Fly Vet wins comparison queries but loses when vets describe their problems: first in 69 of 228 ChatGPT questions in September 2026

Where it wins

69 of 228Spontaneous questions in which ChatGPT places Fly Vet first among its recommendations. One engine, one repetition, a run on 2 September 2026. This is its current position, not a trend: on 22 July it was first in 63 of the same 228 questions. The internal index records an unresolved discrepancy between that figure of 69 and a raw recount that returned 72.

→

Where it disappears

2 of 80Records in which ChatGPT recommends Fly Vet when a veterinarian describes a problem in the first person: 40 questions with 2 repetitions, in the targeted June 2026 campaign. For Claude, in the same subset, 0 of 80.

period

Content has been live since 22 May 2026, when the website’s /geo/ directory was opened to AI crawlers. Measurement ran from 27 June to 2 September 2026. The first-place figure compares 22 July with 2 September 2026, a span of 42 days. The repository has no Fly Vet campaign measurement from before June 2026: the first campaign on disk is from 27 and 28 June.

company size

Not recorded. Neither the company profile nor the work map provides revenue, employee count or a size category, and the case record deliberately leaves this field blank. The only scale claim approved for publication is the canonical statement fixed on 11 September 2026: more than 600 veterinarians served since 2018, cumulatively, counting anyone who purchased at least one paid plan. This is the company’s statement, not a murmur measurement.

question set

The fixed panel contains 300 questions. Its files dated 19 August and 2 September 2026 are byte-for-byte identical: 139 bottom-of-funnel, 121 middle-of-funnel and 40 top-of-funnel questions. Removing the 42 that name Fly Vet in the prompt, all bottom-of-funnel, leaves 258 spontaneous questions. This page uses the 228 present at both endpoints because the 2 September capture did not cover the whole panel: it returned 249 of 300 questions. The 228 break down into 80 agency combinations, 48 comparisons, 40 first-person problem queries, 25 outcome queries, 19 comparisons between two named competitors, 11 generic queries, 4 traffic queries and 1 alternative query. Of these 228, 50 name a competitor, so this is not an entirely unbranded set. Other subsets used here are the broad 100-question June campaign, the targeted 165-question campaign, and 80 records per engine for first-person problem queries.

engines

Measurement covers four engines: ChatGPT, Claude, Google AI Overview and Perplexity. The first-place figure, 69 of 228, refers only to ChatGPT, with one repetition and one run per date. Each response is assessed by a language-model judge, whose backend is stored in every record. Both endpoints of this window used the same judge, although not all intermediate runs did. In the June campaign, Google AI Overview never recommended Fly Vet: it appeared there as a source, in 81 of 100 responses.

Five days after publication, Fly Vet had zero citations and Fly Med had two

Fly Vet is the agency of the founder of murmur.marketing, a GEO (Generative Engine Optimization) engine in Brazil operated by Mateus Gomes. The interest here is therefore double and it is better to say so immediately: the party that measured is the party that executed the work and owns the measured company. This brings one advantage no client grants: the ability to publish what went wrong without asking anyone’s permission. Content went live on 22 May 2026 in an AI-crawler-accessible directory inside the agency site. The first reading, five days later, was poor.

In the 27 May 2026 probe, no page from Fly Vet’s directory appeared as a cited source in answers. Zero. On the same day, with the same method and in the same run, its sister company had the opposite result: ChatGPT cited two Fly Med directory articles, fly.med.br/geo/clinicweb-alternativas in first position and fly.med.br/geo/o-que-e-builder-de-site-medico-em-consultorio-medico in third, with ChatGPT’s citation tag on the URL. Same machine, same week, opposite outcomes. This is not a recipe.

The Fly brand appeared in 32 of the probe’s 59 successful captures, but through older assets: blog and service pages. Only 2 of those mentions were URLs from the new directory, and both belonged to Fly Med. A brand mention is not the new content working, and the measurement requires that distinction first. The instrument’s state also matters: 65 questions were attempted and 59 captured—40 in Google AI Overview and 19 in ChatGPT—while Claude delivered none because its session failed authentication that day.

There is no earlier Fly Vet measurement in this repository. The founder states that in January 2026 the agency did not appear in any ChatGPT answer, based on his manual check of questions he asked that day. There is no screenshot, stored campaign or recorded universe, so this zero remains a stated claim and receives no denominator in this article.

First in 69 of 228 ChatGPT questions, and this is not an increase

Since July, Fly Vet has been remeasured on a fixed panel of the same questions: the 19 August and 2 September 2026 panel files are byte-for-byte identical, with 300 questions. Removing the 42 that name the brand in the prompt leaves 258 spontaneous questions. The calculation uses the 228 present at both dates because the 2 September capture returned 249 of the 300. What is published here is a level, not progress: in 228 spontaneous questions about hiring a veterinary marketing agency, ChatGPT places Fly Vet first in the recommendation in 69. One engine, one repetition, one 2 September 2026 run. The internal index carries an unresolved discrepancy: the owner fixed the figure at 69 on 9 September while a raw recount returned 72. The discrepancy travels with the number so it cannot become a hidden detail.

It is tempting to call that an increase because 22 July had 63 in the same 228 questions. It is not an increase. Of the 228, 35 began bringing Fly Vet first and 29 stopped; the net six is far smaller than the 64-question back-and-forth. The house standard treats only a change above three to four percentage points in the same set as real movement, and this does not reach it. The honest reading is a high level with fluctuation, not a brand moved into its current position.

