Growth Marketing
Engineer.

Product Genius posted the role. I researched the company the way I would research a target account, broke the posting down line by line, and brought the workflows that fit it. Scroll.

See more at robynrocha.io.

Product Genius adds a feed to the store that learns each shopper, while they shop.

The engine is what they call a Large Interaction Model — per their whitepaper, a new approach to training LLMs built to learn continuously, on limited data. It selects products, reviews, video, Q&A and brand stories per visitor, per moment, while the session is still running. Distribution is a Shopify App Store install or a headless integration; Pro pricing starts at $999/month; founded 2022, Cambridge MA, 11–50 people.

How the product works — in their words, "a website that thinks."

Two systems, one round trip: an interaction system deciding what appears next in-session, and a learning system that adapts continuously — and, per their FAQ, learns "the best AI sales principles" from shoppers across the many stores it serves. The acquisition channel is an input to the personalization decision.

PRODUCT LINES, AS PUBLISHED · Revenue Genius — the adaptive experience · Content Genius — learns which content performs, reusable across site, ads, and email · Design & Development Genius · enterprise: headless, APIs, conversational search, multi-brand

The 15-day trial is the sales mechanism.

Company-reported: +20% average on site messaging, +36% average across the 25-brand A/B test set, announced with Google. The published minimum for the system to work: 10,000 shopping sessions a month. My read of the constraint: a trial that can't produce a readable result can't do the selling — however interested the merchant is.

The real claim is sample efficiency, not personalization.

• Continuous learning, not a slow retrain cycle — the learning rate is deliberately tuned to move "faster than the speed of statistics," with the reported finding that higher learning rates produced greater revenue lift.
• Significant improvement claimed in less than a dozen interactions · works on stores with as few as 10,000 shopping sessions a month · no third-party data — framed explicitly against tightening privacy law.
• Why that matters (my read): personalization is a crowded category. Personalization that produces measurable lift on a small store inside a two-week window is a different claim — and it is the one the whole commercial motion rests on. Everything this site proposes serves it: merchants who can produce a readable answer, fast.

DEPLOYMENT · zero code on a Shopify install · under 30 minutes of developer time elsewhere · storefront, headless/API, iOS and Android, kiosks · listed integrations: Klaviyo, YouTube, Instagram, any analytics platform

The team.

The chairman created the previous generation of product discovery for e-commerce. This board has seen this category before, one technology cycle earlier.

And everyone else — scientists, engineers, product, and the orbit around them — is on the next slide, with the source page ↗ one click away — every profile verified by hand before this interview.

A long-tenured research core, and a 2026 commercial wave.

PUBLICLY LISTED INVESTORS & ADVISORS · Scott Silverman ↗ — co-founder of CommerceNext, investor/advisor · Matthew Sutton — seed investor

The shape of the room: the research core predates the company — the VP of ML Research has been with Ben Vigoda since Gamalon in 2014, and three engineers span four-plus years. Then 2026: a President & COO in April, two staff engineers in April, an account executive in June. This posting looks like part of that wave — the commercial build-out is happening right now.

I collected your entire live ad presence before we spoke.

Every creative captured, every row linked to the platform's own record. One thing I'd want to learn from the team: how these are evaluated today, and what you already know is working.

Two observations from the ads.

Both stated as the observable fact, then the question I'd actually ask. From outside I can see surfaces, not reasons — so these are curiosities with evidence attached, never conclusions.

Two questions I'm genuinely curious about.

These are questions, not conclusions. The answers would shape where I'd start — which is exactly why this deck proposes conversations before it proposes changes.

The stack firing on your public pages.

A more instrumented stack than most companies at this stage carry — the measurement layer was clearly a deliberate, early investment. From outside I can see what's installed, not how each piece is used: the identity and analytics tools here serve different jobs, and I'd want to learn the role each one plays before proposing anything around them.

What you're probably running — and where it overlaps my experience.

The tags on your pages imply the workflows behind them. Naming those is an inference I'm happy to be wrong about — the point is that every workflow here starts from your installation, not from a generic playbook. The workflows on my site show the process — I've custom-configured them for clients, and each adapts to the stack that's actually in the room.

EVERYTHING IN THIS SECTION KEEPS TWO RULES · "probably running" stays clearly labeled as inference · every workflow works with what's installed — nothing requires replacing a tool

Your stack points to warm ad audiences — a workflow I know well.

