Customer Journey Analytics + Claude MCP: From Natural Language to Regression Insights

Ben R Woodard May 20, 2026
cjar mcp ai r

Using Claude’s Model Context Protocol to pull live CJA data and test whether weekday sessions drive deeper content engagement than weekend sessions — from data view discovery through a linear regression model.

Why CJA + MCP?

My previous post covered Adobe Analytics + MCP — using Claude to drive live AA queries through natural language. Customer Journey Analytics is a different beast: it’s built on Experience Data Model (XDM) schemas, data views replace report suites, and the underlying R package is cjar rather than adobeanalyticsr. The MCP layer handles the same translation job, but the schema is richer and the component IDs are longer.

This post walks through a real analysis session on real website data — four natural language prompts, ending in a linear regression that tests a concrete business hypothesis.

Hypothesis: Weekday sessions on the Jimalytics website show higher content engagement (more events per session) than weekend sessions, consistent with a professional readership that’s more active Monday–Friday.

Setup

Authentication uses the same Server-to-Server OAuth flow as the AA connector. The cjar package handles token refresh automatically.

suppressPackageStartupMessages({
  library(cjar)
  library(dplyr)
  library(tidyr)
  library(lubridate)
  library(scales)
  library(knitr)
  library(ggplot2)
  library(broom)
})

cja_auth(type = "s2s")

DATE_END   <- Sys.Date()
DATE_START <- DATE_END - 29
DATE_RANGE <- as.Date(c(DATE_START, DATE_END))

What This Looked Like in Claude

Four natural language prompts drove the entire session:

  1. “What data views do I have access to?” → cja_get_dataviews()
  2. “What dimensions are available in the Jimalytics Website data view?” → cja_get_dimensions()
  3. “Pull 30 days of daily sessions, events, visitors, and time spent” → cja_freeform_table()
  4. “Test whether weekday sessions are more engaged — use a regression model” → Feature engineering + lm() + ggplot2 visualization

The MCP server translates each prompt into the right cjar function call, handles authentication, and returns structured data that Claude then uses to write the analysis. No API documentation was consulted, no token was manually constructed, and no JSON payload was hand-crafted.

The full MCP server source is at benrwoodard/cjar-mcpr and the cjar package documentation is at benrwoodard.github.io/cjar.

Prompt 1 — “What data views do I have access to?”

In CJA, data views are the unit of analysis — each one defines which schema fields are exposed as dimensions and metrics, how sessions are counted, and what container labels to use. The first step in any session is to discover what’s available.

dvs <- cja_get_dataviews(
  expansion = c("name"),
  limit     = 100
)

dvs |>
  select(id, name) |>
  kable(col.names = c("Data View ID", "Name"))

Result (selected rows):

Data View ID Name
dv_xxxxxxxxxxxxxxxxxxxxxxxx Jimalytics Website
dv_xxxxxxxxxxxxxxxxxxxxxxxx Retail Demo
dv_xxxxxxxxxxxxxxxxxxxxxxxx Google Analytics Demo Data View
dv_xxxxxxxxxxxxxxxxxxxxxxxx TechPulse Electronics
… 61 total

The Jimalytics Website data view is the one with real traffic data, his actual site can be found at jimalytics.com. Thanks Jim for letting me use your data for this demo!

Prompt 2 — “What dimensions are available in the Jimalytics data view?”

Unlike AA report suites where dimensions are mostly pre-defined, CJA data views expose whatever fields were mapped from the underlying dataset. Calling cja_get_dimensions() returns the full list — 94 dimensions with data in this view.

dims <- cja_get_dimensions(
  dataviewId = DV_ID,
  expansion  = "hasData"
)

dims |>
  filter(hasData == TRUE) |>
  select(id, name) |>
  arrange(name) |>
  kable(col.names = c("Dimension ID", "Name"))

The dimensions with real data fall into a few clusters: device hardware attributes (Device Type, Screen Size, OS), browser/environment fields, geography (City, Country Code, State), web interaction URLs, time-part dimensions (Hour of Day, Day of Week, Month of Year), and a handful of custom identity fields tied to membership type.

For the engagement hypothesis, the most useful dimensions are the time-parts — they let us aggregate behavior by day of week and hour of day without needing a page-level breakdown.

