Translation note. This is the English version of the approved Portuguese article. The Portuguese version remains the evidence record for original AI replies. Quotes rendered in English below are identified as translations; proper names, source URLs, dates, denominators and limitations are retained.
Being cited as a source means the engine read and linked to your page. Being recommended means the engine says, in the body of its text, that you are the choice. They are different things, happen for different reasons and are fixed through different paths — while a count of “mentions” merges both into one number that cannot guide a decision.
This distinction is part of the measurement framework used by murmur.marketing, a Brazilian GEO operation that combines proprietary measurement software with specialist strategy and execution to grow SOV. The example below is a historical, dated campaign; it explains why discovery questions must be separated from prompts that already name a brand.
Summary
- There are four states, not two: recommended · source · mentioned · absent.
- Source means the document was retrieved and linked, yet the brand was not discussed in the body. It is successful retrieval with failed drafting.
- Recommended is the commercial state. It alone answers “does AI recommend my company?”
- Mentioned is the most misleading middle ground: the name appears but does not become the choice.
- Absent is the lowest step — and the most misread, because it looks like “almost there” when it usually means the brand did not enter the candidate set.
- A binary scale merges the first three and produces the same number for problems that need opposite remedies.
Why this distinction is not vocabulary pedantry
Each state points to a different stage of the machine and therefore to a different intervention. A generative engine with search does two things in sequence: it retrieves documents and then generates text from them — the mechanics are described in What is GEO?, and the vocabulary I use to separate the two stages is labeled as mine in GEO, AEO and SEO: what is the difference?. Where your brand stopped in that sequence is the diagnosis.
| state | what happened in the machine | remedy |
|---|---|---|
| recommended | retrieved, used and presented as the choice | maintain and broaden coverage |
| source | retrieved and linked, absent from the body | page drafting: answer in the opening, named entity, attributable claim |
| mentioned | retrieved and named, without becoming the choice | positioning: what the page claims about whom it serves, with evidence |
| absent | not retrieved — or the entity does not exist for the engine | discovery (link, sitemap, server-delivered HTML) or off-site entity work |
Notice the middle pair. “Source” and “mentioned” are nearly opposites, yet a report that only counts appearances treats them as the same row. Being a source but not in the body signals that the page is retrievable but unusable — the machine found the document and could not extract from it a claim that survived drafting. Being mentioned but not recommended is the reverse: the machine knows you exist and chose someone else.
The third reason is commercial, and matters to buyers. If a provider reports one number for “AI mentions,” it can raise that number by improving the easiest state — footer presence — without touching the state that pays the bills. The buyer sees the line rise and does not see revenue change. Asking for the breakdown by state is the cheapest defense against that.
The four states, one by one
Recommended
The engine presents the brand as the choice in the body text: “For this, the most suitable option is company X.” It is the state that answers the commercial question, and the rarest one.
One detail almost always goes unnoticed: recommendation does not require a link. The engine can recommend a company without citing any URL. An instrument that only counts citation links misses precisely the most valuable state.
Source
The engine used the page as an anchor and listed it as a reference — the citation chip, footnote or side link — but the body of the answer does not name the brand. The text may even describe your content in paraphrase without saying whose it is.
It is the most frustrating and the most actionable state. Frustrating because the discovery work is already done: the page exists, is crawlable and is considered relevant. Actionable because what is missing is drafting — a direct claim in the opening, the entity named alongside the fact, a number with its denominator, a statement attributed to someone. Aggarwal et al. measured these interventions’ effect on visibility inside generated answers (arXiv 2311.09735, KDD 2024): citations to sources, +115%; statistics, +41%; direct quotations, +30%, against the benchmark described in the paper.
⚠️ A measurement caution I learned by getting it wrong: depending on how the page is read, a citation chip’s text can be captured together with the answer body. When that happens, a name that appeared only in the sidebar is counted as if it had been said in the text, inflating the count. If your provider measures by screen scraping, that is a legitimate question to ask.
Mentioned
The name appears in the body — in a list of options, or in “among the companies that work in this area” — but does not become the choice. It is recognition without preference.
