Filter 1 — Retrieval: the binary gate

Nothing else matters if retrieval fails. The checks are mechanical: does the robots.txt allow the engine's crawler; is the page in the relevant index (Google's for AI Overviews/Gemini, Bing's for ChatGPT/Copilot); does the content render without JavaScript execution; does the server respond fast enough for a live fetch. Audit this layer first — it is binary, cheap to fix, and the single most common cause of AI invisibility we find in onboarding audits.

Filter 2 — Passage relevance: the chunk is the unit

Retrieved pages are split into chunks, embedded into vectors, and matched against sub-queries. The model never "reads your page" — it reads 100-word slices of it. Chunks that win share a shape: they open with a direct statement, carry specific nouns and numbers, and make sense with zero surrounding context. This is why the answer-block technique exists: it manufactures exactly the chunk shape retrieval scoring prefers.

Filter 3 — Corroboration: one source is a claim, three are a fact

During synthesis, models treat single-source statements as claims and multi-source statements as facts. A spec stated only on your own site gets hedged ("the company says…"); the same spec confirmed by a directory, a review and a trade article gets stated plainly. Corroboration is why GEO programs invest in the source graph — the off-site web of directories, press, reviews and community threads — and not just in owned content.

Filter 4 — Authority priors: topic-conditional trust

Engines carry learned priors about which domains are trustworthy for which topics: government and reference sources for facts, established trade media for industry news, review platforms for product experience, community threads for lived experience. These priors are topic-conditional — a great marketing blog is not an authority on medical claims. Understanding which tier your claim needs is the difference between publish-and-pray and designed corroboration.

The fifth force — consistency over time

Engines reward entities whose facts don't drift. A brand whose founding year, product names and descriptions are identical across twenty sources resolves cleanly and gets stated confidently; the same brand with three different descriptions across sources gets hedged or skipped. Consistency is the cheapest authority signal to build and the most expensive to repair once fragmented.

The compounding insight: these filters multiply, not add. A technically perfect page with a unique-but-uncorroborated claim loses to an average page with a thrice-confirmed one. GEO strategy is the art of moving all four filters together — which is why our engagements sequence foundation → corroboration → authority, in that order.

Frequently asked questions

Is citation selection just domain authority repackaged?
No. Domain authority is a site-level SEO heuristic; citation selection is passage-level and topic-conditional. A niche forum thread can out-cite a major domain when the question calls for first-person experience. Authority priors exist, but they are per-topic and compete with relevance and corroboration.
Can I see which chunks of my page engines retrieve?
Indirectly. Engines don't expose chunk scores, but citation behavior reveals them: the passage an engine quotes is usually the chunk that won. Testing how answers change when you restructure a page is the practical way to reverse-engineer chunk preferences.
Does content length matter for citation?
Length itself, no — chunk quality, yes. A 600-word page with six well-shaped chunks outperforms a 4,000-word page whose facts are buried. Depth helps only when it creates more independently-quotable passages.