Evidence standard · 10 minute read
AI-search advice is moving fast. The evidence needs guardrails.
Some GEO advice is useful. Some is a vendor correlation mistaken for a universal rule. Some is simply a new label on solid search and publishing practice. Here is what This Is GEO measures—and what we deliberately refuse to turn into a score.
01 · Strong enough to act on
Measure the conditions platforms actually expose.
Google, OpenAI, and Perplexity all publish crawler or eligibility guidance. The wording differs, but the practical foundation is familiar: important public pages need to be reachable, indexable where intended, readable, internally discoverable, and consistent with the facts and markup they expose.
Can the intended crawler reach useful content?
GEO retains response status, redirects, indexing controls, robots rules, canonical targets, sitemaps, and whether meaningful text exists without a click or login.
Can a system resolve the business correctly?
GEO checks names, offerings, people, service areas, profiles, page labels, hierarchy, and visible-versus-structured-data consistency.
Can a passage stand on its own without becoming misleading?
GEO checks early answers where appropriate, descriptive headings, scope, qualifiers, definitions, distinct questions, and understandable tables or media.
Can material claims be checked?
GEO looks for attributable sources, fit-for-claim evidence, named expertise, methods, honest updates, precise uncertainty, and supportable promotional language.
Does the wider web resolve to the same entity?
GEO records independent coverage, corroboration, source diversity, coherent profiles, namesake risk, and durable citations to first-party work.
02 · Useful, with conditions
Research findings are hypotheses—not commandments.
The foundational GEO paper found that citations, quotations, statistics, and fluency could change its benchmark visibility measures. It also found that results varied by domain. That supports making important claims specific and well sourced. It does not support stuffing every section with a fixed number of statistics or promising a 40% lift on a live platform.
Recency, bylines, FAQ sections, lists, comparison pages, and answer placement can all be useful in the right context. GEO checks whether they are truthful, substantive, and appropriate for the buyer task. It does not award points merely because a heading ends in a question mark or a page uses a fashionable format.
03 · Observe, don’t score
Provider behavior belongs in a dated panel.
Which source types appear, how much Claude’s source pool overlaps with ChatGPT’s, whether Perplexity cites several pages, and whether a business is mentioned or recommended are observations about a particular provider, prompt set, locale, model, and date. They are not permanent properties of the website.
GEO preserves provider, model, prompt, locale, date, citations, mentions, and recommendations instead of blending every engine into one invented truth.
Sources are grouped into pages the business owns, profiles it can improve, coverage it may be able to earn, and external references it can only monitor.
AI answers vary. One run is a dated observation; repeated comparable panels build a more useful trend.
A change followed by a better panel is worth investigating. It is not proof that the change caused an outside system to respond.
04 · What GEO refuses to score
No magic files. No hidden copy. No fake certainty.
- llms.txt as a requirement.Google explicitly says no new AI text file is required for AI Overviews or AI Mode. GEO will not award readiness points for an unsupported requirement.
- Hidden machine-only content.Important facts should be useful to people and machines. We will not recommend cloaking or an invisible alternate story for bots.
- FAQ or HowTo schema as a citation shortcut.Visible answers and accurate markup can be useful. The markup itself is not a promise of generative inclusion.
- Keyword, listicle, quote, or statistic quotas.GEO evaluates meaning, support, and task fit—not arbitrary content recipes.
- One-run visibility as a stable rank.A dated answer is evidence of what happened once under disclosed conditions, not a permanent score.
05 · The coverage map
Fifty readiness checks. Separate visibility evidence.
Readiness asks whether the website and its evidence are prepared. AI Visibility records what released providers actually returned. Keeping those separate prevents a polished website from being presented as a guaranteed recommendation—and prevents an absent observation from becoming a fake zero.
See the evidence on your site
Start with one defensible next move.
Geo reads five public pages and returns evidence-backed findings without requiring a website login or pretending to know a provider’s secret algorithm.
Sources and boundaries
Read the underlying material.
- Google Search Central: AI features and your website Google says AI Overviews and AI Mode use established Search requirements; no special AI file or schema is required.
- OpenAI: Publishers and Developers FAQ OpenAI documents OAI-SearchBot access and ChatGPT referral tracking without promising inclusion or placement.
- Perplexity: crawler documentation Perplexity distinguishes its search crawler from user-requested fetches and publishes current crawler guidance.
- Google: structured-data guidelines Structured data should match visible content and can enable search features; correct markup does not guarantee an appearance.
- Google: spam policies Serving materially different content to crawlers and people to manipulate search is cloaking.
- Aggarwal et al.: GEO: Generative Engine Optimization The foundational benchmark found that several content treatments changed measured visibility, with results varying by domain.
- Schulte, Bleeker & Kaufmann: Don’t Measure Once The paper argues that variable AI answers require repeated measurements rather than one precise-looking snapshot.
Research and platform documentation can change. This article describes the evidence standard used on September 3, 2026. It does not claim access to any provider’s private ranking, retrieval, or source-selection system.