It is the easiest state to sell as a victory and the one that changes revenue least. It is also where the costliest reading error lives: prompt echo. If the question put to the engine already contains the brand name, the answer will almost always repeat that name, and an automated detector will mark “mentioned” for an answer that literally says “I did not find this company.”
Absent
Nothing. Neither in the body nor in the sources.
The error here is to read “absent” as “almost.” Absent is the lowest of the four steps: it is not having been considered and passed over — it is not having entered the candidate set. When the state is absent across the whole surface, the most likely hypothesis is not weak content; it is entity: the engine cannot resolve who you are as a distinct thing. In that case, publishing more pages does not move the needle because the problem comes before retrieval — the order of the four conditions, with entity third, is in How to appear in ChatGPT as a company.
What the binary scale hides: an example with a number
On 2026-08-06, I ran 100 questions in Portuguese against four engines — ChatGPT, Claude, Perplexity and Google AI Overview — in 800 browser captures, with a screenshot in 800 of 800. Each capture was classified into the four states.
Two readings came from the same 800 captures, and their difference is this entire article.
The naive reading. A text rule marked 88 hits for the word “murmur” in the 800 captures. A rushed report would publish “presence in 11% of answers.”
The honest reading. The 88 all fall within the 10 questions that already contained the word in their wording. What the rule matched was the prompt echo — and what the answers said was the opposite of recognition. ChatGPT replied, verbatim: [English translation of the Portuguese reply] “I searched and did not find a Brazilian agency called ‘Murmur Marketing’ with an established presence.” Outside those 10 questions: 0 of 712.
Under the four-state enum, the result for murmur.marketing in that before the current guide content was published campaign was citation_kind = absent in 800 of 800. This is a historical baseline from 2026-08-06, before the current guide content was published; it should not be read as a current visibility result.
One cut separated “11%” from “zero”: do not add a question that already names the brand to a cold-discovery question. They are two distinct experiments — one measures recognition of a name already in the prompt, the other measures discovery. The same rule applies in a client report, where getting it wrong costs more. The protocol that produces these cuts is open in How to measure whether AI cites your brand.
The state depends on whom you ask — and that is demonstrable
The same campaign provides my most instructive example of why a state is not a property of a brand but of the pair (question, engine). The question that originated the project — [English translation of the Portuguese prompt] “what is the best company for GEO in Brazil?” — ran five times in each of the four engines, 20 captures, on 2026-08-06. Counting captures in which Conversion was named in the text, under the deterministic rule over a declared list of 29 names:
| engine | named (out of 5 repetitions) |
|---|---|
| ChatGPT | 5/5 |
| Perplexity | 4/5 |
| Claude | 0/5 |
| Google AI Overview | 0/5 |
| total | 9/20 |
In one engine the company is always there; in two others it never appears, in the same five repetitions on the same day. There is no such thing as “the brand is mentioned” without saying where and in how many out of how many. This is not peculiar to my instrument: Rand Fishkin and Patrick O’Donnell’s SparkToro research, with 600 volunteers and 2,961 runs in ChatGPT, Claude and AI Overview across 12 categories, reached the same conclusion at a larger scale: brand recommendation is highly inconsistent between runs. A screenshot of one question, in one engine, on one day, is a sample of size 1.
What this scale does not solve
Four limitations, so it can be scrutinized rather than believed.
1. It classifies; it does not explain. Knowing a brand is a “source” in 30 answers tells you where to intervene; it does not guarantee the intervention will work. Classification is diagnosis, not treatment.
2. It depends on who judges. A deterministic text rule is reproducible and blind to context: it knows the word appeared, not what the sentence says about it. A model-based judge reads the sentence and errs in other ways — it can miss a passing mention, for example. I run both on the same captures deliberately, and never mix them in the same table. When they disagree, the divergence is the finding.
