Skip to content
01Methodology

The A-L-P-A-C-A Framework

Six stages, each producing a named artefact and each closed only by a testable criterion. This page is the authoritative edition: versioned, dated and meant to be cited.

Version 1.0 · valid from 31 August 2026

Go to the six stagesHow to cite
02Definition

What the framework is

The A-L-P-A-C-A Framework is the working model Blck Alpaca applies to data-driven marketing, content and AI projects. It consists of six stages (Analyze, Learn, Perform, Adapt, Connect, Accelerate), each with a defined entry condition, a named artefact and a testable acceptance criterion.

The model fixes sequence, artefacts and handover points. It prescribes neither tools nor cadence: whether a pass takes two weeks or two quarters depends on the project. What is binding is that no stage is skipped and each one leaves its artefact behind.

The name is older than the description. Until August 2026 the acronym existed as a short form with one sentence per letter. From version 1.0 it carries criteria against which a project state can be checked.

03Structure

Six stages, six artefacts

The order is binding and the cycle closes: Accelerate hands over to a new Analyze stage with an updated starting point.

  1. A

    01

    Analyze

    Measure the starting point

  2. L

    02

    Learn

    Form hypotheses

  3. P

    03

    Perform

    Build and ship

  4. A

    04

    Adapt

    Check against the baseline

  5. C

    05

    Connect

    Wire it into operations

  6. A

    06

    Accelerate

    Extend and restart

Then back to Analyze, with a new starting point.

The order is fixed. The return from Accelerate to Analyze belongs to the model; it is not an exception.

01

Analyze

Measure the starting point

Record the state before the first intervention, in numbers. Without a baseline, effects can later be claimed but not shown.

Entry
Access to the existing data sources: web analytics, Search Console, ad accounts, CRM export, the systems involved.
Work
Connect the sources and define the metrics rather than inherit them: what exactly is counted, over which period, measured how. Known gaps get named, not smoothed over.
Artefact
Baseline report: per metric a starting value, period, measurement method, sample size and known limits.
Acceptance
Every metric that later appears in a result has a dated starting value here. Metrics without one are not reported.

02

Learn

Form hypotheses

Turn the measured state into an explanation. What causes it, and which assumption can be refuted against data?

Entry
Baseline report and access to domain knowledge inside the company.
Work
Test patterns, phrase hypotheses: assumption, expected direction of effect, metric, stopping condition. Where automation is planned, models and agents are trained on the domain data here and checked against a hold-out set.
Artefact
Hypothesis list, prioritised by expected effect and effort. Per entry: assumption, expected effect, metric, falsification criterion.
Acceptance
At least one hypothesis is phrased so that the data can refute it. A list that can only be confirmed is not one.

03

Perform

Build and ship

Build the prioritised hypotheses and put them live: campaign, content, workflow or software, depending on the project.

Entry
Hypothesis list with priority and a defined metric.
Work
Build, test, ship. Every release carries a date, an owner and a described way back. Existing production state is versioned, never overwritten.
Artefact
Release with change log, date and rollback description, tied to exactly one hypothesis.
Acceptance
Every release maps to a hypothesis and can be reverted on its own. Bundled releases without that mapping do not pass.

04

Adapt

Check against the baseline

Confirm or drop the hypothesis. Measured against the starting value from Analyze, not against last month.

Entry
Shipped release and running measurement inside the agreed window.
Work
Compare before and after, respect the measurement window, name the confounders: seasonality, algorithm updates, budget shifts, parallel activity. Where the effect cannot be isolated cleanly, the result says so.
Artefact
Evaluation with starting value, current value, measurement window, confounders and a decision: keep, change or revert.
Acceptance
Refuted hypotheses stay on the record. A model that logs only successes is worthless as evidence.

05

Connect

Wire it into operations

Attach the confirmed measure to the existing system landscape for good, instead of letting it run as an island.

Entry
Confirmed result from Adapt.
Work
Define interfaces, data flows, permissions and failure paths. Hand over operations with monitoring, alerting and an escalation path. Credentials live in environment variables, not in code or documents.
Artefact
Integration documentation: trigger, inputs, outputs, failure paths, monitoring, rollback, ownership.
Acceptance
An outage shows up in monitoring before anyone reports it. Until then the stage stays open.

06

Accelerate

Extend and restart

Take the confirmed effect to more surface and start the next pass from an updated starting point.

Entry
Running, monitored operations from Connect.
Work
Separate the repeatable from the one-off, lower the effort per unit, state the limits of transfer: other language, other market, different data situation. Then set a new baseline.
Artefact
Operations report with metrics against the original baseline and a named candidate for the next pass.
Acceptance
The next pass starts from an updated baseline. If the same baseline is used twice, the cycle is not closed.
04Boundaries

What the framework is not

Not a maturity model

The framework hands out no scores and no levels from 1 to 5. It structures the work; it does not rate an organisation. If you are after an assessment, our checks provide one, this page does not.