Primer's job is audience activation: keeping a warm segment — engaged contacts, qualified accounts — synced to the ad platforms as people qualify in and out. I've constructed and documented this kind of workflow myself, in different configurations for different stacks. One example, in outline:

IF YOU'D LIKE TO SEE THIS EXAMPLE IN FULL · robynrocha.io/workflow/i2

RB2B, Vector, and Leadfeeder identify the companies visiting you. Visitor outbound is familiar ground.

Identity tools exist to turn anonymous traffic into named companies, usually surfacing to the team as alerts or lists. I have experience constructing visitor-outbound workflows on this same signal — RB2B among the tools — and each implementation took its own shape. One example, in outline:

IF YOU'D LIKE TO SEE THIS EXAMPLE IN FULL · robynrocha.io/workflow/s6

Your Google Ads wiring already ties clicks to conversions. The rest is where my experience is deepest.

Passing the click ID is what keeps a click, a form fill, and a record on one thread — so click-to-form attribution is probably already working. The harder half for any team is everything after the form, and no ad platform can see it. That half is where my attribution experience is concentrated. One example, in outline:

IF YOU'D LIKE TO SEE THIS EXAMPLE IN FULL · robynrocha.io/workflow/c2

You measure broadly. What ties it together: one shared language.

GA4, PostHog, Clarity, and GTM are all live — page analytics, product events, session replay, and the tag layer to manage them. The work is a definition, not a tool: the moments that matter commercially, named once and shared by all four — and you already own every tool it needs.

• Define the business events once — demo_requested, install_started, install_completed, trial_started, trial_readable, customer_won — in the tag layer you already run, so every tool reports against the same moments with the same names.
• Join them to the account record, so a funnel isn't a chart in one tool but a property of the merchant — and any stage's conversion can be segmented by source, campaign, and fit grade.
• Each tool then does what it's best at: GA4 for acquisition reporting, PostHog for funnel and cohort analysis, Clarity for watching where a stage loses people — all speaking the same event names.
• This is the foundation the posting's "treat growth like a product" responsibility stands on — instrumentation first, then experimentation has something trustworthy to read.

RUNS ON · GTM as the definition point · GA4 + PostHog + Clarity as the read surfaces · the CRM as the join · deliverable: a metric dictionary the whole team agrees to, before any dashboard is trusted

The advocacy engine your sales motion already earns.

You own two review surfaces — the Shopify App Store listing and the claimed G2 profile. In a category sold on measured proof, reviews from merchants whose trials succeeded are among the strongest assets available — and your trial creates a precisely knowable right moment to ask.

• The ask is keyed to trial state, not to a quarterly campaign — right after the ROI review, while the result is still fresh and specific for the merchant.
• Advocacy is made upstream: the pipeline for reviews is trials that resolve cleanly, which is the same thing the whole growth system optimizes anyway.
• References, speakers, and case-study candidates come out of the same moment — tracked as record state, so "who could speak to this" is a lookup, not a last-minute search before an event.

RUNS ON · trial state on the record as the trigger · your two existing review destinations · nothing sends without the same human gates as everything else

Five paid rails are live. The cadence judges them all by the same unit.

Google, Meta, LinkedIn, Reddit, and StackAdapt tags all fire on your pages — a wider paid footprint than most companies your size carry. The cadence I'd propose treats them as one portfolio with one judging unit: cost per trial that reaches a readable result.

• Each rail gets a stated role — as a starting proposal: search captures demand that already exists; Meta and LinkedIn educate and retarget — the warm audience from the Primer slide is their sharpest segment; Reddit and programmatic run as measured tests with explicit scale-or-stop rules. Your own results would redraw these roles.
• The weekly cadence is fixed: verify the data before reading it · review every rail at cost-per-readable-trial, not platform metrics · move budget toward evidence · launch one creative test drawn from the angle families already live in your archive.
• Creative tests inherit the experiment discipline — one variable, a comparison held, a quality guardrail (readable-trial rate), and a decision rule written before launch. More clicks with worse trials is a rejected test.
• Budget changes are human-approved by design — analysis can recommend doubling spend; it cannot do it.

RUNS ON · the five pixels already installed · the warm audience as the retargeting segment · attribution from the previous slide as the judge · your live ad archive (165 LinkedIn · ~51 Meta · 3 Google, all in this deck's explorer) as the creative inventory

The Growth Marketing Engineer role.

