When marketing runs itself

Overview

For most of marketing’s history, humans drove every step of the cycle: plan the campaign, write the copy, ship it, read the report, adjust, repeat. But across Enrich Labs’ customer base, lean teams are delegating a growing share of that cycle to Helena, our AI marketer, which is speeding up their work.

Taken far enough, that trend points to a marketing operation that improves itself: a system that not only executes campaigns but audits its own automations, schedules its own follow-up reviews, quarantines its own bad data, and proposes its own next experiments. This is self-improving marketing. AI researchers call the general version of this pattern recursive self-improvement: a system capable of improving the system that improves it. Anthropic’s recent essay “When AI builds itself” documents that loop beginning to close inside AI development; this research documents a bounded version of it closing inside marketing. We are not fully there yet, and it is not inevitable. But it could arrive sooner than most marketing organizations are prepared for.

Using live production data from more than 1,000 active customer accounts, plus previously unreported reconstructions of every scheduled task, memory, and learning file at four of them, this research shows the loop is already partly closed. To take just one example: at a telehealth company, Helena has run a self-scheduled SEO improvement cycle every two hours for the past nine weeks (543 runs), each one reading the learnings file written by the previous runs before deciding what to do next.¹

The trends discussed in this piece suggest marketing systems are going to become much more capable in the coming years. That has huge implications. Marketing that improves itself compounds: every audit makes the next campaign better, every post-mortem makes the next launch cheaper. But it also changes the job. When execution costs almost nothing, the humans who win are the ones who set direction, judge results, and decide which experiments are worth running at all.

The delegation has happened in stages, each one moving more of the marketing cycle from the human to the system. The diagram below shows those stages as they appear in our customers’ accounts today: each row is one stage in that delegation, and the starburst at the right is the output: the campaigns, content, and reports the business actually ships.

☺

Marketer

◫

Tools

→

✹

Stage 1: A marketer drives disconnected tools by hand, copying numbers between tabs.

☺

Marketer

◫

Tools

✎

Chatbot

→

✹

Stage 2: Chatbots draft captions and subject lines, but the marketer still pastes the output into the tools.

☺

Marketer

◫

Tools

✎

Chatbot

Helena

→

✹

Stage 3: Helena connects directly to the ad accounts, email platforms, and analytics, and executes campaigns and content herself.

☺

Marketer

Helena

⌛

Schedules

✍

Learning files

↻

Self‑review

→

✹

Stage 4: where the loop closes today. Helena schedules her own follow-up reviews, writes what she learned to persistent learning files, audits her own automations weekly, and starts each run from what the last one knew.

At each stage the human’s role narrows to setting direction and judging results, while the loop does the rest.

Evidence from across the customer base

The study base is more than 1,000 active customer accounts across 11 industries and more than 35 countries. For channel-level analysis, we sampled the 50 highest-spend accounts per channel from the hundreds with connected ad, search, and email platforms; more than $4.5 million in Meta and Google ad spend flows through those sampled management windows alone, along with 92,740 purchases, 81,634 leads, and roughly $1.6 million in Google ad-attributed revenue: live production accounts, not pilots. And because each figure covers only the account’s Helena window, not its lifetime, the totals understate what sits under management. After Helena went live, click-through rate improved on 26 of 50 sampled Meta accounts, and cost per acquisition fell on 12 of the accounts with measurable conversions. Weekly organic sessions rose on 21 of 50 accounts with a measurable baseline; Search Console clicks rose on 13 of 28. Twenty accounts run two or more standing ad-management automations (audience-fatigue alerts, weekly negative-keyword mining, conversion-tracking audits), and eleven email accounts run recurring flow-health audits. Roughly $4.5 million in Meta and Google spend sits under these management windows: live production accounts, not pilots.² The lift concentrates on accounts that let Helena build rather than just watch.³

The most telling evidence, though, is not any single delta. It is the shape of the work: standing audience-fatigue alerts, weekly negative-keyword mining, conversion-tracking audits that hunt for spend with zero conversions, recurring flow-health checks: always-on guardrails humans forget, analyst work on a cron.

