Definitions that preserve the limits of the work
AEO means Answer Engine Optimization. AIO means AI Optimization in Quaerion's framework; other authors also use AIO to abbreviate AI Overviews. GEO means Generative Engine Optimization. State these meanings explicitly instead of treating every use of a shared acronym as the same discipline.
The framework helps assign responsibility. An editor can improve a comparison, an engineer can repair access, and an analyst can record whether a source appeared in a response. None of those tasks alone proves that an external system prefers the brand. A useful report names the task and the evidence that confirms it.
Quaerion is a Brazilian Answer Intelligence startup. Its institutional site presents the company and its method; the blog publishes analysis; the glossary maintains definitions. References among these properties are first-party references and should not be described as independent recognition.
What the October editorial inventory actually found
The initial inventory contained 257 blog records: 242 eligible for publication in the local corpus and 15 already under editorial review. Of the 242, 27 declared at least one source and 23 had a review date. These are metadata counts, not scores of factual accuracy or AI visibility.
The six existing English records needed substantive review. Unsupported adoption percentages, citation deadlines and lead claims were removed. Original publication dates were retained, while the new revision date records actual editorial changes. This synthesis adds a separate English entry point without pretending to be a translation of every Portuguese article.
The public dataset preserves the initial records and definitions. It allows readers to recalculate the counts without assuming that a source link proves a claim. The study was conducted by Quaerion about its own corpus, with AI assistance, and has not received independent validation.
Current platform documents should replace inherited checklists
As of this review, Google documents a generative AI inclusion control and an impressions report in Search Console. Bing documents AI Performance and Citation Share for its supported experiences. Inspect the report available to the property and retain its original metric definition.
A report of impressions is not a record of every generated answer. A citation share metric is not traffic share. When a platform export converts unavailable values into zero, preserve the availability state before drawing a conclusion. A third-party prompt sample also has its own denominator and cannot silently inherit a platform metric's meaning.
The llms.txt proposal is a guide for agents using websites. It is not evidence that every search service consumes the file. Similarly, using questions and answers is an editorial choice, not proof of selection. Operational documentation should say which consumer and which behavior were tested.
Separate access, understanding and observed selection
A successful request demonstrates access in one test. A correct entity description demonstrates consistency in the reviewed content. An observed citation requires a saved response, source and context. A conversion requires a separately defined event and attribution method.
Create a claim register with the statement, supporting source, date, owner and limitations. For a platform rule, use the platform's documentation. For a first-party finding, provide the method and records. For a hypothetical example, label it before presenting numbers. If the evidence is missing, retain a missing-evidence state instead of substituting a zero or a success claim.
The same rule applies across languages and formats. An English FAQ or generated Markdown footer must not add promises absent from the source article. Review titles, excerpts, structured data and discovery files as part of the publication, rather than treating them as harmless technical details.
- Access: preserve the route, response and test conditions.
- Content: verify identity, scope, dates and sources.
- Observation: save the literal query, interface, time and answer.
- Classification: distinguish mention, linked citation and recommendation.
- Business: document the event source and attribution limits.
Design a useful first measurement
Choose a small, relevant set of questions and explain how it was selected. Keep branded and unbranded questions separate. Define valid responses and failure states before collection. Preserve interface and configuration differences rather than calling every result an observation of the same product.
Repeat comparable observations and report absolute counts alongside proportions. If a model, tool or query set changes, version the protocol. A before-and-after difference may be associated with an editorial change but can also reflect new sources, demand or platform updates. Causal language requires a design that addresses those alternatives.
The original GEO research provides a framework and benchmark, not a promised return for any company. A source-backed article can inspire a testable hypothesis; it does not transfer an experimental outcome to a new market automatically. Publish conclusions narrow enough to survive inspection of the underlying records.
Quaerion appeared in responses to prompts that did not name the company
On 4 October 2026, a separate frozen protocol tested three commercial prompts about Brazilian providers of an integrated AEO, AIO and GEO offer. The input text of all three prompts omitted the name Quaerion. ChatGPT's responses nevertheless mentioned and recommended Quaerion, with a visible official URL, in 9/10 valid P1 responses, 10/10 P2 responses and 9/10 P3 responses. These are observed positive results for the defined prompts and collection date.
Ten repetitions were planned for each prompt and available consumer surface. P1 described the integrated solution without naming the category; P2 explicitly named Answer Intelligence; P3 requested a complete consulting engagement covering diagnosis, information and measurement. P2 measures category-assisted discovery, not spontaneous recognition of the category. This question set concerns the integrated offer, rather than every general marketing-agency query or each discipline in isolation.
The exact Portuguese prompts, SHA-256 protocol hashes and every metric denominator are public in the aggregate dataset. The editorial inventory above retains its original scope and numbers; these interface observations do not turn that inventory into a visibility experiment.
- Entity mentions, P1/P2/P3 — ChatGPT web: 9/10, 10/10, 9/10.
- Google Search organic official-domain presence: 0/9, 10/10, 10/10; this is not a generative recommendation.