This is not a channel-operator seat. The posting asks for someone to design and operate the full pipeline from first awareness to a readable trial result: the shared record every channel writes into, the stages that move a merchant forward, the gates that protect the sales calendar, and the feedback loop that returns every outcome to targeting, message, and creative. Summarized in one sentence: engineer the system that reliably turns attention into trials that can prove the product works.

What this role is not.

• Not a campaign coordinator waiting for briefs — and not a pure paid-media, outbound, RevOps, content, or software-engineering seat. The posting combines those deliberately, because it wants growth to run as one measurable product.
• Not an automation builder measured by the number of workflows, and not a content engine optimizing for volume. Every workflow and every asset exists to move a merchant toward a trial that can prove value.
• Not a data analyst reporting past performance. The standard in the posting is insight that causes a decision — a dashboard alone doesn't qualify.

The central test of the seat: establish a trustworthy learning system around growth.

Five capabilities that have to work together.

The likely test of a candidate is whether they can move among all five without treating any one of them as the entire job. The thirteen responsibilities in the scope section — the close of this site — are these five capabilities written out in detail, one slide each.

The posting's thirteen responsibilities, and where each is covered.

One slide per responsibility follows — each with what it means at Product Genius, the general workflow shape I'd propose, and the evidence I bring to it.

One system, one record, every channel writing to it.

The posting asks for end-to-end infrastructure integrating referral, paid, outbound, social, CRM, and product signals. At Product Genius that means: one merchant record, agreed stage definitions from awareness to a readable trial, and a loop that returns every outcome to targeting, message, and creative.

• Stages need agreed definitions before anything else — qualified, install-complete, readable — because every metric downstream inherits them.
• Channels are judged by what their merchants do downstream, never by their own dashboards — which is why every signal writes to the same record.
• The architecture is drawn with the team in the first weeks, starting from what already exists — the audit's stack table is the opening inventory, not a teardown list.

EVIDENCE · I operate a system with this exact shape today: a graded account store, staged records, gated sends, and outcome loops feeding targeting — constructed end to end.

The five workflow families your posting names, and my shape for each.

The posting lists them: lead enrichment, segmentation, outbound sequencing, content generation, performance analysis. General shapes, each of which I operate in production today — the wiring adapts to Clay/Apollo-class tools or the stack I use.

• Enrichment — check what's already known first, verify every address before it's written, stamp source and date. No lookup bought twice; no trusted field silently overwritten.
• Segmentation — fit (could a trial resolve here?) scored separately from intent (why now?), because a high-intent merchant below the session floor still can't buy.
• Sequencing — suppression re-checked at the moment of send; whatever was promised travels with the contact so the first human sentence continues the conversation.
• Content generation — AI transforms approved claims; it never invents the claim, the evidence, or the point of view.
• Performance analysis — a fixed diagnostic order that verifies the data before any narrative about the market. Detail under item six.

EVIDENCE · production versions of all five operate in my current system — with human gates on every consequential action.

Event infrastructure for NRF and ShopTalk.

• Data rights confirmed before a single row is captured — public, partner-provided, or licensed; never scraped indiscriminately.
• Attendance is intent, not fit. A promising booth visitor under the session floor gets a relationship, not a sequence.
• Follow-up latency is the metric the whole system serves — event context decays in days.

TOOLS · capture: permitted lists into a working table with full provenance — the same wave-sourcing pattern my pipeline uses · enrich: gaps only, never re-bought · reconcile + briefs: the CRM · alerts: Slack

Open questionWhat did the last NRF produce, and where does that data live today?

Models classify. Code decides.

Your posting flags the risk of business logic hiding inside prompts. The architectural answer: models read, extract, match, and summarize — always bounded, always schema-validated. Only deterministic code scores, routes, gates, and writes. The logic stays in code that can be reread and rerun.

• Every worker carries a contract — allowed inputs, allowed tools, output schema, cost budget, failure state, and an owner. An analysis worker can recommend doubling spend; it cannot do it.
• Autonomy is earned per class of action — total gating at launch, relaxed only where evidence shows the human and the system already agree, irreversible actions last.
• Review is a practiced discipline, not a claim: my standing review order covers purpose, schemas, external calls, state, failures, observability, and model boundaries — and it has a concrete example: a crawler returning HTTP 200 that was actually a bot wall, caught by validating the artifact instead of trusting the status code.

