Danny Sullivan: Stop Chunking Content to Rank in LLMs

• Key Takeaways:
• Google’s Danny Sullivan warns creators not to fragment content into “bite-sized chunks” to game LLM rankings.
• He told Search Off the Record engineers and listeners that this approach may work short-term but won’t endure.
• The advice: write for users, not for specific ranking systems; systems will evolve to favor human-focused content.

What Google said on the podcast

On the Search Off the Record podcast, Google public liaison Danny Sullivan addressed a trend he’s seeing among content teams: restructuring pages into tiny, standalone pieces because some large language model (LLM) systems appear to reward bite-sized inputs.

“We don't want you to do that,” Sullivan said, adding he had discussed the issue with Google engineers. He warned that while fragmenting content might yield short-term gains in some systems, it’s not a sustainable strategy for Search.

Why chunking looks tempting — and risky

Marketers and SEO teams are experimenting with concise, discrete content blocks to increase the chance that LLMs will pull and surface their text. The idea: smaller units are easier for models to ingest and repurpose.

Sullivan cautioned this can create two problems. First, creators may end up optimizing for a particular ranking behavior rather than for human readers. Second, ranking systems will likely evolve, eventually prioritizing content written for people — making the earlier investment obsolete.

What this means for SEO and content teams

The takeaway for SEOs and content strategists is straightforward: focus on user intent and clarity, not tactics aimed solely at current model quirks. Sullivan framed the issue as evergreen SEO advice updated for the era of AEO/GEO/LLM-driven features.

If you’ve seen a temporary boost from chunked pieces, Sullivan warned that depends on “edge cases” and transient system behavior. When models and search systems improve, they will favor comprehensive, human-centric content.

Practical guidance for creators

Prioritize readability, usefulness, and thorough answers to user questions. Keep structure logical, use headings and clear sections, and avoid creating duplicate or narrowly tailored versions solely to target LLM outputs.

Measure success by user engagement and long-term traffic trends rather than short-lived gains from tinkering to match a particular model’s current behavior.

For context, listen to the episode of Search Off the Record referenced by Sullivan for the full conversation.