How Wave turns product evidence into shipped improvements

One Amplitude user spent 76 seconds typing a question while three suggestion cards failed to anticipate the task. The replay exposed a mismatch between intended guidance and the actual experience.

At Amplitude, we built Wave to answer hard questions about digital product use by combining session replay, a record of user actions, with product analytics, aggregate behavior data. Teams within Amplitude have used it to save text answers, clarify setup requirements, change how agents investigate requests, and remove documentation-search friction.

Wave’s first investigation followed a bug that deleted sessions when a user clicked during an agent response. It linked replay evidence with customer feedback and produced a coding agent’s pull request, which the team reviewed and merged.

What does Wave change for product teams?

Amplitude’s team sets the rule: “use replays to form a hypothesis and analytics to check how widely the pattern occurs.” Each Opportunity holds the evidence, proposed change, and measurement plan. Teams inspect it, revise the plan, assign implementation to a teammate or coding agent, then return to it after release to see what happened.

How did personalization change suggestions?

The 76-second failure came from an analysis of 100 replays. Users manually typed questions in 80 sessions, while only 12 selected a card. The cards rendered correctly, but their prompts were too generic for the work at hand.

A user’s arrival from a dashboard or return to an existing chart supplied context. The team updated the page to offer personalized cards when that history existed and generic cards otherwise, while preserving the option to ignore them and type a question.

The July experiment produced 2.43 suggestion sends per 100 exposed users for personalized cards, versus 0.94 for generic cards, a 160% increase. Amplitude treats this as an initial readout that still needs validation.

Can one useful answer become repeatable work?

Custom Agents already handled recurring analysis, but users had no clear route from a useful response to an agent. Wave recommended a contextual suggestion into Custom Agent creation when a response represented repeatable work. The team also added chat source attribution to distinguish creation clicks that came from an answer.

Over four weeks, about 70 unique users per week clicked the new entry point, roughly 8.3% of creation clicks. The attribution made a previously ambiguous path measurable.

Users could save charts, but not structured text explanations, even when they did not want automation. Wave proposed options to save an answer to a Notebook or export and copy it; the team shipped both. Saved Notebook content now carries citations and a link to the original response; the closing readout counted 152 save events over 30 days.

Session Value Rate, Amplitude’s metric for judging user experiences, excluded the save event. The team must update its definition before evaluating the behavior. A separate Opportunity extended access to synthesized specialist-agent outputs through Model Context Protocol resources, letting MCP clients retrieve them. Saving, agent creation, and MCP access now reuse agent work in new places.

What did the setup experiments show?

Another Amplitude product, AI Feedback, searches reviews, surveys, support tickets, and other customer channels, then consolidates their suggestions. It needs those connections to produce meaningful results, but its empty state did not guide first connections. Wave identified the gap, recommended clearer guidance, and the team shipped it.

From June 30 through July 29, the measured Opportunity recorded 121 unique users starting setup and 84 completing a source connection, a 69.4% completion rate.

The connector gallery then exposed an administrator-permissions blocker. The team added requirement badges to Intercom and Zendesk when no sources were connected. Completed connections remain the success metric, but Amplitude is still waiting for data that establishes an increase.

Another test labeled CSV and HTTP API recommendations without changing tile order. The HTTP API recommendation rolled out on October 27; its direct readout showed 24.61% developer setup activation versus 17.75% in control. The directional result was statistically inconclusive and remains a hypothesis.

Wave also proposed a guided onboarding checklist that would recommend an action after role selection. The team reviewed the Opportunity; implementation and impact measurement remain unconfirmed.

How did agent behavior change?

The Setup Agent used to scan data broadly before asking what the user wanted. The team reversed that order: establish intent first, then investigate, offer fast and well-researched paths, and produce shorter proposals.

The Web Vitals investigator produced repeated, overlapping performance reports during a run. A tighter prompt contract changed the instructions governing its output to address that duplication. For both changes, the team reviews Wave’s proposal and decides what ships.

What fixes reached documentation and the blog?

Documentation search could fetch results before a user finished typing the first characters of a query, which also recorded premature search events. The fix sets a configurable three-character minimum and roughly 200 milliseconds of debounce, a delay before requests run. Shorter inputs now show a prompt without fetching results or starting a search.

On the Amplitude blog, the team used a post’s topic to tailor its call to action through the existing Sanity inline ad component. The team still needs attribution that separates topic matching from traffic mix and broader form activity.

What stays connected after release?

Wave lets teams review an interface proposal beside a prompt contract or instrumentation fix, with the triggering evidence attached. Teams choose what to pursue and review what ships. Unclear results lead back to the original question and the next measurement.

The saved-answer feature has a clear next task: add its action to the Session Value Rate definition, then evaluate the intended behavior. Investigation, implementation, and result remain in one Opportunity.

Amplitude closes with a direct instruction: “Bring a product area and a goal to Wave, and review the opportunities with your team.” Product teams will increasingly expect proposals to retain their evidence and measurement plan through release, rather than separate analysis from implementation.