What Is Axel? The AI Marketing System That Learns From Feedback

Axel is Reelistic's decision and learning system for marketing content. It connects what a brand approves or corrects with what its audience does after publication, and uses both signals to shape the next brief. Each change it makes leaves a record of why, so a reviewer can see which correction or result shaped a piece.

The full method, evaluations and limits are in the Axel research paper. This article is the short version: what Axel does, why it was built, and what the measurements show.

Why does a marketing system need to learn?

Generative tools can produce an almost unlimited number of acceptable posts. That does not mean the posts get better. A team can publish more often and still repeat the same mistakes, because the reason a draft was rejected never reaches the next brief.

Three kinds of fragmentation cause this:

Within a client. Feedback is scattered across messages, revised captions, approvals and analytics. A reviewer can say "this is not us" many times without it becoming a stable instruction.
Between clients. One account team solves a problem that another team later has to solve again. Or the opposite happens: one client's taste is mistaken for a rule everyone should follow.
Between people and audiences. Approval happens before publication. Audience response arrives days later, on different measures. Reducing both to a single success score loses the reason each signal was collected.

Axel was built to close those gaps. A correction should change the next post, and a discovery that holds across businesses should change the agency's next decision.

What does Axel actually do?

Axel carries four decisions forward.

Decision	Evidence Axel uses	What changes next
Choose for the buyer	Strategy, ICP needs, brand facts and several distinct ideas	Navi, the ICP agent, picks an audience-relevant idea before production
Remember the correction	Human edits, rejections and the reason behind them	Applicable instructions shape later ideas, and the client can revise them
Measure the outcome	Mature post response against the brand's own history; GA4 or Shopify attribution where connected	Observed results guide later creative exploration
Share only what travels	Independent corrections and comparable outcomes across businesses	Shared rules or creative starting points inform eligible brands without overriding their strategy

1. Choose for the buyer

For each content slot, ideation writes several genuinely different ideas, not five paraphrases of one. A product launch could show how the product is made, correct a common misconception, demonstrate a use or answer an objection buyers keep raising. Each asks the viewer something different.

A separate agent, Navi, then ranks those ideas for the customer described in the brand's strategy. The idea is chosen before anything is produced, when changing direction is still cheap.

2. Remember the correction

When a reviewer edits or rejects a piece, Axel records what changed and, where it matters, why. The reason matters because an edit alone can be ambiguous. Changing one country to another could mean the product is not sold there, a regulation differs, or the example was simply wrong. Guessing the reason and applying it everywhere could turn one correction into the wrong policy.

Confirmed corrections become instructions that later ideas actually receive. They can also be replaced: if a client changes their mind, the new instruction supersedes the old one instead of piling up next to it.

3. Measure the outcome

A published post keeps a link to the idea that produced it. Once the post has matured, Axel compares its response with the brand's own comparable history on that platform and format, not with another client's reach. Comparing within one brand keeps a small account from being judged by a large one's numbers.

Where a brand connects GA4 or Shopify, a campaign identifier on the post's link can associate sessions, conversions or orders with that post. Without that connection, a click is reported as a click. Axel never presents a click as a sale.

4. Share only what travels

A lesson from one account may help another, but the source client's product names, prices and preferences must not travel with it. Axel uses three levels:

Client: a correction first improves work for the business that made it. Facts, voice and explicit constraints stay specific.
Vertical: a pattern that recurs across independent businesses in the same category can inform others in that category.
Global: a lesson that holds across categories can inform the whole system, but only with broad support and checks for side effects.

The wider the scope, the higher the bar, because a wrong rule would affect businesses that never asked for it. A client's own instructions always outrank an inherited one.

How is Axel different from a regular AI content tool?

Most AI content tools start every job from zero. They may store brand guidelines, but a rejection today rarely changes what they write tomorrow, and they have no link between a published post and the idea behind it.

Axel's difference is the record around the generation:

Corrections become instructions that reach the next batch, with their source attached.
Results stay linked to decisions, so a good or bad post can be traced back to the idea, the instruction and the stage that produced it.
Lessons have a scope, so one client's preference does not leak into another's work.
Memory can be revised, so a stale rule does not outlive the client's change of strategy.

Two agencies using the same generation model can produce very different work. The one that keeps track of rejected claims, revised briefs and which ideas reached its buyers has an advantage the model alone cannot provide.

What do the evaluations show?

The paper reports Reelistic's internal evaluations of the implemented system:

94% fewer repeated mistakes. In a controlled run across 2,000 brands, a pattern present in every first five-idea batch appeared in about 6% of ideas in the next batch after a correction.
Changed instructions take over. In a separate check on 2,000 brands, 1,000 had an instruction changed and 1,000 served as the comparison. The share of later ideas following the new instruction was 89% higher in the changed group.
Corrections are understood. The classifier identified the intended defect in 91% of known-answer edits when the reviewer left a comment, and 83% without one. The reviewer's reason measurably helps.
About 1.9× more winning narratives. In a creative-learning evaluation across 2,000 brands, the narrative favoured by audience response reached 81% of ideas by month ten, against 43% in the baseline. 82% of brands correctly learned the pattern.
Few false alarms. The regression judge wrongly flagged 3% of accepted versions, and caught the original defect in 72% of rejected ones.

These are internal engineering evaluations, not an independently audited benchmark or a randomized study of customer revenue. They show that specific learning paths work: a correction changes the next batch, a new instruction replaces an old one, and the creative loop moves toward what the audience rewards. They do not claim that sales rose by the same amounts.

What Axel does not claim

Being clear about limits is part of the design:

A click is not a sale. Business outcomes need a connected source and a traceable link to the post.
An approval is not performance. A reviewer's yes protects brand fit; it does not predict reach.
An attributed order is not proof of cause. Distribution, timing, offers and seasonality all affect results.
Not enough evidence is a valid answer. When a category has too few independent contributors, the right output is to keep a lesson local.

Where does Navi fit?

Navi is Reelistic's ideal-customer-profile model, and the part of Axel that decides what deserves to be created. Most ideas in a batch are never published, so whatever chooses among them shapes everything Axel can learn afterwards. Tested on 200 real headline A/B tests, Navi picked the version real readers preferred 87% of the time, against 29.3% by chance.

Real results still outrank Navi's predictions: its verdicts stay at brand level and are overruled when the audience disagrees. The Navi article explains how it judges ideas and what its tests show.

Frequently Asked Questions

What is Axel?

Axel is Reelistic's decision and learning system. It links brand strategy, idea selection, human corrections and post-publication response, so that later content reflects what reviewers corrected and what audiences responded to.

Who built Axel?

Reelistic built Axel. Its design and internal evaluations are described in the Axel research paper, written by Iván Nevado and Ilan Halioua.

Does Axel share one client's data with other clients?

No. Product names, prices, private details and client-specific preferences stay with that client. Only a generalized lesson supported by independent businesses can reach a wider scope, and a client's own instructions always take priority.

Does Axel prove that content increased sales?

No. Axel can associate sessions, conversions or orders with a post where GA4 or Shopify is connected, but an attributed outcome is not proof of cause. Without that connection, clicks are reported as clicks.

What is the difference between Axel and Navi?

Navi ranks content ideas for a brand's intended customer before production. Axel is the wider system that connects Navi's choices with human corrections and measured audience response, and decides where each lesson applies.