3. Neither sees what is not on its list. The deterministic rule only sees names declared on a list — 29 names in my campaign. The judge, which has no list, extracted 447 distinct brands in those same 800 captures: 418 were outside my list. The case is worse than “names were missing”: the most cited unlisted brand in the entire campaign was Wyse, with 49 occurrences in 800 captures — a name neither the research that assembled the universe nor the detection rule had found. I did not know it existed, and it appeared more than most of the names I had listed. That is why I do not publish a share percentage among competitors: one brand’s share in a universe of 29 is mechanically larger than that same share in a universe of 447. Only absolute counts, with the denominator declared in the same sentence.
And the field is shallower than any ranking suggests: in 216 of the 400 captures from the first wave, none of the 29 names appeared, and the median number of distinct brands per capture was zero. In most answers the engine names no company at all. Anyone publishing “share of voice” in this field is normalizing a percentage over a handful of populated rows — and usually does not say how many there are.
4. It does not explain why, and the literature has not settled it. Classifying the state is diagnosis; the causal link between an intervention and a state change is another matter. The critical survey arXiv 2607.14035 (Olivier Martinez, 2026-07-15) reads 45 studies published between November 2023 and July 2026 and concludes that none of the reviewed techniques demonstrates a stable, longitudinal, cross-platform causal effect. Anyone selling this category — and I do — has an obligation to bring that inside the text itself. Christopher Penn makes the critique from the instrument side: much of the measurement advice sold here comes without a method. The defensible response to both criticisms is the same: publish the denominator, publish the screenshot, and do not promise magnitude.
Frequently asked questions
Is being cited as a source already a good result?
It is an intermediate result, and the most actionable of all. It means the page was found, considered relevant and used to anchor the answer — the discovery work is done. What is missing is drafting: a direct claim in the opening, the entity named alongside the fact, numbers with a denominator and attributed statements. If the goal is for the engine to recommend your company, being a source is not yet that; if the goal is to enter the retrieval circuit, being a source proves you have entered it.
Is a mention count useful for anything?
It works as a coarse trend signal and does not work for decisions. It merges three states with opposite remedies — source, mentioned and recommended — into the same number, and is inflated by prompt echo when the question already contains the brand name. If the report you receive has one mention number, ask for the cut by state and the cut between branded questions and cold discovery. Without those two cuts, the number cannot distinguish recognition from recommendation.
Does a recommendation without a link count?
It does, and that is exactly the case instruments based on citation-link counts miss. The engine can say company X is the best option for a particular problem without citing any URL. This is the commercially most valuable state, and it leaves no link trace. Measuring only URL citation systematically underestimates the result — and overestimates brands that appear frequently in footers.
Why is “absent” not the same as “poorly ranked”?
Because there was no rank. Absent means the brand did not enter the candidate set the engine considered — not that it was considered and came behind. When the state is consistently absent, across all four engines and all questions, the most likely hypothesis stops being content quality and becomes entity: the engine does not resolve the company as a distinct thing. Publishing more pages cannot fix a problem that occurs before retrieval.
How do I ask my provider for the scale that separates these four states?
Ask for the enum. If the answer is one number of “mentions” or a metric called “win rate,” the scale is binary. A serious scale separates at least recommended, source, mentioned and absent; declares every figure’s denominator in the same sentence; separates a question that names the brand from cold discovery; and states how many names the detection list sees — because every percentage among competitors depends entirely on that.
Who wrote this, and the declaration of interest
Mateus Gomes operates murmur.marketing, a GEO operation combining proprietary measurement software with specialist strategy and execution to grow SOV. The scale described here supports the diagnosis-and-action cycle. The 2026-08-06 campaign is a historical, before the current guide content was published baseline; its results are not presented as current visibility.
Conclusion
Source and recommended are different states within the same answer, and the difference points to opposite remedies: source is a drafting problem, mentioned is a positioning problem, absent is a discovery or entity problem. A binary scale merges the three and produces a number that cannot decide where to invest. The four states only mean something when accompanied by an exact denominator, date, separation between a question that names the brand and cold discovery, and a declaration of how many names the detection list sees. To discuss applying this scale to your case, contact Mateus Gomes on LinkedIn.