Aggregate deltas say a lot about diffusion. But they cannot reveal whether the system is improving itself. For that, we need direct evidence from within the accounts.

Evidence from within four customers

Running a marketing operation takes two broad categories of work. There is execution: writing the ads and the articles, building the flows, shipping the reports. And there is judgment: deciding what to test, interpreting what comes back, and figuring out which ideas to try next. To see how far Helena has come on both, we reconstructed every scheduled task, memory, and learning file at four customers (a telehealth company, a craft-supplies e-commerce brand, a pain-management practice, and a multi-location medical group), supplemented by channel deep-dives across the 50 highest-spend Meta, Google, and email accounts sampled from that base. The evidence forms a ladder: Helena runs experiments to goals humans set, steers campaigns with her own judgment, proposes her own experiments, checks her own work, and has begun improving the system that does all of the above.

Helena is good at running experiments to hit a goal someone else has set.

The clearest place to watch this is Google Ads. The cleanest experimental loop in the data belongs to a craft-supplies e-commerce brand. In May, Helena built a Google Ads search campaign from scratch (it is literally named “Helena Search” in the account) and then scheduled her own graders on it: a 14-day ad-group split check, a 7-day responsive-search-ad review per ad group, a 48-hour post-fix report after a bidding change. Twenty-five runs of a standing two-week check-in followed the campaign from May 14 to July 5.

The iteration is legible in Helena’s memory log, dated like a lab notebook. June 6: conversion tracking for May 22–28 was confirmed broken, so that week is quarantined: “Do not use May 22–28 data as a performance baseline. Use May 1–21 for clean historical benchmarks.” June 15: the campaign switches from manual CPC to Maximize Conversions with a $45 target, and the memory warns future runs not to compare against pre-switch data. June 24: an underperforming secondary-product ad group is paused and budget cut. And on July 1, the post-mortem that closes the loop:

The 14-day review showed the campaign was not exiting Smart Bidding’s learning period because the $29/day budget equaled the $29 target CPA – should be ~10x for Maximize Conversions to learn properly – yielding 1.69x ROAS vs. the account benchmark of 5.09x.

Helena

Memory log, craft-supplies e-commerce brand*

A root-cause diagnosis of a failed experiment, recorded so it can never be repeated, sitting alongside the note that one sub-group inside the paused campaign was actually working at 5.97× ROAS. Meanwhile the account-level numbers moved the right way: Google Ads conversions rose from 40 to 95 per week (+136%) as CPA fell from $26 to $14 (−46%), and blended ROAS climbed from 4.1× to 6.5×.³

The same loop shows up where the starting point is worse. A pain-management practice arrived in April with ad accounts that had spent $0 for months. Helena diagnosed every blocker (billing, verification, quality score) and rebuilt the paid engine: $11K in spend now serves profitably, producing 1,100 conversions at roughly $10 each, with the flagship campaign at an $8 CPA. Then she kept iterating: her weekly learning file contains a section most human marketers never write down, “Tried & Ruled Out”: hyper-local keyword campaigns killed for low search volume, a manual $8-CPC-cap strategy replaced by Performance Max. When that pivot produced a 4.5× jump in daily paid clicks (from about 60 a day to about 776, at a $0.90 average CPC), the same file did not celebrate. It flagged its own result for audit:

Conversion quality is unverified. 1,039 conversions in 7 days is very high for a local medical practice… Need to audit the conversion action definitions.

Helena

Weekly learning file, pain-management practice*

Across the cohort the pattern repeats in smaller strokes: at a B2B SaaS company, a standing conversion-tracking audit (“identify campaigns with zero conversions despite significant spend”) fed a weekly report loop that took CPA from $308 to $91 (−70%). At an online education marketplace, a weekly negative-keyword mine runs on rolling 60-day search-term windows. Ad hygiene, in other words, is no longer an occasional project. It is a schedule.

