Fullstory Fullcapture Gold

Semantic events, derived from data you already export.

Every click Fullstory captures goes into your warehouse as raw Silver — four hundred events per session that say what happened, but not what it meant. Gold mines those patterns and writes named events back onto the session timelines they belong to. No instrumentation. No release. Retroactively.

✦ No instrumentation required ✦ Retroactive coverage ✦ Requires Anywhere Warehouse

TL;DR

Gold closes the semantic gap between raw behavioral capture and what AI agents can actually reason about.

Fullstory has no Gold layer today. Every semantic enrichment in the product requires a human to author it: element names, URL groupings, custom event definitions. That model was built for human analysts. The consumer has changed — Story AI, MCP, and custom agents now receive raw Silver data, and an agent given raw ingredients returns a plausible answer, not a correct one.

Gold changes this by treating the customer's warehouse data as a mine. The pipeline identifies the shortest ordered chain of steps that unambiguously signals a business moment — a booking completed, a login finished, a signup sequence — and writes that named event back onto every session timeline where it occurred. The event carries its own provenance: which pattern matched, how many raw steps it compressed, how frequently that pattern occurs, and a cohesion score. When someone asks why an event fired, there is always an answer.

The first production run against the CarGo demo org processed 298 million Silver events and produced 17,528 Gold events across five named event types. Zero errors. The pipeline completed in approximately eight minutes. The events appeared on session timelines within the hour.

⟁

The Semantic Gap

Why raw Silver data fails AI consumers and how Gold addresses the structural gap

↻

How Gold Works

Seed resolution, n-gram + nPMI mining, curation, validation, and daily scheduling

★

Event Vocabulary

Two event kinds — sequence and cluster — and a reference event dictionary

◉

In Fullstory

What Gold events look like on a real session timeline

✓

Prerequisites & Engagement

Four gates that must clear, and what the engagement actually looks like

The Problem

Silver data is rich. And completely unsemantic.

Fullstory captures every click, navigation, and form change. A session can be four hundred events — each one recording what happened, none of them recording what it meant. This is fine for human analysts who bring their own domain knowledge. It is not fine for AI systems that must infer meaning from the data they receive.

The semantic gap widens as AI consumption grows. Story AI, MCP-based analyst agents, and custom automation all receive raw Silver data from the warehouse. A sequence mining agent sees custom:SubmitClick → nav:/dashboard and must guess whether this is a login, a booking, or an account deletion. Gold removes the guess — it names the moment.

The same pattern shows up in a second place: Warehouse churn. A significant cohort of multiyear contracts include Anywhere Warehouse as a line item but have never activated it. The existing activation motion — outreach to data teams, no promised outcome — reminds customers of an unused line item without giving them a reason to use it. Gold is a concrete, immediately visible reason. The customer connects their warehouse, the pipeline runs, and named events appear on their session timelines within hours.

Phase 0 is complete. The pipeline ran end-to-end on the CarGo demo org, importing 17,528 Gold events against 298M Silver events with zero errors. The pattern is proven. Phase 1 is the first external co-creation customer.

Before Gold

What an AI agent sees

            custom:SubmitClick

            nav:/checkout-success

            page:seen

            custom:elementClick

            nav:/home

            ... 391 more events

The agent must infer from raw event tokens whether this session contains a successful booking. It may be right. It is not certain.

After Gold

What an AI agent sees

              * Booking Flow Completed

              * Login Sequence

              custom:SubmitClick

              nav:/checkout-success

              ... 393 more events

The named Gold events rise to the top. The booking is confirmed. The analysis is grounded in a fact, not a pattern match.

The Mechanism

Discovery, mining, curation, and daily delivery.

The process starts with a human conversation — not a tool. Business stakeholders identify which funnels and moments matter, and that output anchors the mine. The miner then runs deterministically against existing warehouse history, surfacing candidate patterns that the customer curates and approves before anything goes live.

Stage 1

Discovery Interview

Human ↔ human: FIL/SE team interviews customer stakeholders to identify which funnels and moments matter most to the business. The output is a documented list of anchors — the miner uses these to discover patterns with business importance, not just statistical frequency.

›

Stage 2

Mine

n-gram + nPMI sequence detection runs over the full warehouse export history using the discovery anchors. Candidate patterns surface with frequency scores, cohesion scores, and variance across sessions.

