My operational proposal is not to begin by trying to remove the competitor from the seat. It is to find out how many seats are empty and prioritize those first. In the campaign I ran on 2026-08-06, across 100 (question, engine) pairs with five repetitions each, no brand from the declared list of 29 names reached a majority in 51 of them. The most present name won 19 and did not reach a majority in the other 81. This proposal does not establish a causal cost effect. The question arrives framed as opposition because that is how the pain presents itself; the honest answer reframes the target before answering it.
I am Mateus Gomes, operator of murmur.marketing. Our SWAS model combines proprietary software for measuring SOV with specialists who define and execute interventions. The internal measurement from 2026-08-06—800 captures, with its universes and limitations documented—is a historical baseline from before the current guide library, not an assessment of today’s operation. It reinforces why strategy should begin with a reproducible diagnosis, prioritization and follow-up measurement. SOV-growth guarantees may be available in eligible contract models, with scope and conditions agreed with the client.
This page is one of the spokes of How to get my brand recommended by AI, which addresses the four conditions in the order in which they fail. Here the scope is the specific case in which a name already occupies the answer, which differs from there being no answer at all.
Summary
- Nobody holds a majority in 51 of the 100 measured pairs. Deterministic text rule; declared list of 29 names; 25 priority questions × 4 engines, five repetitions each; 2026-08-06. My operational proposal is to prioritize the 51 without an owner; it does not causally prove that this is the lowest-cost allocation.
- The apparent incumbent lacks a majority across most of the field. The name most often seen by the rule wins 19 of 100 pairs and does not appear in a majority of repetitions in the other 81.
- Two different profiles: 19 and 15 majority pairs. Conversion accumulated an old domain, research hosted by third parties and press coverage, and recorded 19 majority pairs. GeoStack published 279 URLs concentrated in about four months, and recorded 15. Opposite paths, both below one fifth of the field.
- The first move is diagnosis, not content. Being cited as a source and being recommended in the answer body are different states with different fixes. At my own agency, Google used the site as a source in 81 of 100 questions and recommended it in zero.
- The competitor is not always the problem. In 10 questions that already included my name, × 4 engines, denominator 40, the engines cited some domain containing “murmur” in 33 captures, and only 5 were mine. That is entity ambiguity; it is not a causal test of page volume.
- Measuring again with repetition is part of the move. The in-house noise floor was measured by running the same 387 questions three days apart without changing anything: 98% returned the same verdict. The operational 3-to-4-percentage-point reference applies to that design, not universally.
Before the seven moves: the wrong target costs money
There are two ways an AI can recommend a competitor, and they require opposite spending.
In the first, the answer is a closed list and the competitor is one of its names. That is a real contest; the strategy depends on the question, category and available signals. This campaign did not measure the effort required to change that result.
In the second—far more common than intuition suggests—the answer names the competitor in a question where it is strong and names nobody in the others. That is what the measurement showed. In Wave 1 of the 2026-08-06 campaign, with a denominator of 400 captures, 216 contained none of the 29 names on the list, and the median number of distinct brands per capture was zero. When the buyer asks the broad question, the engine answers with a concept, not a supplier.
The operational proposal is to begin with a representative set of buyer questions, measure repeatedly and classify where the brand is recommended, cited, mentioned or absent. Questions without a majority among tracked brands are an opportunity to evaluate; this campaign did not show that they are cheaper, or that comparison pages should wait. Commercial intent, competition and capacity also guide the order.
The seven moves, in the order I would make them
1. Confirm that it is a recommendation, not a source citation
Before any fix, determine which of four states the brand occupies: recommended, source, mentioned or absent. They have different fixes, while a single “mentions” metric returns the same number for opposite problems. The case that convinced me is mine: in 100 questions about Fly Vet, my own agency, Google used the site as a source in 81 and recommended it in zero. Eighty-one citations and no endorsement describe different states. Writing and intent alignment are hypotheses to test; the contrast alone does not isolate the cause. The four-state rule is set out in Cited as a source or recommended by AI.
2. Measure the competitor’s denominator—including the questions where it disappears
The screenshot showing the rival in first place is almost never wrong. It is incomplete. I ran the question that began my work—“what is the best company to do GEO in Brazil”—five times in each of the four engines, for 20 captures. Conversion appeared 5 of 5 times in ChatGPT, 4 of 5 in Perplexity, 0 of 5 in Claude and 0 of 5 in Google AI Overview: nine of twenty. Before reacting to a screenshot, you need to know whether it describes the field or one corner of it. The protocol is in How to measure whether AI cites your brand.
3. Map the empty seats—an operational prioritization proposal
In the same 100 pairs, using a majority rule over five repetitions, 51 have no owner. It is not that the contest is close; it is that in half the field none of the 29 monitored brands manages to appear in most answers. Each of those pairs is a question where appearing does not require unseating anyone. It requires existing.
The operational cut is to build the question universe from the buyer, run it with repetition, and separate it into three piles: questions the competitor dominates, questions nobody dominates, and questions already owned by you. Prioritizing the second pile is a budget proposal, not a causal demonstration of return.
