Get SEEN.
A framework for making an entity easy for AI to find, read, verify, and recommend.
SEO optimized pages for rankings. SEEN treats GEO as corpus engineering: owned content, platform-hosted proof, and third-party corroboration made coherent enough for machines to use.
The failure mode is category blanking: an AI gives an answer in your category, in front of your buyer, and acts like your entity does not exist. Not lower-ranked. Absent.
The four layers
-
S: Structure
Make the answer extractable. Put definitions, pricing, scope, comparisons, and core facts in visible HTML with real hierarchy.
ALM Corp / Indig, 2026: 44.2% of ChatGPT citations came from the first 30% of webpage content.
-
E: Evidence
Make claims verifiable. Pair numbers, outcomes, and positioning with methodology pages, case studies, bylines, and primary sources.
Wan, Wallace, Klein, ACL 2024: LLMs score evidence primarily on direct relevance to the query.
-
E: Entity
Make the entity unambiguous. Use one canonical name, one description, consistent profiles, schema, and a clean sameAs graph.
MultiHal, 2025: knowledge-graph-grounded retrieval improved hallucination scores across evaluated models.
-
N: Notability
Make the wider web corroborate the recommendation. Earn third-party mentions, platform-hosted reach, links, and organic community presence.
Otterly.AI, 2026: community platforms captured roughly 52.5% of AI citations in a 1M+ citation sample.
Why the weakest layer wins
A 9/10 in Structure paired with a 2/10 in Notability performs worse in practice than a 6/6/6/6 distribution. A clean answer with no reason to be trusted gets skipped. Strong proof AI cannot extract goes unread. Both, attached to the wrong entity, get credited to a competitor.
The quick audit
Ask an AI system to explain your entity, who it is for, why it should be trusted, and which sources support the answer. Then inspect what breaks: extraction, proof, identity, or external corroboration.
SEEN checklist
Open PDFScore each item 1 if it is clearly true today, 0 if it is missing, vague, or only partially true. Fix the lowest layer first.
S Structure 10 items
E Evidence 10 items
E Entity 10 items
N Notability 11 items
Quarterly GEO report
A recurring readout on which engines cite what, what is moving, and what still looks like hype. One email per quarter.
Sources and resources
- SEEN preprint paper The paper version of SEEN: evidence base, model, checklist, and validation plan.
- The SEEN manifesto The full essay: failure modes, mechanisms, sources, and the bet behind corpus engineering.
- The SEO-to-GEO Gap SSRN paper on how generative engines change what findable means.
- awesome-geo A curated repo of AI-discoverable surfaces, crawler notes, and platform scoring.
What SEEN does not claim
- Not a guarantee of citations, mentions, or rankings. AI answer surfaces are stochastic.
- Not a replacement for SEO. Different surface, different signal, shared substrate.
llms.txt, schema, and one-shot rewrites are infrastructure, not magic levers.- AI behavior varies by prompt, surface, model, account state, location, and time.
- Single screenshots are weak evidence. Durability across repeated runs is what matters.
Interested in SEEN?