There is also an instrument caveat against any trend reading, including ours. Internal forensics on 10 September 2026 found the 2 September ChatGPT capture truncated before the panel ended, with the missing block being the weakest-performing one, and concluded that no time series is valid until this is corrected. It also found a judge change in rounds in the middle of the window. What supports the number was also measured: the two comparison endpoints, 22 July and 2 September, used the same judge. So 69 remains a legitimate level in the 228 questions present on both dates; it remains not a time-series point.

Each engine also plays a different game. In June’s broad 100-question campaign, ChatGPT recommended Fly Vet in 30, Claude in 41, and Google AI Overview in none. Google showed Fly Vet in 81 of the 100 answers, but as a source the engine had read, never as a recommendation the engine made. Treating the engines as one produces an average of things that are not the same.

In first-person pain queries, Fly Vet is recommended 2 times in 80 and never by Claude

The costliest gap is where the client first appears. When a veterinarian describes their pain in the first person—such as “my clinic does not appear on Google”—Fly Vet practically does not exist: 2 of 80 records in ChatGPT and 0 of 80 in Claude, measured through 30 June 2026 in 40 questions with 2 repetitions per engine. The AI answers there are SimplesVet, Vetus, MarketVet and Agência Pet. The agency wins the comparison, which is asked by someone who already knows it, and loses the pain query, which is asked by someone who does not. The campaign producing these cuts is marked partial in the work map, with roughly half judged; the two cuts cited here are complete within it.

We opened the two hits one by one because the temptation was to make a recipe from them. They are questions fvm2177 and fvm2179, both in the first repetition, both first-listed, and both absent when the same question ran again on the same day. Both came from the sentence “I want to fill my veterinary clinic’s appointment book,” a positive request rather than a complaint. The tempting reading would be that asking positively works better than complaining, but the data rejects it: the other three variants of that sentence were absent in both repetitions, and the other two positive sentences were 0 of 10 each. Two in 80 is noise, not a pattern. The finding is that there is no finding, and publishing that is more valuable than inventing an explanation.

There are two further results. In a direct comparison—Fly Vet or another agency—there were 0 recommendations in 28 ChatGPT records, a figure the internal index itself marks pending a new count. There is also the pattern of doing work without winning: competitor EvolueVet was cited in 28 of 100 ChatGPT answers in the June campaign; within those 28, Fly Vet was recommended in 10; and in 8 “best veterinary marketing agency” questions Fly Vet was absent while the competitor came first. The agency content cluster was already published when this was measured, pushing diagnosis beyond content and into the entity.

324 articles in five collections, a pilot batch born from the blackout, and five claims the operation removed from itself

The order was measurement before writing. First came a broad 100-question campaign in three engines on 27 and 28 June 2026, followed by a directed 165-question campaign through 30 June that separated question types: first-person pain, comparison, outcome and brand. Only then came content: 324 articles in five thematic collections of 71, 81, 16, 137 and 19 pieces. The on-disk count matches the live count, 324 to 324.

The 2 July pilot batch of 33 articles came directly from the pain-query blackout measured two days earlier. Every article went through independent skeptical review, article by article, which caught real fabrications left by writing and first review: a fabricated market statistic, a factual contradiction between sibling articles about a competitor’s CRM, and an unsupported claim. This batch was checked one by one in production after publication and closed at 33 of 33.

On 11 September 2026 a facts sweep entered the entire site: one client total, one title for the founder, one category. It removed claims the operation itself had repeated without support, including “more than 500 clinics” in 91 files, “more than 100 clinics” in 68, “CEO of Fly Vet” in 79, “SaaS digital agency” in 44 and “largest agency” in 27. What remained standing was “more than 600 veterinarians served since 2018,” cumulative. Correcting what the company says about itself is part of the work because that is what the engine reads.

The sweep did not finish, and that must be written too: 26 “Fly Vet leads” claims remain in 12 ranking articles because they sit in the direct answer at the top of the page and rewriting them by regular expression would break the prose. They were left for the next round. While they remain, the page continues crowning itself in 12 places that the measurement does not support.

The average deal value figure on the card is the owner’s statement, not murmur measurement

This case card shows business outcomes: average deal value and qualified-lead conversion. Those numbers did not come from any murmur measurement. They are a statement by Fly Vet’s owner, who is also murmur.marketing’s founder, made on 11 September 2026; he is therefore a legitimate source for his own revenue and nobody else’s. They have no period or comparison base, and nothing in the repository connects average deal value to AI-content work. They also do not yet have a row in the operation’s released-numbers index, the condition for a number to appear in a public piece. That is why they are not repeated in this text: the card is the owner’s statement, the body is measurement, and the two are not mixed even by accident.

The boundary in writing is this. What murmur measured at Fly Vet: 69 of 228 spontaneous questions with Fly Vet first in ChatGPT on 2 September 2026, with the recount discrepancy declared; 2 of 80 ChatGPT records and 0 of 80 Claude records in the first-person pain cut; 0 of 28 in direct comparison; 30 and 41 of 100 recommendations in ChatGPT and Claude and 81 of 100 as a source in Google in the June campaign; and zero new-directory pages cited five days after publication, against two from the sister company in the same probe. What the company declares is on the card. Until card figures have a period, denominator and an index row, they stay on that side of the line. Stating this explicitly is better than leaving the reader to connect the two on their own.