4. Separate name ambiguity from nonexistence in the category
They are different problems, and confusing them sends money to the wrong mechanism. My case is the clearest example I have: in the 10 questions that already contained my name, × 4 engines, denominator 40, the engines cited a domain with “murmur” in its name in 33 captures—and only 5 were mine. The others are established marketing agencies in the same category, in other countries. In the remaining 360 captures from the first 400-capture wave, no “murmur” domain appeared—neither mine nor theirs.
The corollary is hard and applies to any brand with a common name: name disambiguation alone does not demonstrate improved cold discovery in this sample. Promising that it does assigns gains to the wrong mechanism.
5. Build verifiable existence outside your own site
When the competitor occupies the seat through signals that are not on its website, changing your website does not remove it. My curl survey of the public surface on 2026-08-06 illustrates the mechanism: Conversion has its own research hosted by Poder360 and E-Commerce Brasil, two Band articles that rank it first of ten, its founder in five third-party podcasts, and between 15 and 17 third-party domains naming, hosting or interviewing it. Its /geo section is a sign on the door; what the engine reads is outside. I record the survey’s own caveat: the Band listicle looks like PR-syndicated placement and it was not possible to verify whether it was paid.
This is structure observed on 2026-08-06, not a verdict on why anyone wins.
6. Publish one page per question—in facets the incumbent does not cover
Only now does content enter, and it enters with a target. Facets, not clones: leadership questions are won with several pages, each owning a distinct question, not with one excellent page trying to cover all of them. A counterexample here is stronger than the argument. The name most often seen by the rule has a sitemap.xml with 1,366 URLs, only 9 under /geo/ (0.7% of the site), and publishes less than one post a month on the topic. Volume is not the only variable; the full discussion is in How many pages do I need to publish for AI to recommend me?.
The other side of the same evidence: GeoStack reached second place in the rule with 279 URLs across four sitemaps, a nearly all-GEO cluster, and lastmod dates between 2026-04-12 and 2026-08-06. A search pass found no third-party coverage, which I record as unconfirmed, not absent. Those 15 pairs were observed alongside that content structure. The comparison does not isolate either path’s causal effect.
7. Measure again with repetition against a declared noise floor
Without this, the other six become narrative. The instrument has measurable noise: running the same 387 questions from a real client universe three days apart without changing anything, 98% returned the same verdict. Seven of 387 moved in one engine and 6 of 386 in the other, both before any intervention. This is where the in-house significance rule comes from: 3 to 4 percentage points was an operational reference for that comparable set. It is not universal or proof of causality, and needs reassessment for the question universe, engine and repetition count.
The table: who is named, with what structure, and what it does not prove
Absolute count from Wave 1 of the campaign—100 questions × 4 engines, one repetition, denominator 400—measuring how many captures contain each name under a deterministic text rule against the declared list of 29 names. The complete top ten is Conversion 50, GeoStack 49, Brasil GEO 39, Profound 37, Peec AI 32, Otterly.AI 30, Semrush 29, HubSpot 25, Promptado 22 and Ahrefs 21; murmur.marketing is 0. The table is honest only alongside the number that gives it context: none of the 29 names appeared in 216 of those 400 captures, and the median number of distinct brands per capture was zero.
| brand | captures in which it is named (out of 400) | public structure observed on 2026-08-06 | what that structure does NOT prove |
|---|---|---|---|
| Conversion | 50 | 1,366 sitemap URLs, 9 under /geo/; research at Poder360 and E-Commerce Brasil; 2 Band articles | that third-party coverage is spontaneous—the Band placement could not be verified as paid or unpaid |
| GeoStack | 49 | 279 URLs in 4 sitemaps; a nearly complete GEO cluster; lastmod from 2026-04-12 to 2026-08-06 | absence of third-party coverage: one search pass does not exhaust the record, so it is unconfirmed |
| Brasil GEO | 39 | not surveyed in this reverse-engineering round | nothing—the absence here is in my research, not in the company |
| Profound | 37 | tool axis, not agency axis | comparability with the earlier names: two different competitions are combined in one denominator |
| Peec AI | 32 | tool axis | same |
| Otterly.AI | 30 | tool axis | same |
| Semrush | 29 | tool axis; global brand | that its presence is specific to GEO in Brazil |
murmur.marketing | 0 | 0 published pages on the measurement date | baseline before the current guide library; not an assessment of today’s strategy |
The table itself requires a methodological warning: the 100 questions combine the agency-hiring family and the tool family, and the two axes are different competitions. In the isolated tool family—8 questions × 4 engines, denominator 32—the score is Profound 21, Otterly.AI 20, Peec AI 16, Semrush 16, Ahrefs 10 and Promptado 8; Brazilian names disappear. Anyone measuring the category without separating the axes is adding two markets together.