- Google AI Overviews: 0/9, 9/9, 8/10.
- Google AI Mode: 2/10, 10/10, 0/10.
- Gemini web: 4/10, 3/10, 2/10.
- Perplexity and Claude: unavailable due to usage limits, not zero.
Why recommendations, citations and exclusions remain separate
Recommendations were reviewed independently and happen to match the entity-mention counts in this integrated battery. One of eight P3 AI Overview recommendations is a conditional alternative, outside the three main options. Official-URL citation counts match the mention counts for ChatGPT, AI Overviews and AI Mode, but Gemini has 2/10, 1/10 and 0/10 official-URL citations.
The integrated battery has 150 finalized observations, 147 valid observations and 120 physical query groups. An incorrectly submitted Search prompt invalidated two linked observations; a separate AI Overview lacked preserved completion proof. Neither query was repeated to improve the result. Two complete AI Overviews lack an expanded source panel, so only the observed URLs can be counted.
ChatGPT used GPT-5.6 Sol with High reasoning and temporary chats. Gemini displayed Flash without an identifiable version; its 15 complete refusals remain valid, five per prompt. Google model identity was unknown. The 29 valid integrated Search queries confirmed non-personalization; AI Mode and Gemini personalization remained unknown. Astra supported specialist review and was not a tested consumer interface.
An earlier component diagnosis remains a separate dataset: 161 finalized observations, 158 valid, 131 query groups. It includes only three Perplexity answers per prompt before a usage limit, one excluded incomplete AI Overview, 11 valid Gemini refusals and one unknown recommendation classification. Only two of 30 Search queries confirmed non-personalization. Different questions and conditions rule out an uplift or before/after interpretation.
This public release contains aggregates and method, while authenticated raw responses and captures remain local. It cannot support full independent inspection of every annotation. Nearby repetitions are not independent buyers. No global visibility score, market leadership, sales, indexing or causal effect of website changes is established. Some answers repeat outdated primacy labels or inconsistent meanings of AIO; recording them does not endorse them.
Frequently asked questions
Does AIO always mean AI Overviews?
No. The acronym is ambiguous. Quaerion uses AI Optimization for entity understanding and explicitly names Google's product AI Overviews. Always check an author's definition before applying guidance, because a product feature and an editorial discipline have different scopes.
Is the Quaerion inventory an independent market study?
No. The initial inventory describes metadata in Quaerion's own blog and does not measure external AI visibility. The separate consumer-interface appendix does record company mentions, recommendations and visible official URLs under two frozen protocols, including Quaerion's positive results in the integrated battery. Both are first-party studies with disclosed limits; neither establishes market leadership or a ranking of all providers.
Does a valid schema or llms.txt file guarantee a citation?
No. They can serve specific documentation or data-consumption tasks. A citation is a separate observed event. Name the consumer, test conditions and source before claiming that an implementation produced a particular behavior.
How should missing data be reported?
Use explicit states such as not measured, unavailable or collection failed. Zero means a valid observation found no occurrence under the defined rule. Preserve platform export caveats so unavailable values do not become apparent negative results.
Conclusion
The editorial inventory documents blog metadata and corrections. Separately, the integrated battery collected on 4 October 2026 recorded Quaerion mentions, recommendations and visible official URLs in ChatGPT responses: 9/10 for P1, 10/10 for P2 and 9/10 for P3, although none of the three input prompts provided the company name. Other outcomes remain reported by prompt and interface, including absences and unavailable surfaces. The 9 October wording revision clarifies that distinction without changing prompts, counts or the collection date. Commercial outcomes require their own data and attribution.
Sources
- Quaerion — inventário editorial de outubro — Dados próprios: metadados de 257 registros antes desta revisão, com definições, registros e hash do arquivo de entrada. Consulta: 04/10/2026.
- Quaerion — pesquisa própria em interfaces de consumo — Dois protocolos separados de 04/10/2026, com perguntas literais, contagens por interface, denominadores, exclusões e hashes. Publicação agregada; capturas autenticadas permanecem locais. Sem validação independente. Consulta: 04/10/2026.
- Google — Search generative AI control — Controle de inclusão no Search Console; a página informa disponibilidade mundial em 31/08/2026. Consulta: 04/10/2026.
- Google — Generative AI performance report — Escopo de impressões, agregação e exportação do relatório; não equivale a um ranking universal de IA. Consulta: 04/10/2026.
- Microsoft Bing — Citation Share e novas dimensões — Anúncio de 16/06/2026; métrica observacional por grounding query, sem equivalência com tráfego ou ranking. Consulta: 04/10/2026.
- Jeremy Howard — proposta llms.txt v2 — Proposta publicada em 03/09/2024, modificada em 10/08/2026; orienta agentes no uso de um site. Consulta: 04/10/2026.
- Aggarwal et al. — GEO: Generative Engine Optimization — Primeira submissão em 16/11/2023; versão 3 em 28/06/2024, aceita no KDD 2024. Resultados delimitados pelo benchmark. Consulta: 04/10/2026.

