GEO Methods: An Evidence-Led Taxonomy

GEO methods fall into five useful working groups: technical access, content quality, structured information, reputation and measurement. These are editorial categories, not five proven ranking factors. A valid implementation can be valuable without having a demonstrated causal effect on AI citations.

Read the evidence before choosing a method

The GEO paper studies interventions using a defined benchmark and visibility measures. Its reported improvements of up to 40% should stay attached to those conditions, not become a promise for a new website or an individual formatting tactic. It does not establish that every technique in a practitioner checklist works in every current product.

Google's guidance for AI features says established SEO practices remain relevant; no special additional optimisation is required for AI Overviews or AI Mode. That is guidance for Google Search, not proof about every system. The llms.txt proposal describes a publishing interface, while our adoption study measures file availability. Neither establishes a citation lift.

We use three evidence labels below: documented mechanism means a source specifies a supported interface or eligibility condition; research result means a study supports an effect within its stated scope; editorial recommendation / hypothesis means a sensible practice whose independent citation effect we have not established. Measurement methods are procedures, not efficacy claims.

Family 1 — Technical access

Give a reader or authorised crawler a usable page: successful responses, readable HTML, stable canonical URLs, internal links and an accurate sitemap. Diagnose a real access failure before applying speculative fixes. Check rendered content as well as status codes: a 200 response with an error message does not meet the reader's need.

Robots policies and access controls have specific roles. Allowing a crawler is not an instruction to cite you. A User-Agent claim needs appropriate identity verification, and successful verification still does not prove use in training or generation. Our case study preserves this distinction for the historical traffic.

Markdown alternates and a curated llms.txt can be useful interfaces where the content is maintained consistently. They do not replace HTML, canonical links, source quality or ordinary discoverability. Test both human and machine channels when correcting a claim so an old claim does not survive in a text bundle.

Family 2 — Content quality and structure

Answer the actual question, define unfamiliar terms, give sources for important claims and make limitations visible. A standalone opening helps a reader understand the page quickly; headings and tables can make comparisons easier. These are editorial recommendations. We do not label them proven citation-ranking factors.

Primary research is valuable when its method and evidence are inspectable: sampling frame, dates, missing data, calculation and limitations. A polished article without a dataset does not become an experiment by using "we observed". Likewise, adding numbers without evidence makes a claim less reliable, not more citable.

Use concise passages where they improve understanding. There is no demonstrated universal paragraph length or chunk size for every retrieval system. Preserve context around a number so a quotation does not detach it from its denominator or date.

Our answer-first experiment does not establish a citation effect. Its primary observable page-level signal was crawler re-fetching, and the citation panel measured a whole domain. Those outcomes cannot be substituted for one another.

Family 3 — Structured information and links

Schema.org Article describes article properties; FAQPage describes a page of questions and answers. Use markup that matches visible content and verify it. A valid schema block establishes structure, not a universal improvement in AI answers.

Our JSON-LD recipes provide practical examples. The FAQ experiment is inconclusive, so it supplies no basis for a claim that adding FAQPage raises citations. Do not invent questions or reviews purely to add a type.

Internal links should help a reader reach the next relevant explanation or evidence. Link the explainer to the method, the method to the dataset and the result to its limitations. An orphan can be harder to discover, but its absence from a generated answer does not identify the single cause. Evaluate canonical and crawl evidence before merging or removing URLs.

Family 4 — Reputation and transparent authorship

Make the responsible publisher, correction route and relevant experience clear. Use personal or organisational attribution that matches reality. A fabricated named expert is worse than an honest organisational byline. GeoSalience currently uses its organisational byline; this guide does not claim an unpublished author-identity experiment.

External references and independent discussion can make work discoverable, but we have not established a universal exchange rate between a Wikipedia mention, a backlink and an AI citation. Seek useful editorial contributions and avoid manufactured endorsements or unsupported authority scores.

Dates, disclosures and correction notes let readers assess a page. Update a review date after a real review, retain the publication date and say whether new data was collected. These practices support accountability; they are not a guarantee that a retriever prefers a page.

Family 5 — Measurement and iteration

Start with a decision the measurement will inform. A citation panel, a website visit, a search impression and a bot hit do not share a denominator. Choose a small prompt set for a defined audience and retain its version, product surface, model if known, search settings, dates and language. Record successful answers and errors separately.

Citation rate needs a specified denominator: per provider, use valid responses in a fixed prompt set; a site-wide union across providers answers a different question. A failed response is missing measurement. A valid response with no citation can contribute a zero. Share of voice also needs a definition of which mentions or citations are counted and which comparison set is included.

For each run, report absolute counts as well as rates, plus coverage and cost. A trend that changes the prompt set, providers or grounding method needs an explicit break or a comparable subset. Our Lab displays the latest panel separately from historical citations and website analytics.

A before/after change is an observation. It does not isolate the cause from index updates, seasonality, another content edit or a changed model. Where a causal claim matters, pre-register the intervention, controls, window, analysis threshold and missing-data rules. Keep page snapshots and stop or version a study before altering its control.

Evidence matrix

A practical first-month review plan

First, establish the baseline. Choose a useful page and inspect access, source quality, canonical and links. Save its current version. Define a diagnostic measure before changing it. Selecting a business KPI is an owner decision, not something a tracker can infer.

Then correct the content. Remove unsupported claims, distinguish dated findings from present guidance, and make the next useful source easy to find. Check HTML, Markdown and any full-text bundle together. A correction can be worthwhile even if it produces no measurable traffic lift.

Then review instrumentation. Confirm that each chosen action is captured once under a written definition, QA is separate and missing observations do not become zero. Check both arms before interpreting an experiment. Use knowledge cutoffs and live search when comparing surfaces.

Finally, assess what can be learned. Compare equivalent periods, show sample sizes and say when the result is inconclusive. This is a workflow proposal, not a promise of results within a month. Use AI visibility tools compared when a documented collection method fits your needs; we do not recommend buying a tool to compensate for an undefined question.

FAQ

Is GEO separate from SEO?

It adds questions about generated answers and attribution, while sharing access and content foundations with search. Start with What is GEO? and the terminology guide.

Which method is proven to work for every engine?

None is established here as universal. Source documentation, a benchmark result and a local experiment have different scopes. The evidence category is more useful than an unqualified ranking.

Should every article have FAQPage schema?

Only use a type when it accurately describes the visible page. Our experiment provides no established citation benefit from adding it.

How was this review produced?

This is an AI-assisted editorial synthesis of the primary sources linked above and our published research record, reviewed on 21 October 2026. No new paid model run or causal study was performed. See Methodology, Editorial policy and the Foundations hub.

GeoSalience (2026, May 19). GEO Methods: An Evidence-Led Taxonomy. GeoSalience. https://geosalience.com/foundations/geo-methods-taxonomy