Prompt 3 — “Pull 30 days of daily sessions and engagement metrics”

One cja_freeform_table() call returns the daily time series. The key metrics here:

  • visits — sessions (the CJA equivalent of visits in AA)
  • occurrences — total events fired across all sessions
  • visitors — unique people
  • adobe_timespent — total seconds spent across all sessions

From these four numbers we can derive two engagement ratios: events per session (breadth) and time per session (depth).

raw <- cja_freeform_table(
  dataviewId  = DV_ID,
  date_range  = DATE_RANGE,
  dimensions  = "daterangeday",
  metrics     = c("visits", "occurrences", "visitors", "adobe_timespent"),
  top         = 0,
  metricSort  = "asc"
)

Result — 30 days of daily data:

Date Sessions Events People Time Spent (s)
2026-04-20 86 105 83 2,600
2026-04-21 83 94 83 860
2026-04-22 169 184 165 5,820
2026-04-23 186 195 183 3,573
2026-04-24 131 142 129 249
2026-04-25 113 134 113 2,372
2026-04-26 58 58 58 0
… … … … …
2026-05-19 173 190 167 1,154

Prompt 4 — “Test whether weekday sessions are more engaged — use a regression model”

Deriving Engagement Metrics

daily <- raw |>
  rename(
    date       = daterangeday,
    sessions   = visits,
    events     = occurrences,
    people     = visitors,
    timespent  = adobe_timespent
  ) |>
  mutate(
    date              = as.Date(date),
    dow               = wday(date, label = TRUE, abbr = FALSE),
    is_weekend        = dow %in% c("Saturday", "Sunday"),
    events_per_session = events / sessions,
    time_per_session  = timespent / sessions
  )

The key derived metric is events per session — how many events (page views, clicks, interactions) a visitor fires in a single visit. A value of 1.0 means every session is a single-event bounce; higher values indicate deeper exploration.

daily |>
  group_by(dow) |>
  summarise(
    n          = n(),
    avg_sessions = mean(sessions),
    avg_eps    = mean(events_per_session),
    avg_tps    = mean(time_per_session)
  ) |>
  arrange(desc(avg_eps)) |>
  mutate(across(where(is.numeric), ~ round(.x, 2))) |>
  kable(col.names = c("Day", "Days in Range", "Avg Sessions", "Avg Events/Session", "Avg Sec/Session"))

Events per session by day of week:

Day Days in Range Avg Sessions Avg Events/Session Avg Sec/Session
Tuesday 5 102.8 1.151 11.91
Wednesday 4 113.5 1.127 17.04
Monday 5 90.4 1.113 8.31
Thursday 4 139.3 1.083 14.83
Friday 4 131.5 1.061 3.39
Saturday 4 84.8 1.081 7.67
Sunday 4 84.5 1.028 0.61

The pattern is directionally consistent with the hypothesis — weekdays cluster higher than Sunday, which sits at the bottom. Saturday, however, nearly matches Thursday and Friday, which is a wrinkle worth noting.

Visualizing the Pattern

dow_order <- c("Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday")

daily |>
  mutate(dow = factor(dow, levels = dow_order)) |>
  ggplot(aes(x = dow, y = events_per_session, fill = is_weekend)) +
  geom_jitter(aes(color = is_weekend), width = 0.15, size = 2.5, alpha = 0.6) +
  stat_summary(fun = mean, geom = "crossbar", width = 0.5,
               color = "#333333", linewidth = 0.7) +
  scale_fill_manual(values  = c("FALSE" = "#2E86AB", "TRUE" = "#F4A261"),
                    labels  = c("Weekday", "Weekend"), guide = "none") +
  scale_color_manual(values = c("FALSE" = "#2E86AB", "TRUE" = "#F4A261"),
                     labels = c("Weekday", "Weekend"),
                     name   = "Day type") +
  scale_y_continuous(limits = c(0.95, 1.40), breaks = seq(1.0, 1.4, 0.1)) +
  labs(
    title    = "Events per Session by Day of Week — Jimalytics Website",
    subtitle = paste0("Last 30 days  ·  ",
                      format(DATE_START, "%b %d"), " – ",
                      format(DATE_END, "%b %d, %Y")),
    x        = NULL,
    y        = "Events per Session",
    caption  = "Source: CJA · Jimalytics Website data view  ·  Crossbar = daily mean"
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title         = element_text(face = "bold", size = 15, margin = margin(b = 4)),
    plot.subtitle      = element_text(color = "#666666", size = 10, margin = margin(b = 14)),
    plot.caption       = element_text(color = "#999999", size = 9, margin = margin(t = 10)),
    panel.grid.major.x = element_blank(),
    panel.grid.minor   = element_blank(),
    legend.position    = "top",
    legend.justification = "left",
    plot.margin        = margin(16, 24, 12, 16)
  )

Simple Model: Weekday vs. Weekend

m1 <- lm(events_per_session ~ is_weekend, data = daily)
summary(m1)
Call:
lm(formula = events_per_session ~ is_weekend, data = daily)

Residuals:
     Min       1Q   Median       3Q      Max
-0.09688 -0.06018 -0.01411  0.03953  0.19489

Coefficients:
               Estimate Std. Error t value Pr(>|t|)
(Intercept)     1.10940    0.01609  68.94   <2e-16 ***
is_weekendTRUE -0.05462    0.03118  -1.75    0.091 .
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.07551 on 28 degrees of freedom
Multiple R-squared:  0.0987,   Adjusted R-squared:  0.0666
F-statistic: 3.065 on 1 and 28 DF,  p-value: 0.091

The weekend flag has a coefficient of −0.055 (p = 0.091). That means weekend sessions average ~5.5% fewer events than weekday sessions. The effect is directionally consistent with the hypothesis but does not clear the conventional 0.05 threshold with 30 observations.

