Inside Backbase's blog agent: the exact workflow behind 100% organic traffic growth

Why consistency beats creativity in AI content

An LLM's job is to predict the next most probable word in a sentence. Train it on a confined knowledge base with a clear rule set, writing examples, and tone-of-voice guidelines, and that same prediction engine becomes a brand consistency machine. As the speaker put it: "AI sticks to brand regardless of what your human does at the end of the day. And if humans just edit, there's way less brand skew that actually happens."

You give the model the right inputs, constrain the environment, and it will reproduce your brand's voice far more reliably than a rotating cast of writers ever could. The human's job shifts from production to quality control - and that shift is where the time savings compound.

What content marketers were actually doing in 2023

In 2023, the average content marketer spent roughly 15% of their time on strategy, 70% on what the speaker called groundwork, and 15% on quality assurance. That 70% in the middle - copy-pasting from Google Docs into a CMS, finding data, building reports, loading images - was never high-value work. It was just necessary. Loading a single blog post into Webflow or any decent CMS took 15 to 20 minutes per person per day. Multiply that across a team and a month, and you are looking at a serious chunk of productive time going to pure grunt work.

Agents handle the grunt work, leaving the thinking to humans. And once that 70% is automated, the team's time distribution flips to something closer to 40% strategy, 30% groundwork, and 40% quality control - more creative, more considered, and with far fewer excuses for mediocre output.

The blog agent, step by step

The Backbase blog agent is a multi-step agentic agentic content workflow that runs from keyword gap to near-final draft without a human touching the keyboard.

Step 1: keyword and gap identification

The process starts with a search gap - a topic the brand wants to rank for. Say the team identifies "conversational banking" as an opportunity. That becomes the brief input. Simple, one-line, human decision.

Step 2: SERP analysis

The agent pulls the top five ranking blogs for the keyword and analyses what they are saying. It builds a structured read of the competitive content space for that topic.

Step 3: Google AI snippet analysis

This step runs through an API three times. It analyses what Google is surfacing as the general answer for the topic - the AI Overview, the featured snippets, the implied intent. It also scrapes around 30 links that show up in those results, building a picture of what Google currently rewards for this search. All of that gets saved into a database.

Step 4: knowledge base synthesis

An agent takes everything gathered in steps two and three and writes it into an internal knowledge base entry for the topic. This becomes the competitive context layer the writing agent draws on later.

Step 5: proprietary angle identification

Here is where the workflow separates from anything a generic AI writing tool does. The agent looks at the brand's own knowledge base and asks: what proprietary angle do we have here? Do we have a data point, a point of view, a transcript of someone speaking about this? It goes looking for source material - and if there is a relevant transcript, it pulls that in. The brief it produces specifies the topic, the competitive context, the internal source material, and the data to use. The non-commodity content angle is baked in at brief stage, before a single word of the draft is written.

Step 6: draft writing

The agent writes the full draft based on the brief. The agent writes from a structured brief that includes competitive context, brand knowledge, source transcripts, and data points. The output is grounded before it starts.

Step 7: internal linking

The workflow runs internal linking based on the brand's cluster strategy: three evergreen blogs, three new blogs, and two product pages relevant to the topic. These are not suggested links - they are embedded by the agent according to a predefined content cluster logic.

Step 8: external linking via live scraping

External links are handled through live scraping, not from memory. The agent finds the page, reads something on it, and only then uses it as a source. This eliminates hallucinated URLs and dead links entirely. No 404s, no made-up citations.

Step 9: slop clean and tone-of-voice check

Before the final output, the draft runs through what the speaker called a "slop clean" - a tone-of-voice framework that eliminates certain writing patterns, double-checks brand voice, and applies what they call RAD rules. This is the brand consistency layer, automated. For anyone building a system to prevent AI slop, study this step - it is where brand consistency is automated.

Step 10: final output with H1

The H1 is written last, once the agent knows what the full piece actually says. The final output follows a specific template and arrives with the headline, body, links, and tone all checked. Total runtime: around 12 minutes. Time needed to edit and publish: about five minutes.