Helena is getting better at steering campaigns toward outcomes.

Paid social is where judgment shows. The day-to-day work of Meta ads is a chain of judgment calls: is this creative fatigued or fine? Is this dip a trend or noise? Scale, hold, or kill? A multi-location medical group with 170+ providers, live only since June 6, shows what those calls look like when they are written down. Within 90 minutes of onboarding, Helena had recorded operational guardrails a contractor might take weeks to learn: which insurance plans never to reference in content, which nine cities to target. Its sister pediatrics account carries explicit anomaly thresholds: flag any campaign with CPA above 2× its 7-day average, flag zero conversions for 3+ days, flag day-over-day spend spikes above 30%, and do not scale Performance Max until conversion tracking is verified.

Each week, an autonomous-learning run reads the account’s Meta learning file (the standing instruction is to “build on prior insight rather than starting cold”), pulls fresh platform data, and rewrites the file with what changed. Six weeks in, Meta CTR is up 16% (1.39% to 1.61%) and ad frequency is down 23% (4.2 to 3.2) with engagement up: the signature of an account being actively managed rather than left to burn out its audience. Unit costs confirm it: pulled live from the Graph API, cost per landing-page view fell from $0.66 to $0.55 a week over the same window, even as budget guardrails cut spend 35 percent in mid-July. And the July 21 learning file reads like a competent analyst overruling their own earlier hypothesis:

The prior concern about creative fatigue appears premature… CTR and CPC are actually improving, which confirms the audience seeing the ads is still engaging. But the pool of targetable users is shrinking. If this pattern continues into Jul 20–26, the audience needs to be broadened or new creative needs to be added.

Helena

Weekly learning file, 170-provider medical group*

Where Helena writes the ads herself, she iterates against explicit baselines. At a wellness studio, new treatment creatives were drafted with the target written into the task: the original ad set runs CTR 2.0% at $24 per lead: beat it. Over the management window, purchases per week rose 144%, CPA fell from $392 to $174 (−56%), and frequency dropped 27%; one lead campaign delivered 272 leads at a $5 CPL. At the education marketplace, the loop runs at analyst depth: a “Catalog Hidden Gems” deep dive surfaced under-funded tier-2/3 catalog campaigns that now run at 28.9× ROAS on $99K of spend, alongside a creative-diagnostic rewrite of the ESA program and a daily month-to-date CAC report that lands before the team’s standup.

And the loop replicates to newly connected channels. The craft-supplies brand wired up Meta on July 2; by July 17 Helena had already scheduled a standing weekly creative-performance check on it. The account runs at 7.0× Meta-attributed ROAS with purchases up 28% week-over-week against the pre-Helena baseline. And cohort-wide, standing audience-fatigue alerts (flag frequency above 3.5, or a 20% CTR drop) are among the most common automations customers run.³

And given budget authority, the loop scales spend, not just efficiency. At the craft-supplies brand, spend rose 20 percent while purchases rose 28 percent and CPA fell, holding a 7.0× ROAS; at a conversion-optimization agency, weekly Meta spend grew from CAD 388 to CAD 1.9K, almost five-fold, with click-through up 40 percent.

Helena is getting better at proposing her own experiments.

In search, Helena no longer waits to be told what to optimize. The most fully closed loop in production belongs to a telehealth company. Since May 15, Helena has run a self-scheduled task called “SEO Autopilot – Every 2 Hours”: 543 runs by July 21.¹ Each run opens by reading a state file and a learnings log written by the previous runs. The standing prompt is explicit about the loop:

Read learnings from previous runs – apply these to improve quality… Finish it fully, verify quality, capture learnings, update state, then stop. The next run triggers in 2 hours.

