For three decades, "search engine optimization" meant one thing: convince an algorithm to put your link higher on a ranked list of ten blue ones. The game was discrete. The optimization surface was small. The winners were known.
That game is ending. The generative engines — ChatGPT, Perplexity, Claude, Google's AI Overviews — do not return ranked lists. They return synthesized prose, and silently choose which sources to attribute. A site can rank #1 on Google for a query and not appear, anywhere, in the answer Perplexity gives for the same query. Position-based SEO is necessary and no longer sufficient.
We've been calling the new game Generative Engine Optimization, or GEO. It is genuinely a different game, and the optimization surface — what to change on your site — looks different too. Here's what we've learned across six months of testing on twenty-two client sites in legal, e-commerce, B2B SaaS, and content publishing.
What the generative engines actually reward
At Avakata we run a monthly LLM audit on every client: a fixed list of queries, asked of five engines, parsed for which sources get cited and which don't. Six months of those audits, averaged across verticals, tell a consistent story.
The three signals that move citation most reliably:
- Freshness stamps. Pages with a visible "last updated" date get cited 2.3× more often by Perplexity. Pages with structured date metadata and a visible stamp get cited 3.1× more often. Half of this is the LLM's own preference for fresh sources; the other half is that fresh pages tend to be better.
- Source-of-truth structure. Pages that look like authoritative reference — definitions, then evidence, then citations, in that order — get pulled into answers more often than pages that argue a position. The engines are not looking for op-eds.
- Provenance signals. Author name, role, credentials, last-reviewed-by line. Cheap to add, disproportionately rewarded. Anonymous content gets cited less; named-with-credentials content gets cited more.
What doesn't move citation
Two things we thought would matter — and didn't, much.
Word count. The LLM-era SEO industry has decided long-form is the answer. Across our audit data, length above ~1,200 words shows diminishing returns on citation. A 2,400-word piece is not 2× as citable as a 1,200-word piece. The ceiling is faster than people think.
"Schema" markup beyond the basics. Article schema, Author schema, basic FAQ schema — all useful. Beyond that, the long tail of schema types makes essentially no measurable difference in citation rate. We expected more.
How the engine does this in practice
Avakata's engine watches the LLM audit feed on a rolling basis. When a page that should be cited stops being cited — or starts to be cited by the wrong engines — the engine drafts a candidate update, runs it past a critic against the page's authority brief, and ships it to a slice of traffic.
The update almost always boils down to two things: refresh the visible date and tighten the lead-in paragraph so the first 90 words can be lifted verbatim by a generative engine without losing meaning. (Engines are very good at lifting clean opening paragraphs and very bad at synthesizing scattered prose.)
What to do this week, if you're doing it manually
- Add a visible "last updated" stamp to your top-twenty pages. Today.
- Put an author byline with role + a one-line credential on each.
- Rewrite the first 90 words of each so they can stand alone as a self-contained answer.
- Run your own LLM audit once a month — even a manual one. You can't optimize what you don't measure.
If you'd like the audit run for you, with the engine taking the diff and shipping the changes — that's what we do. Book a 30-minute discovery.