Full Model: Day-of-Week Factors

The simple model collapses Saturday and Sunday together. The full model lets each day speak for itself.

daily <- daily |>
  mutate(dow = factor(dow, levels = dow_order))

m2 <- lm(events_per_session ~ dow, data = daily)
tidy(m2) |>
  mutate(across(where(is.numeric), ~ round(.x, 4))) |>
  kable(col.names = c("Term", "Estimate", "Std. Error", "t value", "p value"))
Term Estimate Std. Error t value p value
(Intercept) — Monday 1.1126 0.0338 32.95 < 0.001
Tuesday +0.0383 0.0478 0.80 0.436
Wednesday +0.0147 0.0491 0.30 0.769
Thursday −0.0292 0.0491 −0.59 0.561
Friday −0.0516 0.0491 −1.05 0.311
Saturday −0.0314 0.0491 −0.64 0.532
Sunday −0.0846 0.0491 −1.72 0.104 
glance(m1) |> select(r.squared, adj.r.squared, p.value) |>
  bind_rows(glance(m2) |> select(r.squared, adj.r.squared, p.value)) |>
  mutate(model = c("is_weekend", "day_of_week"), .before = 1) |>
  kable(digits = 3)
Model R² Adjusted R² p value
is_weekend 0.099 0.067 0.091
day_of_week 0.255 0.018 0.371

The full day-of-week model explains more variance (R² = 0.26) but the adjusted R² collapses to near zero — adding six day dummies to 30 observations is overparameterized. The simple weekend flag is more honest at this sample size.

tidy(m2, conf.int = TRUE) |>
  filter(term != "(Intercept)") |>
  mutate(
    day       = str_remove(term, "^dow"),
    is_weekend = day %in% c("Saturday", "Sunday"),
    day       = factor(day, levels = dow_order[-1])
  ) |>
  ggplot(aes(x = day, y = estimate, color = is_weekend)) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "#999999") +
  geom_pointrange(aes(ymin = conf.low, ymax = conf.high), size = 0.8, linewidth = 1) +
  scale_color_manual(values = c("FALSE" = "#2E86AB", "TRUE" = "#F4A261"),
                     labels = c("Weekday", "Weekend"), name = "Day type") +
  labs(
    title    = "Day-of-Week Coefficients for Events per Session",
    subtitle = "Reference level: Monday  ·  95% confidence intervals",
    x        = NULL,
    y        = "Coefficient vs. Monday",
    caption  = "Source: CJA · Jimalytics Website data view"
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title    = element_text(face = "bold", size = 15, margin = margin(b = 4)),
    plot.subtitle = element_text(color = "#666666", size = 10, margin = margin(b = 14)),
    plot.caption  = element_text(color = "#999999", size = 9, margin = margin(t = 10)),
    legend.position = "top",
    legend.justification = "left"
  )

Conclusion

The hypothesis holds directionally but weakly. Weekday sessions average ~5.5 more events per 100 visits than weekend sessions (1.109 vs. 1.055), and Sunday sits lowest at 1.028. The simple weekend flag produces p = 0.091 — suggestive, but not conclusive at 30 daily observations.

A few things to note about the data and the limits of this analysis:

  • Saturday behaves like a weekday in terms of engagement, nearly matching Thursday and Friday. The weekend penalty is driven almost entirely by Sunday. That may reflect a specific posting pattern — content published midweek that draws Saturday readers who are still in “professional” mode.
  • The 12:00 AM hour bucket in the hourly breakdown contains a suspicious spike (584 sessions vs. the next highest at 172). This likely reflects timezone normalization in the data view assigning unclassified timestamps to midnight. It’s worth investigating before using hourly data in any model.
  • adobe_timespent is noisy at the daily level — many days show near-zero totals, suggesting the time-spent implementation only fires on page exits and may miss a large share of sessions. Events per session is the more reliable engagement signal here.

With a longer data window — 90+ days — the weekday/weekend split would likely harden into a clean, significant result. The direction is consistent enough to act on for scheduling content or optimizing campaigns by day type.

Disclaimer: This post was developed using Claude AI with live data from the Jimalytics Website CJA data view. Regression results are based on approximately 30 days of observations and should be interpreted as exploratory rather than confirmatory. Always validate findings against a longer baseline before using them to drive decisions.