Essay · GEO Strategy

AI Slop Is Doomed. Here's Why.

Your content strategy comes down to a simple choice: chase "GEO at scale" for a short-lived spike, or build genuine information gain and compound your advantage as the models tighten.

Originally published on LinkedIn ↗

We're in the slopacolypse. The internet is filling with AI-generated content faster than it's filling with anything else, and most marketing teams in 2026 are pouring fuel on the fire — spinning up five, ten, fifteen articles a week with generic AI workflows. Content-farm agencies promise "five expert articles a week, all original, all on-brand," running the same prompt under the hood.

That is the exact strategy that's about to stop working. Here's the case, in three reasons.

Reason 1 — The models are being tuned to reward new information

Frontier labs have a data problem. As models are recursively trained on AI-generated text that merely paraphrases existing training data, they lose the edge-case nuance and factual precision that define high-quality intelligence. This is a documented failure mode, not a hunch: Shumailov et al., writing in Nature in 2024, showed that generative models trained on recursively generated content suffer model collapse — the tails of the distribution disappear, novelty falls, and outputs homogenize.

The people building these systems have been unusually candid about the fix. The consensus from the leaders of OpenAI, Anthropic, Google DeepMind, and NVIDIA is that the next phase of AI depends on better data, not just bigger models — OpenAI's Sam Altman has publicly questioned whether generating vast amounts of synthetic data and feeding it back in is the right path, reframing the real challenge as how to learn more from less data. When the people writing the algorithms warn against flooding models with low-novelty content, the brands shipping exactly that are betting against the house.

To compensate for a training corpus increasingly dominated by synthetic noise, retrieval and citation systems are being tuned harder and harder toward novelty. Content that contributes nothing distinctive is the first thing those systems learn to ignore.

Reason 2 — Traffic from AI slop is short-lived

The pattern for "GEO at scale" is the same every time: traffic climbs, sometimes sharply, then falls back toward where it started — or below it.

Lily Ray, of the SEO consultancy Algorythmic (and VP of SEO & AI Search at Amsive), analyzed 220+ websites publicly identified as customers of AI content-scaling platforms. The shape was consistent: organic pages grow fast; organic traffic peaks within a few months of that content peak; then traffic declines steeply, erasing most of the gain — and frequently dropping below the prior baseline — within the following year. Tracking dozens of sites producing pure AI-generated content since 2023, Thomas Peham of Otterly.AI observed the same thing: no sustainable long-term gain.

Existing slop may keep getting cited for a while — it landed in the indices and training corpora before the filters tightened. But every new piece produced from here is a worsening bet.

Reason 3 — Information gain is proven to work better

Information gain measures how much new, attributable information a piece of content adds to an LLM's answer — beyond what the model already knows from pretraining, and beyond what competing retrieved sources already offer. The higher a piece scores, the more likely it is to be quoted, named, or cited.

This isn't just intuition; it's measured. Princeton's Generative Engine Optimization study (Aggarwal et al., KDD 2024) tested nine content strategies across roughly 10,000 queries and found that adding statistics, quotations, and citations were the top-performing methods — boosting visibility in AI answers by up to about 40%, with statistics the single strongest lever. At Profound's Zero Click event, iPullRank's Mike King presented data mapping an information-gain score against two outcomes at once — AI citations and organic search rank — and both rose together as information gain rose.

More unique content earns more AI citations — and better rank. It would be redundant for an engine to cite five sources that all say the same thing.

The evolution: the Active Information Gain (AIG) Framework

Google frames information gain narrowly, in the context of a search patent designed to reduce redundancy. The move to GEO requires a broader, more granular definition — so I developed the Active Information Gain (AIG) Framework: a methodology that measures content against six technical dimensions, the "survival score" of your content in an AI-gated world.

• Novelty vs. pretraining
• Distinctiveness vs. competitors
• Attribution survivability
• Position weight
• Entity exclusivity
• Definitional density

I break each of these down — and show how to engineer for it — in a companion piece: How to Stop the Slop: The AIG Framework.

What to do right now

Scaling slop is easy; scaling information gain is hard, because it depends on human input. Six concrete moves:

1. Stop summarizing the internet. If an LLM can already generate your entire article accurately without reading it, your novelty score is zero. Shift to edge cases and internal data.
2. Lead with expert opinion. AI trends toward the median because it's afraid to be wrong. Human experts can take positions — and those positions are what engines crave to differentiate an answer.
3. Coin terms, and define them. If a named concept appears only in your content among the retrieved set, any answer that uses the term has nowhere to point but back to you.
4. Use formats that resist compression. Comparison tables, decision trees, and proprietary diagrams create a new attribution surface that engines disproportionately lift.
5. Cite primary research with year and source. As the Princeton data shows, statistics and quotes are citation magnets. Anchor every claim to a specific, verifiable entity.
6. Score before you publish. Run every draft against the six AIG dimensions. The discipline forces the originality that makes content indispensable.

The divergence, in one table

Feature	Synthetic Noise	Information Gain
Source material	LLM-paraphrased training data	Primary research, new ideas
Entity density	Low; relies on common terms	High; exclusive terms & stats
Engine treatment	Deduplication: brand is stripped	Citation: source is attributed
Competitive moat	Zero; easily replicated	High; creates unique novelty
User experience	Skim-and-forget	Save-and-cite
Long-term outcome	Model collapse	Compounding authority

The brands building information gain into their content now compound their advantage as filtering tightens. The rest are training tomorrow's models to ignore them.