SEO Autopilot

Standing prompt, telehealth company*

The autopilot scores every page by a formula (impressions times the gap between expected and actual click-through rate), fixes the highest-leverage page, and moves on through three phases: blog optimization, keyword-gap articles, internal linking. A companion weekly loop, “Orphan Fix,” started from a baseline audit that found 2,111 published posts with zero clicks and works through the ten highest-impression orphans each week, counting recoveries against that baseline. The result is quality, not just output: click-through rate up 11.9% versus the trailing four-week average, average position improved from 10.87 to 10.03, impressions down 7.9%: the account is ranking for fewer but better queries.

The same discipline shows up at smaller scale. The pain-management practice’s organic side grew alongside its ads rescue: weekly organic sessions up 85%, Search Console clicks up 112%, with the site now ranking #2 for a high-intent treatment query. At the craft-supplies brand, a weekly SEO CTR monitor audits high-impression pages on a rolling 28-day window. At a baking-supplies retailer, the loop is editorial: write a post targeting a named 8,100-searches-a-month keyword, run a monthly keyword-opportunity report on what actually ranked, then retro-fit internal links across 35 of her own previously published posts.

And a new surface is emerging: GEO, ranking in answer engines. GA4’s AI-referral channel is already measurable across the cohort: 6.9K sessions at the education marketplace (which runs a standing “Weekly SEO & GEO Retro”), 5.9K at a vacation-rental platform, and at the craft-supplies brand, 33.8 AI-referral sessions per week where there were zero before Helena’s answer-engine-optimized articles. The education marketplace’s average position improved from 13.4 to 11.9 over the same window. The loop that improves classic SEO is the same loop pointed at a new target.

The loop also runs at local scale. At a local financial institution in Indonesia, weekly organic sessions rose 23% within two months of go-live while local Search Console clicks quadrupled; at an Australian Italian-foods retailer, a 27-task local program has traffic up 32% over the trailing four weeks.

Helena checks her own work.

Email is where Helena’s method is most visible, because the pattern on winning accounts is so consistent: audit first, then build, then schedule a check on what you built. At an apparel retailer, the sequence was an initial Klaviyo audit, a rebuild of the weakest flow, then a recurring monthly flow-health audit. Eleven email accounts across the cohort now run recurring audits like that.

A skincare brand shows the full loop on one account, across 99 email tasks. Helena built the flow stack (abandoned cart in brand hex codes, browse abandonment, post-purchase) and then scheduled her own graders: a “Browse Abandonment Flow – 2-Week Check-In” that pulls entries, opens, and clicks per email ID on the exact flow she built, and a “Memorial Day Email Performance Check” that reviews three named campaigns and emails the owner a summary. Nothing ships without a scheduled look back at whether it worked.

Where the loop runs, the numbers move. A sporting-goods brand ran a full “Father’s Day Campaign Post-Mortem” (a dated retro across the ads and emails of a single campaign window), and over its management window, email went from an afterthought to the engine: email’s share of revenue rose from 6.4% to 46.3%. At a digital-products brand, every recurring content task begins with the instruction to read a persistent strategy file (“FIRST ACTION – Read the master strategy file… files/strategy/blog-master-strategy.md”) before executing, and a “Blog Performance Audit – 2-Week Check-In” plus a Monday-morning digest close the loop; email revenue per week is up 21% and conversions up 31%. One caveat we repeat from the underlying analysis: GA4 understates flow-driven email revenue, so these figures are floors.²

The revenue follows the flows. At the skincare brand, $7.9K of email-attributed revenue followed the new flow stack, each flow with its own scheduled check-in; at a stationery brand, email went from nothing to $2,700 a week.

Closing the loop: Helena improves the system that improves the marketing.

