Retrieval selects information that can support an answer; generation writes the answer from available context. An accessible page may not be selected. It may also be used as a source without its publisher being recommended.
The distinction helps choose the next test. It does not make every internal engine step observable from the final answer.
How information reaches an answer
Search-enabled systems may interpret the question, find documents, select passages and write an attributed answer. Some expand a query into subquestions; Google describes such an approach in its AI Mode introduction.
Sequence and implementation vary. Previously known content, indexes and live queries may contribute differently. “Crawled”, “indexed”, “retrieved” and “cited” are therefore not synonyms.
murmur’s operational taxonomy
murmur uses the following areas to organize investigations. This is our working taxonomy, not a mandatory industry standard or a claim about every architecture.
| Area | Investigation | Useful evidence |
|---|---|---|
| Access and discovery | Does the content exist and can it be found? | HTML, links, sitemaps and server responses |
| Identity | Is the company distinguished from namesakes? | Correct descriptions and coherent public sources |
| Retrieval | Which documents appear as candidates or sources? | Visible URLs and accessible-index checks |
| Answer use | Which facts and brands are included? | Full text and reviewable classification |
| Attribution | How are references presented? | Links, cards and citation context |
Problems in several areas can coexist. The table guides investigation and intervention; it does not assign untested causes.
What each signal demonstrates
A server request shows access. Agent identification requires validation because user-agent names can be spoofed. Access does not show that the document entered an answer.
An index result documents presence in that index under the examined condition. It does not guarantee retrieval for every question.
A cited URL shows visible attribution in the observed answer. It is closer evidence of content use, but does not automatically mean supplier endorsement.
A prose recommendation shows the engine presented the brand as an option in that answer. Measuring recurrence requires a preserved universe and repetition.
A missing URL does not prove no website information was used. Brand absence alone cannot distinguish identity, selection or presentation problems either.
An example that changes the action plan
For Fly Vet, a June 2026 Brazilian measurement of 100 Portuguese questions recorded 81 Google AI Overview answers using the company as a source and no recommendations. ChatGPT and Claude recorded 30 and 41 recommendations, respectively, in their own 100-answer sets.
The case illustrates the difference between serving as a reference and being suggested as a supplier. Investigate the questions and evidence of offer fit rather than interpreting this as permanent inability to be recommended. The Gemini app was outside this scope.
Testing a hypothesis
Start with a question family and a recorded content version. Check access and identity; examine visible sources and competitors. State the expected intervention: improve the opening answer, clarify the category, add evidence or cover a requirement.
Record the date and change, repeat comparable questions and examine whether the expected signal appears. Preserve engine-level results. Improvement after an edit is temporal association; controlling other changes strengthens attribution.
The founding GEO study informs content hypotheses. Vercel’s crawler research informs access. They provide evidence about different parts of the process.
Frequently asked questions
What is RAG?
Retrieval-augmented generation combines information retrieval with answer generation. Its implementation varies; the term alone does not identify sources, indexes or selection rules.
Does crawling mean citation?
No. Access, selection and attribution are separate events. Keep evidence for each observable stage.
How can I identify a retrieval problem?
Combine access checks, identity, question coverage and visible sources. Final absence is a reason to investigate, not proof of a particular failed stage.
Is a citation without recommendation valuable?
Yes, it can indicate informational participation. If the goal is hiring, also track whether the company appears as a solution and whether recommendations recur.
How murmur works
murmur works through a SWAS model, combining proprietary measurement software with specialists who diagnose, plan and execute GEO to grow Share of Voice (SOV). Eligible contract models may include an SOV-growth guarantee, with the metric, baseline, period, scope and conditions agreed in writing. This does not guarantee an individual answer, a fixed position or sales.
Author: Mateus Gomes, murmur operator. This article explains our approach and discloses our commercial interest.