Revenue-based analytics and A/B testing for marketing teams that run paid ads. See which campaigns actually make money. Free 14-day trial at https://xitter.gq/t.co/RyAZsIKAH7
Joined April 2023
- Tweets104
- Following2
- Followers486
- Likes281
Humblytics retweeted
Obsessed with checking the new @humblytics dashboard btw. I encourage everyone to switch from Google Analytics to Humblytics
Check the split before you read an A/B test result.
Microsoft's experimentation team found about 6% of its A/B tests had a sample ratio mismatch. LinkedIn saw about 10% on one type of test.
SRM means the visitor counts per variant don't match the split you set.
Microsoft's MSN example: showing 16 cards instead of 12 looked like it hurt engagement. The test failed its SRM check. Variant users engaged so much that a bot filter removed them from the data. Once fixed, the result flipped to positive.
The check:
1) Compare visitors per variant to the configured split.
2) Run a chi-square test. Eyeballing the ratio ignores sample size.
3) If p is below 0.01, find the cause before reading any metric.
52/48 on 10,000 visitors fails. 684/644 on our last homepage headline test passes (p = 0.27).
Give Claude one job every Monday:
“Review live funnels and experiments. Find the largest revenue-weighted leak. Recommend one test and prepare the exact implementation. Never activate, edit, or stop a test without approval.”
The loop:
1) Read current tests.
2) Query the funnel and Sankey.
3) Generate one selector-safe recommendation.
4) Stop at approval.
5) Run and log the result.
Dedication Agents moved from 2–3 to 8–12 tests per month. Eleven of its first 15 tests produced significant improvements.
Autonomy without a gate is not a CRO system. The agent does the repetitive work. A person controls the mutation.
ALT A Humblytics weekly CRO agent card showing Observe, Diagnose, Prepare, Approve, and Learn, with human approval highlighted as the required gate.
The Monday marketing report, before Humblytics was in Claude:
1) Export spend from Ads Manager
2) Pull sessions from analytics
3) Pull payments from Stripe
4) Match them by campaign in a sheet
5) Hunt for the rows that don't match
After:
"Last week's spend, sessions, and Stripe revenue by campaign. Flag anything that doesn't match."
Same report. One message. You spend the time reading it instead of building it.
One Humblytics key can give Claude a CRO queue across 30+ client sites.
Prompt:
“Sweep every property. Rank the five largest opportunities by absolute people lost, not conversion-rate drama. Give me one evidence-backed test per client.”
1) `list_properties`
2) `aggregate_traffic`
3) `get_pages_breakdown` per property
4) Page, click, form, and funnel calls only where the opportunity is large enough
The result is a client queue with traffic, loss, affected users, evidence strength, test, and ICE score.
Dedication Agents used two-week discovery sprints across 30+ sites, then moved from 2–3 to 8–12 tests per month.
The agent removes the blank-page work. The strategist still chooses what ships.
RT @amirmxt: Introducing Humblytics in Claude.
- See which Meta and Google Ads campaigns actually made money, matched to real Stripe reven…
Five questions to ask once Humblytics is connected to Claude:
1) "What changed in traffic this week, and which source caused it?"
2) "Which high-traffic page loses the most visitors?"
3) "Where does my signup funnel lose the most people?"
4) "Which ad campaign produced revenue, not just clicks?"
5) "Which variant of my running test is ahead?"
All five are read-only. Nothing on your site changes.
Number 4 is the one your ad dashboard can't answer.
One sentence can become a live revenue experiment without opening a visual editor.
Prompt:
“Recommend one copy-only test for this page. Use live behavior, keep the offer and layout fixed, and show the exact DOM mutation before activation.”
1) Read `get_page_details` and `get_clicks_details`.
2) Run `get_split_test_recommendations`.
3) Keep recommendations above 70 confidence.
4) Map selector, operation, and value into a preview.
5) After approval, call `create_split_test` with `goal: revenue`.
The Host Report found a 63% winner by rewriting features as outcomes. An exact deadline added another 16% in a separate test.
The agent prepares the mutation. You control activation.
ALT A Humblytics test payload card showing selector .hero h1, a text operation, the variant value Build a direct-booking business, a revenue goal, and an awaiting approval state.
Ask Claude to turn one funnel leak into a pull request.
“Build the signup funnel. Rank steps by absolute people lost. Inspect detours. Find the matching form stage in the codebase. Draft the smallest reversible fix.”
1) `query_funnel` finds the loss and splits it by device.
2) `get_funnel_sankey` exposes detours and drop-off paths.
3) Claude locates the implicated component.
4) It returns a diff, QA checks, and measurement plan.
MatchDay moved its primary funnel from 2.4% to 6.11%. Deferring one high-friction step increased form completions by 180%.
Analytics should end in a testable change, not another dashboard.
One question in Claude, with Humblytics connected:
"Which Meta campaign actually made money last month?"
Claude calls `get_ads_attribution`, joins ad spend to Stripe payments, and answers with one table.
In this demo account, the campaign Meta credits with the most purchases returned $0.75 for every $1 spent. The one with the fewest returned $3.83.
Meta reports purchases. Stripe reports revenue.