Above the channel loops sits a loop about the loops, and this is the step that separates automation from self-improvement. It is also the closest thing in this data to recursion: Helena improving the machinery that improves the marketing. Every account runs a weekly task whose entire job is to review Helena’s own automations and propose one change: a new automation, an update, or a deletion. At the pain-management practice it has run 24 times; at the craft-supplies brand, 12.⁵ At the telehealth company, in mid-July, it noticed that Helena’s own publishing queue was being triggered manually and proposed automating itself out of the bottleneck: the quote at the top of this research.

A companion weekly task, Autonomous Learning, maintains a per-platform learning file (What’s Working, What’s Not, Tried & Ruled Out, Open Experiments), with the standing instruction to build on prior insight rather than starting cold. Helena’s memories function as an experiment log across channels: bad-data quarantines (“do not use May 22–28 as a baseline”), comparability warnings after bid-strategy changes, and human-intent annotations (“paused by the owner on July 1 – intentional, do not flag as needs-review in future reports”). Even channel expansion is a data-gated decision: Bing sits queued at the craft-supplies brand to be switched on “once Google data justifies it.”

And the honest number: 23 of the 50 sampled Meta accounts, and roughly 95 of the 116 email-connected accounts, have wired Helena up but given her little building to do yet. Connection is the on-ramp; activation is the upside. The gap between what the system can do and what customers have switched on is currently the largest untapped source of improvement in the data.

What might the future of marketing work look like?

The evidence suggests the human role is narrowing at each step of the marketing cycle. Once AI-built campaigns and content reach parity with agency work (and on cost and consistency they arguably already have for lean teams), humans stop producing and shift to reviewing. But review has its own ceiling: at the telehealth company’s pace of 900 posts a month, no founder reads everything. The question then shifts from “is this piece good?” to “is the system that produces these pieces pointed at the right goal, with the right guardrails?” Put simply: the doing (writing the ad, building the flow, shipping the report) now costs almost nothing in human time.

An area of human comparative advantage, for now, is marketing taste and judgment: choosing which market matters, which numbers to trust, and when a strategy is a dead end. The customers in this piece exercise exactly that. The craft-supplies brand’s owner paused a Helena campaign on July 1 based on Helena’s own 14-day review; the medical group’s team set the insurance-plan exclusions Helena treats as law. The founders supply the goal; increasingly, they no longer need to supply the method.

What if we’re wrong?

A natural objection to the evidence presented above is that the work still in human hands (choosing what the business should say, to whom, at what price) is what matters most. Without that judgment, Helena is a capable executor, not a system that could drive marketing on its own. That objection is partly right, and worth taking seriously.

Our data has honest limits. Every before-and-after in this piece compares a 90-day pre-window to a management window in which many things changed at once; the deltas reflect correlation, not controlled attribution. GA4 understates email flow revenue, so those figures are floors. One customer’s conversion count is flagged in our own analysis as almost certainly inflated by micro-events, a flag Helena herself raised. And nearly half of connected accounts haven’t activated deeply enough to show anything yet. A standing loop is also only as good as the integration beneath it: at the medical group, the Google Ads connection’s token expired in early June, and the daily campaign report ran 43 times returning only re-authorization requests: the loop kept showing up, but the data pipe stays broken until a human reconnects it.

But marketing is rarely advanced by strokes of creative genius. In between, most progress is incremental: ship something, measure it, fix it, and try again, on the ad account, the search rankings, and the email flows all at once. That is exactly the kind of workflow the system now excels at, and the perspiration is becoming automated. The humans in this piece supply the goal and the judgment: The craft-supplies brand’s owner paused a Helena campaign on July 1 based on Helena’s own 14-day review; the medical group’s team set the insurance-plan exclusions Helena treats as law. Increasingly, they no longer need to supply the method.

