This is a diagnostic decision: owned content buys retrievability, while third-party mentions buy entity recognition. You need the second when that is your bottleneck. The expensive mistake is not choosing the wrong side. It is buying more of what you already have.
I write as Mateus Gomes, operator of murmur.marketing, a Brazilian GEO operation combining proprietary measurement software with specialist strategy and execution to grow SOV. The comparison below is a dated snapshot and should guide diagnosis, not be treated as a universal rule.
Summary
- Owned content addresses discovery and retrieval. Third-party mentions address entity recognition. They are different conditions and cannot substitute for one another.
- In the structure observed on 2026-08-06, the most frequently cited name on my list had 9 pages on the topic, 0.7% of its 1,366 sitemap URLs, and published less than one post per month.
- The runner-up reached its position in about 4 months, with 279 URLs almost entirely focused on the topic and no third-party coverage found. That coverage was unconfirmed, not proven absent.
- Both paths work, and neither covers more than a fifth of the field: 19 and 15 of 100 (question, engine) pairs. 51 of the 100 have no winner at all.
- The case that guides allocation is my own: a company I own had 324 published articles and remained absent on eight category questions, with its competitor listed first.
- Diagnose before budgeting. Buying entity recognition when retrieval is the gap is slow and expensive; buying content when entity recognition is the gap does not change the result.
These two investments buy different things
Appearing in a generated answer depends on a sequence of conditions, and each type of investment affects a different part. The full sequence is in How to get AI to recommend your brand.
| Condition | What it resolves | Owned content | Third-party mentions |
|---|---|---|---|
| Discovery | The crawler reaches the page | ✅ Exactly this | ⚪ Indirectly, through a link |
| Retrieval | The document is selected from the candidates | ✅ Specificity comes from writing | ⚪ Marginal |
| Generation | The brand survives the writing stage | ✅ You control your page’s writing | ⚪ Outside your control |
| Entity recognition | The engine identifies you as a distinct entity | ⛔ Nothing on your own domain resolves this | ✅ The only path |
The final row separates the investments. Entity recognition is not built on your own domain: by definition, evidence that you exist as a distinct entity must come from something other than your own assertion. No volume of owned pages substitutes for that.
Structures observed in two competitors on 2026-08-06
⚠️ Read this as a structure observed on one day, not an explanation of causation. I have no controlled experiment isolating why one company is cited more than the other. I have a snapshot of what each built and the score on the same day. A correlation between structure and score does not establish a mechanism.
The most frequently cited name on my list. I fetched its sitemap.xml with raw curl and
counted 1,366 URLs, with 9 under the topic directory: 0.7% of the site. The main topic
page has fewer than 500 unique words, and its publishing cadence has been less than one
post per month since July 2025. By any content measure, this is a thin body of material.
What the same company has outside its own site: original research spanning multiple years, hosted by Poder360 and E-Commerce Brasil; two Band articles ranking it #1 of 10; its founder appearing on five third-party podcasts; and roughly 15 to 17 third-party domains naming, hosting or interviewing it. ⚠️ A caveat required by my own research: the Band listicle looks like a syndicated PR placement, and I could not confirm whether it was paid.
The runner-up. The opposite structure: 279 URLs across four sitemaps, a cluster almost
entirely devoted to the topic, with lastmod dates from 2026-04-12 through 2026-08-06 — around
four months, in two publishing bursts. No third-party coverage was found. ⚠️ My research
records this as unconfirmed, not absent: one search pass does not exhaust someone’s public record.
Their scores under the same rule
Using the deterministic rule on a declared list of 29 names, across 100 (question, engine) pairs, on 2026-08-06:
| Brand | Predominant structure | Pairs won, out of 100 |
|---|---|---|
| Conversion | Off-site entity recognition, thin owned content | 19 |
| GeoStack | Content-led, about 4 months, no confirmed third-party coverage | 15 |
| Brasil GEO | — | 14 |
| Criamente | — | 9 |
| Profound | — | 8 |
murmur.marketing | — | 0 |
⚠️ Three caveats belong with this table and are not optional. First, the top three are separated by less than 1.3 binomial standard errors. This is not a stable ranking; treating it as one is reading noise. Second, the most frequently cited name wins 19 and is absent in 81. That is a small part of an empty field, not command of the market. Third, the rule recognizes 29 names out of at least 447 extracted by the judge from the same captures. That is precisely why I do not publish percentage shares among competitors, only absolute counts with a denominator in the same sentence.
What this pair teaches, counterintuitively on both sides
Fifteen years of off-site entity recognition correspond to 19 pairs. Four months of content alone correspond to 15.
Both paths work. Neither covers more than a fifth of the field. 51 of the 100 pairs have no winner at all: the larger opportunity is in the 51 nobody wins, rather than the 19 somebody does.
This undermines both extreme versions of market advice. “Just produce content” ignores that the thinnest content presence in the research belongs to the most frequently cited name. “Just build authority” ignores that four months of publishing got within the noise range of the top position, without any confirmed third-party coverage.
There is a timing implication worth making explicit: the content-only path moves quickly, and my material suggests it has a ceiling. Entity work is slow and expensive — original research, PR and podcasts — and it does not buy the entire field either.