›

Stage 3

Curate

Customer workshops confirm or reject each candidate. The customer is the authority on whether a pattern means what the data suggests. This is the long pole — elapsed time tracks workshop scheduling, not engineering capacity.

›

Stage 4

Validate

Real sessions are pulled. Events must land exactly where they should. No Gold event goes live before a human confirms it on a real session timeline.

›

Stage 5

Push Gold

Events write back onto timelines via the Fullstory Server Events API. The dictionary is versioned — each run records which version produced which events. Daily schedule from go-live.

Three anchor modes shape what the mine discovers. An end anchor — a known completion point — is the most powerful: it discovers all the paths that lead to that moment, surfacing variation you didn't know existed. A start anchor discovers everything that follows a known entry. An any anchor is the broadest and the most prone to false positives.

Retroactivity is the headline capability. The pipeline runs over the full history in the warehouse. An event you wish you'd instrumented in January can exist for January — on the sessions from that month, on the timelines of those users, surfaced in any analysis that queries that period.

The Gold Property

Every event carries its own provenance

Each Gold event carries a gold property with the following fields — making every derived event fully auditable:

↦

pattern_id — which pattern in the dictionary matched

↦

steps_compressed — how many raw Silver events the pattern collapsed

↦

frequency — how often this pattern occurs across the session population

↦

cohesion — confidence score from the mining evaluation

↦

run_id — which pipeline run produced this event

Event Vocabulary

Two event kinds. One naming convention.

Gold events are of two distinct types depending on what the feature surface looks like in the data. The detection method is invisible to users and agents — only the name is exposed. All Gold events carry a leading asterisk to bring them to the top of any event list in the Fullstory UI.

Sequence Events

Ordered paths with a known start or end

The feature follows a defined, ordered path. The mine identifies the shortest ordered chain of steps that unambiguously signals completion. The chain is as short as possible — if three steps distinguish this pattern from all others, the event fires on three steps, not ten.

Named as past-tense completion verbs: the name describes what finished, not what happened. * Booking Flow Completed. * Login Sequence. * Account Deletion Sequence.

Cluster Events

Non-linear engagement surfaces without a fixed path

The feature is a surface a user engages with rather than a path they complete. The mine identifies an anchor event — a key navigation or custom event — plus a threshold of active user behaviors within that context window. The user doesn't need to follow steps; they need to engage.

Named as present-tense engagement nouns: the name describes what the user is doing, not what finished. * Car Detail Engaged. * Plan Comparison Active.

Event Name	Kind	What It Signals	Enrichment Properties
* Booking Flow Completed	Sequence	A rental booking went end-to-end through checkout confirmation	car_id, trip_length (pending redesign)
* Login Sequence	Sequence	User opened the login modal and completed authentication	`is_host` — whether user is a car owner or renter
* Host Signup Sequence	Sequence	A car owner signed up to list their vehicle	—
* Account Deletion Sequence	Sequence	User completed account deletion flow	—
* Car Detail Engaged	Cluster	User engaged meaningfully with a vehicle detail surface — non-linear browsing	—

Platform coverage mines separately. Web, iOS, and Android session data are processed as distinct pipelines. A moment instrumented on web may not have an equivalent signal on mobile — the app may emit no corresponding event. Gold events on mobile require separate seed resolution against mobile session data.

In Fullstory

Named business events on a real session timeline.

Once Gold runs, the derived events appear on session timelines in Fullstory exactly like any other event — searchable, segmentable, usable in funnels and metrics. The * prefix brings them to the top of every event list.

app.fullstory.com · Session Replay · Gold events on timeline

Fullstory session replay showing Gold events on the event timeline

Gold events appear at the top of the session event list alongside raw Silver events. Each carries the gold metadata property — pattern ID, steps compressed, frequency, cohesion, and run ID — visible in the event detail panel.

Events are additive and permanent once they reach a timeline. There is no bulk undo, which is why curation runs as a review process before anything goes live. A pattern that looks right in aggregate can be wrong on individual sessions — curation catches this. Once you are confident in a pattern, it is stable. If the UI changes and breaks a pattern, the pipeline detects the change at the next run and surfaces it for review.

Prerequisites & Engagement

Four gates. All must clear before scoping begins.

Gold reads a customer's exported data, not Fullstory's production systems. This means the warehouse connection is not a nice-to-have — it is the foundation. A prospect without the export needs an Anywhere Warehouse conversation before a Gold one.