Possible futures

What happens next depends on whether the trend continues, and what teams choose to do if it does. We can imagine at least three scenarios:

1. The trend stalls, but today’s capabilities are widely diffused. The channel loops described here may be near their ceiling: execution and optimization automate, but strategic judgment stays stubbornly human. Even so, the diffusion alone changes the industry: one person sits atop a pyramid of standing loops that audit ads, prune keywords, sharpen rankings, and check every flow. The connected-but-dormant accounts in our own data suggest most of this diffusion is still ahead.
2. Marketing systems continue to see compounding efficiency gains. Each channel loop keeps improving the next run, while humans set direction and judge results. Speeding up one part of a process shifts the bottleneck elsewhere: the constraint becomes the human capacity to decide what’s worth testing and to verify what’s working. We have already met this friction: Helena’s weekly suggestion queue generates more improvement ideas than most customers act on. The rate at which a team can spot and clear its own bottlenecks may become its most important marketing skill.
3. The loop closes. If the trends continue, systems like Helena could become capable of redesigning their own strategy end to end: noticing a market shift in the data, re-deriving the positioning, and rebuilding the channel mix, with humans reviewing rather than directing. That would be recursive self-improvement in miniature: not a model retraining itself, but a marketing system rebuilding the system that does the marketing. The early signals exist: Helena already proposes changes to her own automations, quarantines her own bad baselines, and overrules her own prior hypotheses in writing. Whether strategic taste is a capability that AI systems fail at for a time and then get good at is the open question. We would not bet against it.

What should we do?

This research will land differently depending on where you sit. If you are a marketing operator already tinkering with AI, running prompts through a chatbot, trying a copy tool here and an ad tool there, the evidence points to one specific upgrade: stop using AI as a faster keyboard and start giving it a loop. The lift in this study did not come from better prompts. It came from connected accounts, scheduled reviews, and learning files that carry forward what has already been tried and ruled out. The practical guidance falls out of the evidence:

1. Connect the accounts. Whatever AI you use, give it eyes on your real channels: the ad accounts, the analytics, the search data, the email platform. Nothing in this piece happened on an unconnected channel, because an AI that cannot see performance can only guess at it.
2. Let it build, not just draft. The measurable lift concentrated where the AI owned real campaigns, flows, and content end to end. If your AI only produces drafts that a human still has to ship, you are paying for typing speed, not outcomes.
3. Insist on the loop. One-shot output is where most AI marketing experiments stall. Every build should carry its own scheduled review, every review should write down what worked and what was ruled out, and that record should be readable by you.
4. Make the guardrails explicit. Before handing anything over, write down your thresholds, exclusions, and non-negotiables, and hold whatever system you use to them as law. The best accounts in this study did this on day one.
5. Keep humans on the two jobs the data says still belong to them: setting direction, and deciding which results to trust.

And if you are less sure where marketing is headed, whether agents are a real shift or another hype cycle, that skepticism is reasonable, and this piece was written for it. Every number above comes from production systems, not a demo. Marketing that improves itself is no longer a thought experiment. It is running on the ad accounts, the search rankings, and the email flows: on a two-hour timer, in production, right now.

If you are experimenting with AI in your own marketing, unsure where to start, or want to go deeper on the data and methods behind this research, we would like to compare notes. Write to seijin@enrichlabs.ai: questions, pushback, and collaboration ideas are all welcome.

About the research

This research is built from live production data across more than 1,000 active customer accounts running their marketing on Helena, channel deep-dives across the 50 highest-spend Meta, Google, and email accounts sampled from that base, and full reconstructions of every scheduled task, memory, and learning file for four customers: a telehealth company, a craft-supplies e-commerce brand, a pain-management practice, and a multi-location medical group. Performance data is drawn from the Meta Graph API, Google Ads, Google Search Console, GA4, and connected email platforms, as of July 21, 2026.

All quotes attributed to Helena are drawn verbatim from her production task history, memory logs, and weekly learning files, and are used with customer permission. They reflect the state of each account as of July 2026.

* Quotes attributed to Helena throughout this research are drawn verbatim from her production task history, memory logs, and weekly learning files, and are used with customer permission. They reflect the state of each account as of July 2026.

Want to see how Helena can run your marketing? Start your free trial at hirehelena.com.