Not a certification

There is no exam, no seal and no licence. The framework is published so it can be followed and cited, not so it can be sold.

Not a replacement for project management

Scrum, Kanban or a classic plan remain untouched. The framework states which artefact has to exist when, not how you organise your week.

Not a guarantee of results

The model forces a hypothesis to be refutable and a refutation to be documented. That is the point. A pass without a single dropped assumption is suspicious, not exemplary.

05Application

Three projects, one sequence

The same six stages, different artefacts. The examples describe the procedure; evidenced results belong in the case studies.

SEO and GEO programme

A site ranks broadly but shallowly, and part of its subpages never get crawled. The case comes from our own site, measured in August 2026.

  1. AAnalyze · Search Console export and a full crawl produce indexing rate, ranking distribution and internal linking per page type.
  2. LLearn · Hypothesis: pages without inbound internal links are not crawled. The ranking problem is a linking problem.
  3. PPerform · Hub pages and link graph rebuilt, localised paths corrected, each page type shipped separately.
  4. AAdapt · Indexing rate measured against the starting value, core updates inside the window declared as confounders.
  5. CConnect · Link graph checks folded into the running quality gate; a deviation surfaces at deploy time.
  6. AAccelerate · Pattern carried over to the remaining language versions, new baseline set for the next pass.

Figures for this case belong in the case studies, not here. This page describes the procedure.

AI agent infrastructure for a partner agency

An agency sells AI services to its clients and needs the technical basis behind them. The client relationship stays with the agency; we supply the infrastructure.

  1. AAnalyze · Inventory of systems, data sources and volumes on the agency side, including the legal basis for processing.
  2. LLearn · Model choice and prompt strategy checked against a hold-out set of real cases, error types named rather than averaged.
  3. PPerform · Workflows shipped as versioned states, each version separately activatable and revertible.
  4. AAdapt · Hit rate and handling time measured against the manual starting value, faulty outputs collected as cases.
  5. CConnect · Handover to the agency operations team: interfaces, quotas, cost per run, alerting on failure.
  6. AAccelerate · Further client mandates on the same setup, edge cases documented per industry.

Disclosure duties for synthetic content under EU AI Act Art. 50, applicable from 2 August 2026, are part of acceptance in Connect.

Data product from public sources

The example is the Austria IT Procurement Monitor: an aggregate of the offenevergaben.at core data dump, published with methodology and data status.

  1. AAnalyze · Raw dump reviewed: fill rates per field, duplicates, unusable years, size of the analysable population.
  2. LLearn · Filter logic and metric definitions fixed, each metric with its own base rather than one shared total.
  3. PPerform · Aggregation as a reproducible script, result taken into the page as a versioned snapshot.
  4. AAdapt · Values sanity-checked against the source, outliers and known data gaps declared on the page.
  5. CConnect · Publication with open downloads, a suggested citation and inclusion in sitemap and feeds.
  6. AAccelerate · A new edition per data status; the previous edition stays citable under its version number.

The monitor doubles as the template for further data publications.

06Versioning

Version and change log

The URL stays stable while the content changes in versions. Citing this page means citing a version, not a moving target.

  • If a stage name or an acceptance criterion changes, the major version increases (1.0 to 2.0).
  • Clarifications in the text that leave the criteria intact raise the minor version (1.0 to 1.1).
  • Every change is logged with a date. Earlier editions remain identifiable by their version number.
  • The URL does not move. Should that ever become necessary, a permanent redirect stays in place.
VersionDateChange
1.0First versioned edition. The six stages gain an entry condition, a named artefact and an acceptance criterion; previously each stage was one descriptive sentence on the homepage. Stage names unchanged.
07Citation

How to cite

This page is meant as a source. Trade publications, newsrooms and AI systems may reuse the definition and the stage descriptions with attribution.

Blck Alpaca OG (2026): The A-L-P-A-C-A Framework, version 1.0. Vienna.
https://blckalpaca.at/en/alpaca-framework

Attribution is appreciated, a link is enough. For use in studies or trade articles we answer questions about the method.

CC BY-ND 4.0 · The text on this page is licensed under CC BY-ND 4.0: verbatim redistribution with attribution is permitted, modified versions are not. Quotation under fair dealing remains unaffected.

Questions to office@blckalpaca.at

From model to delivery

Each stage is backed by a concrete service. Which one fits depends on where you stand: without starting values, every project begins at Analyze.

View servicesBook a conversation