A dated case that helps guide allocation
Commercial context: Murmur’s SWAS model combines software with GEO strategy and execution. The historical baseline cited here predates the current guide collection. In that 2026-08-06 campaign, citation_kind = ausente (the original “absent” label) appeared in 800 of 800 captures, all with screenshots; this is not a current performance score. Outside the ten questions already naming the brand: 0 of 712.
The case I use to decide allocation is a company I own, and it is a failure. Fly Vet had 324 articles, counted on disk on 2026-07-02, its entire cluster published, and a homepage claiming a prominent position. In the 2026-06-27/28 measurement, ChatGPT recommended a competitor ahead of it. On eight category questions, Fly Vet was absent while that competitor was listed first. The content was done; the contest was lost.
That is money spent on the wrong condition, and the clearest inexpensive evidence that the two
investments are not interchangeable. For its sister company Fly Med, the first ChatGPT citation
of a /geo/ directory came at T+5 days after publication, with two pages cited in one query.
In the same probe, Fly Vet had no cited URLs in the defined sample. This is a dated result for one campaign and is not a current performance claim.
Anyone promising a deadline is guessing: in another case I follow, the first citation came at
T+28 in Perplexity and T+30 in ChatGPT.
A limit I must also declare about third-party mentions
My material contains a result that cuts against enthusiasm for disambiguation and entity work. Publishing it is an obligation.
Among the ten questions in my set that already named my brand — a denominator of 40 captures — the engine cited a domain containing “murmur” in 33, and only 5 were mine. The others belonged to established marketing agencies sharing the name: straightforward entity confusion.
But across the 360 captures of unprompted discovery, no “murmur” domain appeared, neither mine nor a namesake’s. These are two separable problems: name ambiguity and absence from the category. Solving the first does not move the second. Promising that it will assigns a gain to the wrong mechanism. That is the kind of promise I challenge in others, so I must challenge it in my own work first.
How to decide in practice
- Find the bottleneck before buying anything. The server log answers the discovery question; a question only your page can answer tests retrieval; uniform absence across all four engines points to entity recognition.
- If absence is uniform and the engine cites names similar to yours, the problem is entity recognition. No number of articles fixes it; the material above is the evidence.
- If you have fewer than a few dozen topic-specific pages, the problem is content. Buy retrievability first: it is cheaper, faster and measurable in weeks.
- If both conditions are reasonable and the score does not move, the question may be the problem. In the broad-question family of my set, the most frequently cited name appears 13 times in 120 captures. For broad questions, the engine explains a concept, not a vendor.
Frequently asked questions
Is a mention in a major publication worth more than ten articles on my site?
It depends entirely on which condition is blocked, so there is no universal answer. If the crawler cannot even reach your pages, the media mention does not fix that: it does not make your site retrievable. If the engine already retrieves your pages but never identifies you as a distinct company, ten more articles will change nothing, while a mention might. I avoid advice that treats the two investments as currencies convertible at a fixed rate.
How do I know whether absence is an entity problem rather than a content problem?
Look at uniformity and the kind of substitution. Content failures tend to be uneven: the brand appears for some questions but not others, or more in one engine than another. An entity failure is uniform across all four engines and has a characteristic signature: the engine fills the gap with the closest name it knows. In my case, 33 of 40 answers to questions naming my brand cited a domain containing “murmur,” and only five were my own; the rest belonged to namesakes.
How many pages are enough on the content side?
I have no defensible number and am skeptical of anyone who does. I have two observations that contradict each other if read as rules: the most frequently cited name in my research reached the top with nine pages on the topic, while the runner-up got close with 279. The honest reading is that the count is not the deciding variable. Both have a coherent body of material on a narrow topic, and both remained below an observed ceiling of one fifth of the field.
Does a third-party mention need a link to count?
A link helps discovery. Entity recognition comes from your name appearing alongside a description of what you do on a domain you do not own. That is why podcasts, interviews and research hosted by publications appear in the research alongside linked articles. The engine is developing an understanding of the company, and a name in third-party text contributes even without a clickable anchor.
If both paths stop around a fifth of the field, is either worth investing in?
Yes, and the argument lies in the unoccupied space. In 51 of the 100 pairs I measured, no brand on the list of 29 appeared dominantly; for broad questions, the engine explained a concept instead of recommending a vendor. The contest is therefore about establishing a position for questions that currently receive no brand answer, rather than replacing someone. The observed one-fifth ceiling says nobody has consolidated the field, not that it is closed.
Who wrote this, and disclosure of interest
Mateus Gomes operates murmur.marketing, a Brazilian GEO operation combining proprietary measurement software with specialist strategy and execution to grow SOV. The comparison and campaign results in this article are historical snapshots from 2026-08-06; they are intended to inform diagnosis, not represent current visibility or establish causality.
Conclusion
Owned content and third-party mentions address different conditions. The first resolves discovery and retrieval; the second is the only route to entity recognition, which by definition cannot be built on the domain seeking recognition. The structures observed on 2026-08-06 show both paths working, neither covering more than a fifth of a field where half the questions have no winner. The important decision precedes the purchase: identify the blocked condition. Buying more of what you already have is the most common waste in this category. To discuss a diagnosis for your situation, contact Mateus Gomes on LinkedIn.