🗄

Anywhere Warehouse — Connected

Gold reads the customer's exported session data from BigQuery or another warehouse destination. Without a connected export, there is no data to mine. This is the hard first gate.

🔑

Org API Key — Server Event Write

Derived events write back to Fullstory timelines via the Server Events API. An org API key with server event write permission is required before the pipeline can push anything.

🌐

Data Region — NA or EU

The pipeline is validated for North America and EU data regions. Confirm data region before scoping — cross-region architectures require additional review.

📊

Sufficient History and Volume

A pattern appearing in 1.4% of sessions is workable. A pattern appearing eleven times is noise. Enough history must exist for the mine to produce stable, recurring sequences rather than one-off matches.

Shape of an Engagement

Confirm prerequisites — warehouse export confirmed connected and accessible. API key scoped. Data region verified.

Identify the event dictionary — work with the customer to name 5–10 moments they wish they could search or segment on. These become the seeds for the mine.

Discovery interview — FIL/SE team interviews business stakeholders to identify which funnels and moments matter most. This produces the anchor list that guides the mine — without it, the miner has no business context to orient around.

Mine and curate — automated mining runs, candidate patterns surface, curation workshops confirm or reject them. Get the right people into the workshops early — this step tracks on their availability, not ours.

Validate and go live — events are confirmed on real session timelines, then pushed to the daily schedule. Adding a ninth event later is a configuration change, not a new engagement.

Talk to your CSE or Account Executive about how Fullstory Professional Services can help evaluate and operationalize these capabilities for your data environment.

Behavioral Clustering

User archetypes, derived from behavioral sequences at scale.

Standard analytics collapses a session into counts. Behavioral clustering reads the sequence — what the user tried, what blocked them, how they recovered — and groups users by the shape of their behavior across an entire 30-day window. The output is a named archetype for every user, ready to enrich a CDP, drive a personalization system, or train a fraud detection model.

✦ No cloud cluster required ✦ ~$2 per 100K-user run ✦ Requires Anywhere Warehouse

TL;DR

Clustering makes the structural diversity inside a user population visible — and actionable.

Every analytics team knows that "users who abandoned checkout" is not one group. Some abandoned because the store was closed. Some abandoned because the promo code failed. Some got distracted and came back ten minutes later. Some are testing the flow. Aggregate metrics treat all of them the same. Behavioral clustering does not.

The pipeline pulls from the customer's existing Anywhere Warehouse sync — no separate export required. A backwards-looking engagement window (typically 60 days, with a parallel 30-day run for comparison) is agreed upfront. The pipeline assembles each user's full behavioral journey into a semantic representation, encodes it using a sentence embedding model, reduces to three dimensions with UMAP, and clusters by density using HDBSCAN. A token-efficient summary LLM reads the event timelines of the five most-central users in each cluster and generates an archetype name and one-sentence behavioral description.

Each user in the segment exits the pipeline with a cluster ID, an archetype name, and a behavioral narrative that any downstream system can read and act on. The first production runs validated the approach at scale across millions of events and thousands of users, producing over 100 named behavioral archetypes per run.

⋮

The Problem with Aggregates

Why count-based analytics hides the structural diversity that drives different outcomes

↻

The Pipeline

Warehouse sync → window definition → semantic layer → embeddings → UMAP → HDBSCAN → LLM labels

◎

The Cluster Space

Interactive 3D UMAP visualization of 103 behavioral archetypes from a real production run

◈

CDP Enrichment Output

Enriched user card — what the cluster data looks like attached to a real user profile

→

Use Cases

CDP segmentation, AI prompt enrichment, fraud detection, friction diagnosis

The Problem

Context collapses into counts. And counts lie.

Fullstory captures behavioral sequences. Most downstream systems receive summaries. The individual user's journey — what they tried, what blocked them, what they did next — is reduced to a page view count, a session duration, or a binary conversion flag. At scale, this makes every user look the same.

The mismatch between archetype and reality is the most expensive failure mode. Consider a user whose behavioral cluster says "occasional reorderer" — moderate sessions, moderate spend. In reality they placed four checkout attempts in one session, were blocked by a store availability error, never received a recovery intervention, and their $84 order was lost. The cluster membership was correct in the aggregate. The operational failure was invisible until behavioral clustering surfaced it.