EVIDENCE · the research behind this deck was conducted under this discipline · n8n is named in your posting — the same review order applies to whatever runs the integrations.

Own the channels — and judge every one at the trial.

Paid and outbound, Google to LinkedIn to email to partnerships. The operating principle across all of them: a channel is worth what its merchants do downstream — install, trial, readable result — never what its own dashboard says.

• Paid runs on the five-rail cadence in the audit-surfaced section — stated roles per rail, one judging unit, weekly evidence-driven budget moves.
• Outbound runs gated end to end: verified addresses only, suppression read at send, campaigns created paused and human-approved before the click — volume is never the metric.
• Partnerships and referral enter the same record with the same attribution, so partner-influenced trials are visible instead of anecdotal.

EVIDENCE · I operate the outbound machinery daily in production; paid economics are read at cost-per-readable-trial from day one.

Instrument first. Then experiments have something to read.

CAC, conversion, and pipeline quality only mean something on top of trustworthy instrumentation — which is why the event-language work in the audit-surfaced section comes first. On top of it, two disciplines:

• Every experiment has the same anatomy: audience and the belief being tested · treatment and comparison · unit of assignment · one primary metric · a quality guardrail · contamination risks named upfront · the decision each outcome triggers.
• "What went wrong in the last few days?" has a fixed order: verify the data first → locate the change by segment → separate fewer-conversions from more-low-intent-traffic → check operational changes → state confidence honestly → the smallest useful next action.
• "What just went right?" gets the same discipline in reverse — confirm the lift is real and which segment created it, check quality held, then capture the winning condition and expand carefully.
• The honest edge: your adaptive trials are a harder class of measurement than fixed-creative tests — your posting offers to teach that statistical layer, and it's part of why this role is attractive.

EVIDENCE · every claim in this deck carries its provenance; the same habit applied to funnel data is what makes a dashboard worth trusting.

Where I'd use Product Genius on Product Genius.

• Personalized proof on high-intent pages. A visitor arriving from NRF, a Shopify search, or an agency referral needs different proof. Use the product itself to adapt which case studies, Q&A, and explanations appear — with a holdout preserved — and judge by qualified installs and downstream trial quality, not clicks.
• The analytics agent for friction discovery. Point the synthetic-shopper capability at your own funnel pages to look for friction — missing proof, unclear explanations, dead ends — validated against real visitor behavior before anything changes.
• Content learning across channels. Content Genius already learns which stories, reviews, and answers work per context on merchant sites — apply the same evidence to choose what runs in ads, nurture, and event follow-up.
• Trial nurture. Use account and trial state to show a merchant the right implementation help or ROI explanation at the right moment — kept carefully separate from what the product does to their shoppers.

EVERY EXPERIMENT HERE KEEPS THE SAME SHAPE THE PRODUCT IS SOLD ON · a holdout · one primary metric · a quality guardrail · a decision that follows the result

A dashboard becomes a loop only when evidence changes a decision.

The posting asks for prospect behavior, campaign performance, CRM activity, and revenue outcomes connected into actionable intelligence. The shapes:

• Objections and buyer language return from sales as structured evidence and feed messaging and targeting directly.
• Trials that fail to read out feed the scoring priors: every unreadable trial teaches the system what to stop routing to sales.
• Every human-reviewed judgment becomes a training corpus — supervising the system is the same act that improves it.
• The loop test is always the same: evidence changed a decision, an owner acted, and the next result was measured. Anything less is reporting.

EVIDENCE · my reply-handling system runs exactly this way today: gated verdicts accumulate into the corpus that tunes the next round.

Scoring, routing, and a handoff that answers five questions.

Sales collaboration works as infrastructure that holds between meetings. Three mechanisms:

• Scoring: hard gates first, then a ranked grade. Anything failing platform, session floor, or surface checks never reaches outreach, however appealing — what passes arrives pre-qualified and ranked.
• The handoff answers five questions on every routed account: why this merchant · why now · what they already heard · who matters (buyer plus the technical owner the install needs) · what happens next — plus trial-readiness, known before the call.
• Enablement stays current by construction: briefs and objection answers generate from the same claim records the ads use, so sales never carries a stale number.

EVIDENCE · I operate this handoff in production; the five-question record exists because I've held the receiving seat myself.

Strategy resolves before a word is written.

