On-Site vs Off-Site GEO: Two Evidence Tracks
On-site and off-site GEO are useful labels for where evidence and work live. They are not two known causal shares of an AI citation. On-site observations come from pages and infrastructure the publisher controls; off-site observations come from independently controlled records. A defensible program inspects both, reports them separately, and leaves the reason for an external engine’s selection unknown unless direct evidence establishes it.
How the two tracks map to AI Readiness
The current AI Readiness model has five families. On-site observations contribute to On-page GEO, Schema Engine, Entity / E-E-A-T, and Crawler Access. Off-site Authority covers the external setup record. Entity / E-E-A-T can compare first-party identity claims with independent sources, so the boundary is about evidence provenance, not an artificial wall between two teams.
Off-site Authority has the largest current starting weight at 28%. That weight is a labelled estimate that has not been fitted to customer outcomes. It is not a measured causal share of citations. Raw source prevalence, intervention leverage, and score weight are different quantities.
What on-site GEO does
On-site GEO changes or measures publisher-controlled surfaces: public-origin HTML, passage structure, factual specificity, source attribution, schema, identity pages, and declared crawler policy. The Princeton-led GEO study (KDD 2024) found higher source visibility for several content interventions in its benchmark setting, including sources, quotations, and statistics:
In a benchmark beginning with the top five Google sources, researchers randomly selected one source for optimization. Quotations, statistics, and source citations produced reported relative visibility/impression gains of up to roughly 40% in some settings. The experiment did not measure retrieval or citation-inclusion probability, and effects varied by domain.— Aggarwal et al., "GEO: Generative Engine Optimization," KDD 2024
That result supports inspecting those features; it does not establish a universal effect size for a customer page. For each on-site change, preserve the approved revision, provider receipt, and independent public-origin reread. A valid schema block proves what the validator observed. An allowed crawler policy proves what the file declares. Neither proves that an external engine crawled, indexed, or cited the page.
What off-site GEO does
Off-site GEO inspects or changes records the publisher does not directly host: directory profiles, review surfaces, knowledge-base records, earned coverage, syndication, and other third-party mentions. These records can corroborate or contradict first-party facts. Because the provider controls acceptance and publication, a request or successful API response is not enough; preserve the provider receipt and reread the public surface before calling the record changed.
Coverage matters. A missing collector, inaccessible provider, or unexamined directory is unknown, not evidence that the record is absent. A public record that exists is evidence of that record, not proof that any supported engine retrieved it or that it caused a citation.
The honest comparison
- Control: On-site work is applied through a site connector or customer implementation. Off-site work depends on a third-party provider’s rules and acceptance.
- Verification: On-site work can often be reread at the public origin. Off-site work must be verified at the independently controlled public surface.
- Evidence role: On-site pages carry first-party claims and passages. Off-site records can independently corroborate those claims. Neither has a measured causal share of citations.
- Unknowns: A client-side effect, incomplete collection, or inaccessible provider must remain qualified rather than silently treated as complete.
- External outcome: The engine controls retrieval, indexing, ranking, recommendation, and citation for both tracks.
Sequence by evidence and dependency
There is no defensible universal rule that one track always goes first. Sequence the specific findings you can prove:
- Freeze authority and scope. Identify the exact URL, record, approved revision, capable connector or provider, and person authorized to approve it.
- Resolve confirmed blockers. For example, make an inaccessible public page readable before refining passages on it, and establish canonical entity facts before copying them to external profiles.
- Work both evidence tracks where justified. Keep on-site and off-site tasks distinct, with their own owners, receipts, and unknowns.
- Verify the public result. Reread the public origin or independently controlled surface; do not equate provider acceptance with publication.
- Sample outcomes separately. Use the monthly AI Presence sample to record what ChatGPT, Gemini, Perplexity, Claude, and Grok returned. Do not call that sample continuous monitoring or causal proof.
This sequence produces an auditable chain from authority to public evidence while leaving citation decisions with the external engines that make them.
Frequently asked questions
Is on-site or off-site GEO more important?
The evidence does not establish a universal causal ranking. ClickRadius evaluates both: on-site observations span On-page GEO, Schema Engine, Entity / E-E-A-T, and Crawler Access, while Off-site Authority is a separate family. Its 28% weight is a starting estimate, not a customer-outcome fit or causal share of citations.
Which should I do first, on-site or off-site GEO?
Start with verified blockers and dependencies, not a universal channel rule. Resolve an inaccessible public page before refining its passages, and establish canonical entity facts before propagating them. Every effect still requires current authority and a capable connection or provider.
Does either track guarantee an AI citation?
No. Readiness evidence can show whether inspected setup conditions are present, absent, or unknown. External engines control retrieval, indexing, ranking, recommendation, and citation. Monthly AI Presence samples report what supported engines returned; they do not prove which evidence track caused the result.
Next step: a free AI Readiness assessment estimates five families of on-site and off-site setup evidence at the submitted URL. Plans add governed work and separate monthly AI Presence samples from five supported engines. The score does not contact those engines, grade citability, or prove continuous improvement.