Humblytics is now in Claude's connector directory.
Find it in Claude, click Connect, sign in with your Humblytics account. Then ask:
1) Which page loses the most visitors?
2) Where does signup drop off?
3) Which campaign produced revenue, not just clicks?
4) Which test variant is ahead?
No API key to paste. Included on every Humblytics plan.
Ad platforms report the click. Claude can now read what happened after it.
Claude can tell you whether to rewrite a CTA or move it.
Prompt:
“Find high-traffic pages where average scroll depth suggests the primary CTA is too low. Check CTA click share, inspect the page source, and propose the smallest placement test.”
1) `get_pages_breakdown` ranks pages by traffic, bounce, and average scroll depth.
2) `get_page_details` splits the pattern by device and source.
3) `get_clicks_details` measures the CTA's share of clicks.
4) Claude reads the page component and locates the CTA in the DOM.
MatchDay found that 70% of visitors never reached its booking CTA.
A copy test cannot fix a button most visitors never see.
ALT A Humblytics audit card comparing 31% demo average scroll depth with a primary CTA located at 68% page depth, recommending move before rewrite.
The most expensive ad can look healthy inside Meta and still produce $0 in Stripe.
Give Claude this job:
“Find campaigns spending money without attributable revenue. Trace the worst one from creative to checkout. Return pause, fix, or scale.”
The Humblytics MCP chain:
1) `get_ads_attribution` finds spend without revenue.
2) Unmatched campaign and UTM lists expose broken joins.
3) `get_meta_ad_creative` retrieves the promise and landing URL.
4) `get_page_details` and `query_funnel` locate the first revenue-path leak.
Meta rewards clicks. Revenue attribution has to survive the trip to payment.
The Host Report used this full-funnel discipline. Its later copy program found a 63% conversion winner.
The agent returns the evidence and a pause, fix, or scale recommendation. It does not touch the campaign.
Humblytics has 12 marketing skills you can add to Claude Code or Codex.
The useful part is how they chain:
“Find my worst-performing page and fix it”
CRO Optimizer → Page CRO → A/B Test Generator
“Send a weekly performance report with recommendations”
Funnel Reporter → CRO Optimizer
“Cut the worst ad and replace its landing page”
Funnel Reporter → CRO Optimizer → A/B Test Generator → Ad Expert
The API-backed skills read the live funnel. Any test that changes the site still waits for approval.
We usually start a conversion funnel with three events:
1) Landing-page view.
2) Primary CTA click.
3) Form submission or purchase.
Find the largest step loss. Open the heatmap for that page. Turn the most likely cause into one A/B test. Then read the same funnel again.
Humblytics supports page views, clicks, forms, and up to 15 funnel steps. Stream, Bars, Vertical, and Table views are there when the path gets longer.
Start with the smallest path that can prove where revenue is leaking. Add steps when they answer a specific question.
ALT A technical-grid graphic with a three-step funnel card showing page view, CTA click, and form submitted.
Average scroll depth can hide the section that is actually losing visitors.
Run the audit by bucket:
1) Separate mobile and desktop.
2) Start at 0-10%. It always shows 100% reach.
3) Move down in 10% increments and mark the first sharp loss.
4) Compare the click heatmap at that same section.
5) Save the current heatmap, change the page, recapture it, and compare the two versions.
The average tells you how far people went. The bucket and page section tell you what to inspect next.
Before rewriting the whole landing page, audit the first screen.
1) Check mobile and desktop separately.
2) Read the scroll buckets in 10% increments. Find the first section that loses meaningful reach.
3) Open the click heatmap. Check whether the primary CTA is getting attention or competing with secondary actions.
4) Read the first screen as a new visitor: what is this, why should I trust it, and what should I do next?
5) Test one change against a purchase, form, or destination-page goal.
One customer combined heatmaps and A/B testing to move booking conversion from 2% to 8% over three months. The useful work started with the page visitors actually saw.
ALT A technical-grid graphic with a landing-page audit card organized into promise, proof, and primary CTA regions beside a demo reach indicator.
Your landing page is on one domain. Checkout or booking happens on another. The conversion can still land in one funnel.
1) Install the same Humblytics tracking script on both domains.
2) Add the external domain under Settings, then Domains, when your integration requires whitelisting.
3) Fire the external page-view, form, or click event with `domain` set to the primary website.
4) Complete a test conversion and confirm the event appears under the primary property.
The common error is setting `domain` to the page where the event fired. It should name the property that receives the data.
Most A/B tests go wrong before the first visitor is split.
Use this five-check setup:
1) Change one element.
2) Choose the goal before building the variant. Purchase, form submission, click, page view, or destination page.
3) Run the sample-size math. A 5% baseline moving to 6%, at 95% significance and 80% power, needs about 4,000 visitors per variant.
4) Copy the final URLs after every redirect. `/page` and `/page/` are different test URLs.
5) Run through full business cycles and do not call the winner early.
Then compare revenue per visitor, not only conversion rate. A page can produce more signups and less money.
ALT A technical-grid graphic with a sample-size card showing a 5% baseline, a 1-point lift, 95% confidence, 80% power, and about 4,000 visitors per variant.