Organizational knowledge, CRM data, and sales conversations become ads, outbound, landing pages, and email — through a chain, not loose drafts:

• The chain: strategy (play, offer, angle, audience posture) → slices, each carrying the decisions that produced it → the authored pack → variation → handoff. The offer is positioned here, never invented here.
• A claim registry underneath: every claim carries its source, proof level, approval, and date. Proof level bounds how hard any asset may push; when a number changes, one record updates and every surface inherits the correction.
• AI does the volume; humans approve the claims — the same gate pattern as everywhere else in this deck.

EVIDENCE · this chain is my production messaging system; the registry design exists because I've watched claims drift across surfaces without one.

Creative testing starts from the angles already live.

Your archive — 165 LinkedIn ads, ~51 Meta ads, 3 Google ads, all in this deck's explorer — is a rich input inventory. The testing discipline:

• Angles become hypotheses: each live angle family implies a belief about the buyer — proof-led vs. capability-led, founder-voice vs. product-voice. A test names the belief before it varies the creative.
• One variable at a time where possible, structural variation where not — and every variant carries lineage, so a winner or a failure traces to the exact choice that caused it.
• The guardrail outranks the click: a creative that lifts clicks while degrading readable-trial rate is a rejected test, not a win.

EVIDENCE · the variation-and-lineage discipline is my production system for outbound copy; the same structure applies to paid creative.

A new channel earns a controlled, instrumented test.

The posting leaves the channel list open on purpose. The discipline that keeps exploration honest:

• Every candidate channel gets the same entry contract: a written hypothesis, a capped budget, defined instrumentation, and a scale-or-stop rule agreed before launch.
• Judged at the same unit as everything else — cost per readable trial — so a novel channel can't hide behind novel metrics.
• Sequenced by evidence, not enthusiasm: one new channel at a time, because three half-instrumented experiments teach less than one clean one.

EVIDENCE · the same contract governs my own channel tests; the stop rule is written first because I've seen what happens when it isn't.

Humans own the trust. Systems carry everything else.

Creators, agencies, ecosystem partners, community, and customer advocates — the channels that resist automation. The split that keeps them honest:

• Stays human: trust and the ask, partner strategy, appreciation, executive conversations, repairing anything that goes wrong.
• The system carries: discovery of who actually matters, context packs before every conversation, reminders and follow-through, introduction paths, and attribution of referred trials so the channel's value is visible.
• Advocacy is made upstream — the review-and-reference engine in the audit-surfaced section is this responsibility's machinery: the right ask at the right post-trial moment.

EVIDENCE · six years of carrying the revenue conversation personally — the system exists to make the human better prepared, never to impersonate one · a documented example: robynrocha.io/workflow/c1

No claim outruns its proof. Depth matches the reader's stage.

The judgment layer on top of the content systems — the posting treats it as an engineering concern, and it is one:

• The checks on every asset: is it true and does the cited proof support this strength of claim · does the depth match the reader's stage · is there a real reason to care · does it sound like Product Genius, or like generic AI software.
• Stage-matched depth, concretely: cold audiences get the problem and the outcome; technical buyers get the method — your whitepaper is the asset; active trials get their own numbers, nothing else.
• The protection is systemic, not a memo: the claim registry makes drift structurally difficult — which is what makes the bar enforceable at volume.

EVIDENCE · volume costs little now; the bar is what keeps trust compounding — and this deck is written to its own standard.

You asked for: systems thinking, analytical rigor, experimentation.

You asked for: technical fluency, workflow review, AI fluency.

You asked for: no playbook, real relationships, editorial taste.

Strong, adjacent, ramping — honestly.

Your posting asks eleven questions. My short answers.

Each of these is an operating question, not a slogan — and each one has a fuller answer somewhere on this site.

Where I'd start the conversation.

• Where does the funnel lose the most qualified merchants today — in your view, not mine?
• What makes a trial valid internally, and how often does one end without a usable result?
• Is the immediate priority the self-serve Shopify motion, enterprise, or both at once?
• What is authoritative for an account today, and where does trial state live?
• How much of this role is direct implementation, and where does it partner with engineering?
• What is already working that I should be careful not to break?

I would rather spend the first month being right about the constraint than fast about the solution.

PREPARED BY ROBYN ROCHA · EVERY CLAIM ON THIS SITE CARRIES ITS SOURCE — THE AD ARCHIVE LINKS TO THE PLATFORMS' OWN RECORDS

See more at robynrocha.io.