content, not your budget.

Verification became the product in January 2026

Verification used to be manual labor: a card owner got a reminder, opened the card, clicked Verify. The January 2026 release handed that to agents and pushed it past Guru’s own content — agents now verify or unverify material across all connected sources using usage signals, engagement patterns and AI analysis, with human override and a stated reason for each decision. Cards can be marked never-expire for evergreen content and remain subject to agent verification anyway.

One detail decides whether you actually get this: auto-verify is on by default only for Knowledge Agents created after 14 January 2026. Agents that predate that date require a manual opt-in. Teams that piloted Guru in 2025 and expanded in 2026 are running two verification regimes without knowing it.

February added live Slack retrieval through Slack’s MCP, so agents read current channel conversation under each user’s own Slack permissions rather than only what was written down. March added date-range, folder and source-attribute filters, plus per-channel agent routing when you @mention Guru. April added Fathom and Vitally as sources, pulling meeting transcripts and customer-success signal onto the same answer surface.

Pricing is quote-only, and the public numbers are stale

Guru removed per-seat pricing from its site. getguru.com/pricing routes every tier to a sales conversation and describes the deal as tailored to your organization’s scale, knowledge complexity and AI maturity, bundling solution engineers and knowledge-architecture design with the license. No free trial and no free tier are advertised as of 5 September 2026. Organizations with 501(c)(3) status can apply to Guru for Good.

Third-party pricing pages still publish $25 per seat per month on annual billing with a 10-seat minimum. That is the pre-2025 self-serve list price and it is not what you will be quoted.

Plan against transaction data instead. Vendr’s February 2026 analysis of 167 Guru purchases puts the median contract at $39,168 per year, with a range of $8,159 to $121,023 and average savings of 18% off first quote. At 150 seats that median works out to about $22 per person per month all-in — near the old list price, which is why the quote-only shift reads as segmentation rather than a price rise. Budget the cycle as well as the license: a sales-led motion with a knowledge-architecture workshop attached is a 4-to-8-week purchase, not a credit-card one.

Best for

A customer-support or enablement leader who owns a body of answers that go stale on a known cadence — refund policy, escalation matrix, security-questionnaire responses — and whose failure mode is a rep confidently repeating last quarter’s policy. The scoped case where Guru is the clear pick: CS macros and support answers surfaced in the browser extension next to Zendesk or Intercom, where the reader never leaves the ticket and the card carries a verification date they can see.

The same shape holds for recruiter FAQs and legal playbooks, which is why this entry is cross-vertical. All three are small, high-consequence corpora with real owners.

Not for

• Long-form documentation. The card format is deliberately short. Runbooks that run for pages fight it, and reviewers report the card model straining as documents get longer.
• Companies with no content owners. Verification assumes a named human accepts the ping. Without that, Guru degrades into a wiki with extra notifications.
• Search across everything. If the answers live in 40 systems and none of it has been curated, indexing is the job and Guru is not an index.
• Teams that want to buy this quarter without a sales cycle. Quote-only means procurement.

Versus the alternatives

• Notion — the volume leader in the knowledge-workspace segment, from $10 per user per month. Pick Notion when the corpus is documents people author and read end to end, and AI is a convenience on top. Pick Guru when the corpus is answers people consume mid-task and staleness is the risk being managed. Alternatives to Notion works through that split.
• Atlassian Rovo — the other incumbent by installed base, included with paid Atlassian Cloud plans and metered in credits. If Confluence already holds the corpus, Rovo costs an order of magnitude less than a separate platform. Pick Guru instead when Confluence is where documents go to die and the fix is ownership, not better search over the graveyard.
• Glean — the fastest-growing entrant in enterprise AI search, around $40-50 per seat with a 100-seat floor. Pick Glean when breadth is the binding constraint and permissions across dozens of systems are already clean. Pick Guru when you would be paying Glean to index content nobody trusts.
• Onyx — the open-source, self-hostable route when data residency or per-seat economics rule out all three.

If none of them fit, best knowledge bases for teams covers the wider field, and RAG explains what retrieval can and cannot recover from a corpus nobody maintains.

Watch-outs

• Notification fatigue is Guru’s oldest and most-reported failure. Verification reminders, card suggestions and update pings stack up, people learn to dismiss them, and the one stale card that mattered gets dismissed with the rest. Guard: set verification cadence per collection instead of accepting defaults — quarterly for policy, annually for evergreen — and route reminders to a single owned team channel rather than to individual DMs. Then read the AI Agent Center’s unanswered-question list monthly; that report, not reminder volume, tells you whether the corpus is working.
• Auto-verify’s default splits your agent fleet by creation date. Agents built before 14 January 2026 sit at manual opt-in and behave differently from ones built after it, with nothing on the surface saying so. Guard: enumerate every Knowledge Agent and check the auto-verify setting on each one before you treat a verification badge as evidence of anything.
• Agents now stamp verification on content Guru does not own. Extending verification to Google Drive, SharePoint and Confluence puts an AI judgment on documents whose authors never opted into Guru’s governance. Guard: scope each agent’s sources to folders with an accountable owner, and keep the human override path staffed through the first quarter. A verification signal you cannot explain to an auditor is worse than no badge at all.
• The legacy /boards API endpoint was sunset in January 2026. Anything built against it — an onboarding sync, a card-provisioning script — breaks and has to move to /folders. Guard: grep your integration code for /boards during renewal planning, not after the errors start.
• Everyone who reads needs a paid seat. Guru is an internal platform, so read-only consumption is not free and headcount is the whole bill. Guard: price the rollout at full eventual headcount in the first quote and negotiate the band there, rather than piloting with 25 seats and meeting the real number at expansion.

Related: the AI support agent stack shows where a verified knowledge layer sits relative to the deflection tools that read from it.