The root problem is temporal: behavioral analytics tools are optimized for snapshots, not sequences. They answer "how many users converted" but not "what did the users who didn't convert actually do differently." Clustering answers the second question at population scale — it groups users by their behavioral sequence across an entire engagement window, not by the last event in the funnel.

A second problem is label quality. Most user segmentation relies on demographic data, device type, acquisition channel, or manually authored cohort rules. These labels describe who users are on paper, not how they behave in the product. A high-value user by LTV who is in the middle of a friction loop that will end in churn looks identical to a high-value user who is happily converting. Behavioral clustering sees the difference.

HDBSCAN does not force cluster assignment. Users who don't belong to a stable behavioral group are labeled as noise — a feature, not a failure. A 15–20% noise rate is the target. It keeps archetypes behaviorally coherent rather than forcing every user into a bucket.

Two Pipeline Variants

User-level journey clustering

Groups users by their full behavioral path across an agreed engagement window (typically 30–60 days). Output: per-user cluster membership and archetype label. Primary use case is CDP enrichment, personalization, and cohort-level analysis.

Click-context clustering

Embeds every click with ±8 events of surrounding behavioral context, then clusters by that context. Output: a friction taxonomy — error loops, abandon spirals, dead click zones — resolved to specific UI elements. Primary use case is friction diagnosis and product prioritization.

The Pipeline

Six stages from raw export to named archetypes.

The enterprise pipeline reads directly from the customer's existing Anywhere Warehouse sync — no separate export step. Open-source tooling (HuggingFace sentence-transformers, UMAP, HDBSCAN) handles embedding, reduction, and clustering. A token-efficient summary LLM generates the archetype names and descriptions.

Window Definition — The customer's Anywhere Warehouse sync is already running, streaming all Fullstory session data directly into their data warehouse. A 7-day test window is used first for semantic layer calibration. The primary clustering window is then agreed — typically 60 days looking back from today. Running the pipeline against both a 60-day and a 30-day window is standard practice: comparing the two reveals how stable the archetypes are over time and surfaces any behavioral drift.

Prototyping note: For standalone evaluation without a live warehouse sync, the pipeline can run from a raw NDJSON export loaded into DuckDB locally — useful for validating the methodology on a subset before committing to full operationalization.

Semantic Layer — Named elements and URL patterns map to business-meaningful tokens. Conversion-critical events are weighted 1.5×; high-intent interactions 1.3×; structural navigation noise 0.8×. This step separates signal from noise — the quality of the resulting archetypes tracks directly to the quality of this mapping. It is reviewed and adjustable before the embedding stage.

Journey Fingerprinting — Each user's history across the agreed engagement window is assembled into a journey string: a sequential, weighted representation of their full behavioral path across all sessions. Users who visit the same pages in the same order get similar strings. Users who explore differently get different strings. The string is the unit that enters the embedding model.

Embed → UMAP → HDBSCAN — all-MiniLM-L6-v2 encodes each journey string into a 384-dimensional vector (~23M parameters, runs locally, no API cost). UMAP reduces to 3 dimensions using n_neighbors=15, min_dist=0.1, cosine metric. HDBSCAN clusters the 3D coordinates with min_cluster_size=30, producing density-based groups without a preset count.

Token-Efficient Summary LLM Labels — The 5 most-central users in each cluster are sampled. Their event timelines — with semantic layer applied — are passed to a token-efficient summary LLM via a structured prompt. The model returns an archetype name and one-sentence behavioral description per cluster. No expensive frontier model is required; the task is summarization, not reasoning.

Tuning Targets

Parameters that shape the output

→

Noise rate: 15–20% — Users not assigned to any cluster. Target this range: below 15% means clusters are too broad; above 25% means min_cluster_size is too aggressive.

→

min_cluster_size: 30 — Produces ~100 clusters for ~100K users. Lower produces more clusters with less behavioral coherence; higher produces fewer, broader archetypes.

→

UMAP n_neighbors: 15 — Controls local vs. global structure in the 3D projection. Higher values preserve more global structure; lower values emphasize local cluster geometry.

→

Engagement window: 30 days — Standard for capturing meaningful behavioral variation. Shorter windows miss multi-session patterns; longer windows introduce noise from lifecycle changes.

The Cluster Space

103 behavioral archetypes in three-dimensional space.

UMAP preserves the local and global structure of the embedding space in 3D. Users who behave similarly cluster together. Users on opposite ends of the behavioral spectrum are far apart. The 3D visualization below is from a production run — each point is a user, each color is a cluster, each cluster carries an LLM-generated archetype name.

