Why We Built GrafStack

AI answer engines cut search traffic to publishers. They use publisher content without credit or pay. They flood the web with synthetic content, which erodes reader trust.

The market has built tools for each problem. Most of those tools treat the URL as the smallest unit of content. AI answers quote a sentence, not a page. GrafStack works in the gap between those two scales.

Where this started

I built the first version as a capstone project for an AI Products and Services course. I called it NewsAtom. The pitch was an AI-native CMS, focusing on breaking content down into trackable, discrete "atoms." The pitch was wrong.

Newsrooms already pay for Sanity, Contentstack, Arc XP, Brightspot, or WordPress VIP. They do not need another CMS. They need a method for structuring content inside the system they own.

So I changed the product. GrafStack is a newsroom methodology and reference architecture. You run it on top of your current CMS.

What the research showed

I mapped the market into seven layers: content structure, provenance, rights expression, transport and monetization, direct deals, legal action, and agent discovery. Then I placed each product by the size of content it handles: URL, asset, section, paragraph, or atom.

Three findings stand out:

• Rights and monetization tools cluster at the URL. RSL, extended robots.txt, IAB CoMP, TollBit, and Cloudflare Pay-Per-Crawl all work at page level.
• Few tools reach paragraph or atom level. ProRata, Google's pay-per-value pilot, CMS-hosted MCP servers, and Sanity's Portable Text are the main examples.
• Money flows to the top. Tier 1 publishers earn $20 million to $250 million a year in direct deals. Regional and independent outlets earn close to $0.

Page-level pricing also misses value. Pay-Per-Crawl bills for a fetch. It does not measure what your content is worth inside an LLM's context window.

If your content has no structure below the page, you cannot price it, track it, or correct it at the level where AI systems use it.

A name change

Sannuta Raghu published a News Atom schema through the Reuters Institute in October 2025. Her approach to wrapping content is close to mine. Using her name would confuse readers and bury her work.

We renamed the project. "Graf" is newsroom shorthand for paragraph. "Stack" refers to the seven layers. If you work on structured news content, read her work.

What we offer

The GrafStack AI Editorial Workflow Audit runs four weeks. We review how your newsroom writes, structures, tags, corrects, and publishes content. We score each layer of your stack against a rubric. You get a report and a reference architecture sized to your CMS and your team.

Design choices so far

• The sentence is the core unit. Images, data, and other media need separate units.
• An LLM proposes tags. An editor reviews and approves each one.
• A correction triggers a recheck. We version the change at the graf or sentence level, depending on the size of the correction.

These choices are open to challenge. We expect editors to push back, and we will record what changes.

What we want to do next

• Publish the methodology and research openly at grafstack.com.
• Publish a crawler information page. When our crawler visits your site, you will see why.
• Serve regional and mid-market publishers first. They sit outside direct licensing deals and carry the most risk.
• Work with CMS vendors, so the architecture fits the platforms newsrooms already use.
• Collect feedback from editors and engineers, then revise the methodology in public.

Work with us

If you run a newsroom, product team, or engineering group at a publisher, send me a note through Bite Size Elephant. Tell me which CMS you use and where AI traffic hurts you most.