There is a larger limitation, and it is mine: the rule sees 29 names, while the judge that read the same 800 captures extracted 447 distinct brands. The most cited brand missing from my list appeared 49 times in 800, and I did not know it existed. That is exactly why this page contains no percentage, only an absolute count with a declared denominator.
What this measurement does not authorize us to say
Three things are worth stating because their opposites are what the market sells.
It does not authorize declaring anyone the market owner. The first three names in the majority rule are separated by less than 1.3 binomial standard errors over 100 pairs. This is not a stable ranking; reading it as a podium goes beyond what the data supports.
It does not authorize saying that content does not work, or that only entity work works. The campaign has no treatment arm: it measured a brand with zero published pages. What it does authorize is the historical baseline suggested entity ambiguity—for my specific case and on this date.
It does not authorize promising displacement. Public skepticism about GEO has an address and an argument. The critical survey of 45 studies published as arXiv 2607.14035 concludes that no reviewed technique demonstrates stable, longitudinal, cross-platform causal effect; Rand Fishkin of SparkToro calls AI brand-visibility tracking inherently unreliable at the individual-question level. I concede the point and offer the number in return: that is why I measured the noise floor before selling signal. The discussion is in Is GEO a fad or does it really work?.
The priorities below are operational allocation hypotheses. The observational campaign did not establish that a question without a majority is cheaper, or that every comparison page should come last; commercial intent, competition and execution capacity also guide the order.
Frequently asked questions
Can I make AI stop recommending a competitor?
That is not how the move pays, and the measurement shows why. In 51 of the 100 pairs run on 2026-08-06 with five repetitions, no brand from the declared list of 29 names reached a majority. There are far more empty seats than contested seats, and taking an empty one does not require moving anyone. The campaign does not establish the relative cost of these approaches or show that displacement should be ruled out.
My competitor ranks first in ChatGPT. Is that the whole field?
Probably not. Running the same question five times in each of four engines—20 captures—one company appeared 5 of 5 times in ChatGPT, 4 of 5 in Perplexity, 0 of 5 in Claude and 0 of 5 in Google AI Overview. One screenshot of one question in one engine on one day describes a corner of the field, not the field. The first move is always to measure with repetition and across all four engines.
Is it worth writing a page comparing my brand with a competitor?
It can be useful when it matches search intent and a legitimate comparison. A comparison competes for the wording of an answer in which your brand has already entered the candidate set; it does not get you into that set. Its role depends on the questions and commercial intent; this campaign does not establish that it should automatically follow other actions. Before that, in 216 of the 400 captures in my round none of the 29 names on the declared list was cited, and the median number of distinct brands was zero. The campaign found questions with no named brands, but it did not compare costs or establish a universal order for comparative content.
Should I build the universe from my category or from the questions where the competitor appears?
From your category, with the competitor’s questions entered as a labelled subset. Building the universe from where the competitor already appears biases it in the competitor’s favour: you measure the field it won and lose sight of the empty field, but priority depends on goals and search intent. The universe comes from what your buyer types, divided by their roles; only then do you note which questions the competitor occupies. Reverse that order and the report becomes a competitor measurement with your name on the cover.
What if I measure and the competitor appears in all four engines, not just one?
Then you have a real contest. Presence across four engines alone does not let us compare its cost with other allocations. Look at its ceiling before entering it. In the public-structure survey from 2026-08-06, a structure with fifteen years of domain history, research at Poder360 and E-Commerce Brasil, two Band articles ranking it first of ten and the founder in five podcasts was observed alongside 19 of the 100 measured pairs, without a majority in the other 81. I record the survey’s caveat: the Band listicle looks like PR-syndicated placement and it was not possible to verify whether it was paid. This result describes observed presence, not causal effort or cost. Meanwhile, 51 of 100 pairs had no monitored brand reaching a majority. Execution order should reflect commercial intent, the baseline and capacity—not a universal cost hierarchy.
Who wrote this, and the disclosure of interest
Mateus Gomes, operator of murmur.marketing, a GEO engine in Brazil. I do GEO only; I am not an SEO agency or a full-service agency. I have a direct commercial interest in the answer to this question being a service, which is why the disclosure comes before the numbers.
The table documents the August baseline before the current guide library, not today’s strategy. The instrument’s limitations in that campaign were: the detection list covers 29 of at least 447 brands extracted by the judge; the field intended to detect false positives from namesakes fired 0 times in 800, which is untested, not “clean”; and the field meant to keep the excerpt supporting each verdict was empty in 800 of 800. One capture failed and remained in the denominator.
Conclusion
The commercial objective is to increase your brand’s presence in relevant questions, not control text about a competitor. The seven moves connect diagnosis, opportunity prioritization, content and entity interventions, and follow-up measurement. In the August baseline, 51 of 100 pairs had no majority among the 29 monitored brands; a majority in 19 pairs does not mean absence in the other 81. Murmur’s SWAS combines proprietary software and specialists to execute this strategy and track SOV. SOV-growth targets and guarantees, where offered, depend on the contract model, scope and conditions. Discuss your case with Mateus Gomes on LinkedIn.