Interactive UMAP 3D Cluster Map Production run · 100K+ users · 103 named archetypes · drag to rotate

Cluster U13

Habitual Direct Buyers

High-frequency purchasers with minimal exploration. Navigate directly to saved orders or recent items. Conversion rate near 100% when they reach checkout. Segment priority: retention — they're already the behavior you want at scale.

Cluster U15

Order-Tracking Habitual Buyers

Frequent buyers who engage heavily with post-order tracking. Loyalty-program-aware. High rewards balance engagement. Sensitive to store availability issues — any disruption cascades immediately into tracking behavior. Monitor for operational failures in this cluster.

Cluster U06

Promo-Driven Occasional Buyers

Enter via offer pages or promo links. High abandonment when a code fails to apply. The failed redemption is the conversion killer for this archetype — not the price point. Rechargeable with targeted offers that are confirmed to work before they're shown.

CDP Enrichment Output

Three fields per user. Immediately actionable downstream.

The pipeline exports three enrichment fields per user in the segment. These are designed to plug directly into a CDP, a data warehouse, or an AI system prompt — no transformation required. The schema is consistent across customer deployments.

The behavioral narrative is the most powerful field for AI consumption. fs_cluster_narrative is a one-sentence plain-language description of the archetype generated from real session data. When injected into an AI system prompt at session start — before the first user interaction — the AI has behavioral context it could not derive from conversation history alone.

Enrichment Schema

fs_cluster_id

Stable numeric identifier (e.g., U13). Join key for cohort analysis, A/B targeting, and funnel comparison.

fs_behavioral_archetype

Human-readable archetype name generated by a token-efficient summary LLM (e.g., Habitual Direct Buyers). Ready for AI system prompts and personalization rules.

fs_cluster_narrative

One-sentence behavioral description. Contextualizes the cluster for any downstream agent or analyst without requiring access to session data.

Kaley Lower

kaleydanielle_23@yahoo.com

● U15 · Returning High-Value Buyers

Enrichment Signals

Desktop-dominant 43 sessions / 30d 10 orders / 30d $71.76 avg AOV Repeat purchaser 35% time on order status High-value customer Repeat buyer

Behavioral Breakdown (30d)

Home

26%

Order Status

36%

Other

38%

CDP Fields

fs_cluster_id

U15

fs_behavioral_archetype

Returning High-Value Buyers

fs_cluster_narrative

Frequent high-value buyer who re-engages post-purchase; sensitive to fulfillment friction and order delays.

Applications

What you do with archetype membership.

Behavioral archetype data is most valuable when it flows into a system that can act on it — not when it sits in a dashboard. These are the four highest-value downstream applications, ranked by how quickly a team can move from cluster output to business impact.

CDP Segmentation. Push fs_cluster_id and fs_behavioral_archetype to your CDP and build audiences from behavioral reality rather than demographic proxies or last-touch signals. Route users to retention tracks, growth tracks, or recovery interventions based on who they actually are.

AI System Prompt Enrichment. Inject fs_behavioral_archetype and fs_cluster_narrative into the system prompt of any AI that interacts with the user — a support chatbot, a personalization engine, a voice ordering system. The AI knows the user's behavioral identity before the first message arrives. A user whose archetype is "deal-seeking occasional buyer" gets a different opening than a user whose archetype is "habitual direct buyer who never uses promos."

Fraud Detection. Known fraud sessions have behavioral fingerprints — specific sequences of actions that appear before a high-value claim, promotional abuse event, or account takeover. Train a lookalike model on the behavioral embeddings of confirmed fraud sessions. Score new session embeddings at ingestion time — every few seconds in a Snowflake ML pipeline — and route sessions that score above threshold to an automated hold or review queue before the exposure compounds. The model operates on behavioral sequence, not rule-based signals, so it catches novel fraud patterns that haven't been explicitly coded.

Friction Diagnosis. The click-context variant surfaces which behavioral patterns are correlated with errors, abandonment, and dead interactions — at element level. Rather than knowing that checkout has a 22% failure rate, you know that cluster #1083 (error loop on the checkout footer) contains 675,000 error events from a specific element that should be investigated first. Prioritization stops being opinion-based.

Talk to your CSE or Account Executive about how Fullstory Professional Services can help evaluate and operationalize these capabilities for your data environment.