Essay · Methodology

How to Stop the Slop: The AIG Framework

If AI slop is a dead end, the next question is practical: how do you actually produce the alternative? You need a technical, measurable way to ensure content survives the deduplication and summarization passes of the frontier models.

Originally published on LinkedIn ↗

My last essay made the case that AI slop is doomed. In the age of Generative Engine Optimization, we can't rely on "quality" as a vague vibe. We need a measurable way to know whether our content will earn a citation — before and after we publish. That's what the Active Information Gain (AIG) Framework is for.

What the AIG Framework is

Information gain is Google's patented approach to measuring how much new, attributable information a piece of content adds to an answer. The AIG Framework extends that concept in two ways.

First, it breaks information gain into six specific, measurable dimensions rather than treating it as a single score. Second — and this is why we call it Active — it measures performance at two distinct stages: before publication, when you can still change the content, and after publication, when it's live and indexable. Most content-quality frameworks are static. AIG is a live measurement that follows your content into the wild.

Why the discipline matters: Princeton's GEO study found that adding statistics or expert quotes measurably lifts citation likelihood — statistics were the single strongest lever, improving visibility in AI answers by up to about 40%. Those aren't abstract quality signals; they're direct outputs of scoring well on the dimensions below.

The four pre-publication dimensions

1. Novelty vs. pretraining

Does the content add specific facts, data, or insight the base model can't produce on its own? If an LLM can already generate your entire article accurately before reading it, your novelty score is zero. The most common failure I see when scoring a new client's library for the first time: pieces that feel original but are functionally indistinguishable from what the model already knows.

2. Distinctiveness vs. competitors

LLMs often retrieve 10–20 sources for a single query. Does your content offer claims or perspectives the other top-retrieved sources miss? If you're saying what everyone else is saying, you'll be deduplicated out of the response — the engine has no reason to cite you when five other sources say the same thing.

3. Entity exclusivity

How many named entities appear in your content that appear nowhere else in the retrieved set — coined terms, proprietary statistics, unique frameworks, specific people? If a named concept is exclusive to you, the LLM has nowhere else to point but back to your brand. This is one of the highest-leverage dimensions, because a single editorial decision — coining a term and defining it precisely — can move the score significantly.

4. Definitional density

Generative engines have a strong bias toward explicit "X is Y" structures. We measure the ratio of clear definition sentences per thousand words. High definitional density makes your content immediately usable for an engine explaining a concept — and gives it a distinct, excerptable surface competitors' content often lacks.

The two post-publication dimensions

These only become measurable once your content is live and indexable. They tell you whether the information gain you engineered is actually surviving contact with real LLM retrieval — and where to intervene if it isn't.

5. Attribution survivability

Does your content have enough unique signal to survive an LLM's deduplication pass? High-scoring content is so specific that the model must cite, quote, or name you to keep its answer accurate. One client had a piece ranking well organically but earning zero AI mentions; attribution survivability was the culprit — nothing forced the model to name them. The fix was adding three coined terms and two proprietary data points, after which mentions appeared within the next crawl cycle.

6. Position weight

Not all recognition is equal. A first-paragraph mention carries far more trust and click-through value than a buried citation. We measure where in the response your content surfaces and whether it's named as a primary source or absorbed anonymously. We engineer for it by placing the highest-gain insight at the very top: the claim that most requires attribution goes in paragraph one, not the conclusion.

What you can do today, and what needs a tool

Two dimensions are eyeball-able with no infrastructure. For entity exclusivity, scan your draft and count the coined terms, proprietary statistics, and named frameworks a competitor wouldn't have. For definitional density, count your explicit "X is Y" sentences per thousand words. Both give an immediate read on whether a draft is pulling its weight.

The rest is heavier lifting — running your claims against a base model with no retrieval, decomposing the competitor sources the model actually retrieves, then querying live retrieval systems to track where and whether you surface. Done manually, the full six-dimension workflow runs 30–45 minutes per piece per query. Most teams won't do it consistently, and the ones who would won't do it at scale. That's exactly why we're building a tool to automate it.

Over the coming weeks I'm publishing one deep-dive per dimension — what it measures, why it matters, and how to engineer for it with real client examples.