What the knowledge base actually contains

The workflow runs on a knowledge base that holds everything about the brand - rules, tone of voice, writing style, strategic memos, product marketing, all processes, all data from Google Analytics, Search Console, Salesforce, LinkedIn, and the brand's signal engines. Transcripts, research, industry news. The speaker described it as a GTM operating environment: "We control the knowledge base, everything about the brand, rules, the tone of voice, the writing style, the strategic memos, the product marketing, all our processes, all our rules, all our data from different sources."

Most teams skip this when building something similar. They focus on the agent steps and ignore the knowledge base quality. A well-structured brand knowledge base for AI produces consistent, on-brand output. The agent is only as good as what it is drawing from.

The org structure that makes it work

The workflow does not exist in isolation - it sits inside a deliberately restructured team. The speaker directs content strategy alongside the CMO and CEO: where the brand is going, what it is speaking about, how to translate business narrative into stories. That responsibility stays human. What changed is the content engineering layer underneath it - the systems, the SOPs, the templates, the agent builds. As the speaker put it: "I'm constantly building so they don't have to copy paste something into a CMS."

The team runs one system. The content lead runs the build environment behind it. Non-technical team members execute through the platform without touching the agent infrastructure. Pay attention to this if you are transitioning from a traditional content team to content engineering - this is a real organisational model in production. Read more on that transition here.

The numbers, plainly stated

Pipeline up 85% year over year. Average deal size up 60%. Marketing budget down 25%. Organic traffic up 100% year over year. The team runs 75 to 80% of all their work inside this system. They are dropping around 100 tools they no longer need because the middleware is gone. No AI-related layoffs. Eight months to get there.

These are reported results from a team that shipped this in production. The workflow was not built overnight, and the knowledge base did not build itself - but the compounding effect of consistent, on-brand, well-researched content at volume is exactly what those numbers reflect.

What this means for the human in the loop

The speaker's closing argument was direct: "A good writer is the difference between slop and grit. This is the person that does the final mile and the first mile. The better they are, the better everything else gets." The workflow raises the stakes for good writers. When the grunt work is automated, there are no more excuses for mediocre. The quality bar goes up because the time excuse disappears.

A well-built AI content workflow automation system gives your people more time to think and more time to plan. That is what shows in the work.

Frequently asked questions

How long does it take to build a blog agent workflow like Backbase's?

The Backbase workflow took around eight months to reach the point where it was producing the results shared in the keynote. That timeline includes building the knowledge base, designing the agent steps, integrating data sources, and refining the tone-of-voice checking layer. The build itself was done by a combination of the content lead working in ClaudeCode and a team of AI developers maintaining the platform infrastructure.

Does automating blog content hurt SEO rankings?

The data from Backbase suggests the opposite - organic traffic doubled in eight months. The key is that the workflow is designed to produce genuinely useful, well-researched, on-brand content, not to mass-produce thin posts. The SERP analysis, AI snippet analysis, proprietary angle identification, and slop clean all exist to make sure the output competes on quality, not just volume.

What is a "slop clean" and why does it matter?

The slop clean is a tone-of-voice pass that runs before final output. It eliminates certain writing patterns the brand wants to avoid, checks that the voice matches brand guidelines, and applies a set of editorial rules (called RAD rules in the Backbase system). It is the automated equivalent of a senior editor's eye on the draft, run consistently on every single piece without human effort.

Can a small team or solo operator build something similar?

The Backbase system was built with access to serious AI developers and a well-resourced team. For smaller operators, the underlying logic - confined knowledge base, structured brief, sequential agent steps, tone-of-voice checking - is replicable at a smaller scale. The honest constraint is the build time and technical skill required to do it from scratch, which is exactly why pre-built agentic content platforms exist for teams who cannot afford to build it themselves.

What should go into the knowledge base before running a blog agent?

Brand rules, tone-of-voice guidelines, writing examples, strategic memos, product positioning, and any transcripts or data points that represent the brand's proprietary point of view. The speaker was clear on this: start clean, pick the ten most impactful pieces, and build from there. A messy knowledge base produces messy output. The inputs that matter most are the ones that carry your actual point of view - proprietary data, real transcripts, and specific strategic positions your brand holds that no competitor can replicate.