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.
A brand is recommended by a generative engine when four conditions are met at once — discovery, retrieval, entity and drafting — and they fail in order: no text technique repairs a site the crawler cannot reach, and no page volume repairs a model that does not know the company exists as a distinct thing. The title question is in the buyer’s voice; the answer below is in mine, and it begins by separating two things the market merges.
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 dated campaign below predates the current guide collection; it is included as a historical baseline, not as a current performance claim.
This is the broad page in the method family. Each of the twelve questions hanging from it has its own page; the full map is in “The twelve questions in this family”.
Summary
- Recommended ≠ cited. They are distinct states within the same answer, with different remedies. The four-state scale is open in Cited as a source or recommended by AI; this page uses it rather than re-explaining it.
- Four conditions, in the order they bite: discovery (the crawler arrives), retrieval (the engine retrieves the page), entity (the model separates the brand from namesakes), and drafting (the text survives machine rewriting).
- Order matters more than the list. In my own case the bottleneck was the third: in 33 of 40 captures of questions that already contained the name, engines cited a domain containing “murmur” that was not mine.
- The category has space, and this is measured. In 51 of 100 (question, engine) pairs run five times, no brand from the declared list of 29 names reaches a majority of repetitions.
- Two paths work, and neither exceeds 20%. Conversion, with more than fifteen years of domain history and authority beyond its own site, wins 19 of 100 pairs and is absent in 81. GeoStack, with about four months of concentrated publishing, wins 15. Deterministic text rule over a declared list of 29 names, 2026-08-06.
- There is no honest timeframe to promise. The first citation record for a
/geo/directory in ChatGPT was T+5 days at Fly Med and, at Fly Vet in the same probe and on the same day, zero. In a third case, from a client universe I cannot name, it was T+28.
Recommended is not cited — and this is where most projects choose the wrong remedy
A generated answer puts a brand in one of four states, and treating them as a Boolean is the error that makes a project work in the wrong place for months:
| state | what happened | what to fix |
|---|---|---|
| recommended | the engine presents it as the choice in the body | broaden question coverage |
| source | the engine linked the page and did not discuss it in the body | page drafting |
| mentioned | the name appears without becoming the choice | positioning |
| absent | nothing | discovery or entity |
“Source” and “recommended” are the two most often confused, and their difference is not one of degree: a page can be read, linked and ignored when the engine writes the sentence that decides. The full distinction, with captured examples, is in Cited as a source or recommended by AI. The operational point here is: before choosing what to do, you need to know which of the four states the brand is in, and that is not discovered by impression. It is discovered through How to measure whether AI cites your brand.
The four conditions, in the order they bite
1. Discovery — the crawler must arrive, and it gets no second chance
An AI crawler reads the HTML the server returns and leaves. It does not execute JavaScript and does not reload a page waiting for content to appear. Vercel published a reading of its own network logs in The rise of the AI crawler: none of the main AI crawlers renders JavaScript today. The detail almost nobody repeats is that they download script files without executing them: 11.50% of OpenAI crawler requests and 23.84% of Anthropic crawler requests. Downloading is not running. Text that only comes into existence after a script runs does not exist for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, CCBot or Google-Extended.
Discovery has three paths, and only three: an internal link from an already known page, sitemap.xml declared in robots.txt, and active submission to an index. Content with none of the three is orphaned — it can be technically impeccable and never be seen.
Two costly traps belong here. First: blocking classic search while allowing only an AI crawler is cloaking, the fastest way to burn the entire domain. Generative engines rely on search indexes — ChatGPT with search anchors itself in Bing’s index, making Bing Webmaster Tools a lever almost nobody uses. Second: llms.txt is good faith, not a discovery channel. No engine has yet confirmed consuming it as a crawling route. It costs nothing and does no harm to publish, but relying on it as an appearance mechanism bets on an unconfirmed channel. This position is dated and falsifiable the day an engine states otherwise.
2. Retrieval — being reachable is not being retrieved
Once the site is reached, the engine still needs to choose that page among candidates when the question arrives. Specificity decides here: a page answering the buyer’s exact wording is retrieved; an institutional page that talks about everything is retrieved for nothing. The house formulation is facets, not clones — a leadership question is won with several pages, each owning a distinct question, not one excellent page trying to cover all of them.
A counterexample is worth more than the argument. In reverse engineering I ran on 2026-08-06 with curl over public surfaces, Conversion, the name the text rule saw most, had 1,366 URLs in sitemap.xml, only 9 under /geo/ — 0.7% of the site — and published fewer than one topic post a month. Page volume is not the variable being measured. How much to publish, and why the answer is not “as much as possible,” is covered by How many pages do I need to publish for AI to recommend me?.
3. Entity — the condition no page volume resolves
This is the stage that almost never appears on provider checklists, and it is where I myself am stuck.
Before any technique, the model needs to know the company exists as a distinct thing in the world. When it does not, it does not make a careless error: it fills the gap with the nearest namesake. In my 2026-08-06 campaign, the cut is this — in the 10 questions that already contained my name, × 4 engines, denominator 40: in 33 of those 40 captures engines cited a domain containing “murmur,” and 5 were mine. The others are established marketing agencies in the same category, in other countries. In the other 360 captures from the same run, no “murmur” domain appeared — neither mine nor theirs.
Name ambiguity and nonexistence in the category are two different problems, and confusing them is costly. Disambiguating a name does not improve cold discovery; promising that it does attributes gain to the wrong mechanism. The sign a brand is stuck here is a done but not winning pattern: content is published, the cluster is complete, and AI still recommends the competitor. I saw it in two different businesses, including mine. At Fly Vet, the home says “first and largest agency,” the cluster is live, and ChatGPT recommended EvolueVet ahead of it: EvolueVet was cited in 28 of 100 ChatGPT answers, and in 8 “best veterinary marketing or traffic agency” questions Fly Vet was absent while the competitor was listed first. Content done, game lost: the bottleneck had shifted from content to entity.
What this stage needs is not content: verifiable existence beyond your own site — coherent profiles, a unique and consistent name across surfaces, third-party coverage. Conversion’s mechanism illustrates it: it has research hosted by Poder360 and E-Commerce Brasil, two Band articles that crown 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 is a sign on the door; what the engine reads is outside. The survey’s caveat remains: the Band listicle looks like a PR-syndicated placement and it was not possible to confirm whether it was paid.
4. Drafting — what survives when the machine rewrites
Only in the fourth condition does text form matter, and it matters specifically: the engine does not copy the page; it drafts a new answer from it. What survives this rewriting was measured in academic publication — Aggarwal et al., Generative Engine Optimization, presented at KDD 2024 and published as arXiv 2311.09735, which built GEO-bench with 10,000 queries. Three drafting interventions and their measured effect against that benchmark: citations to sources, +115%; statistics, +41%; direct quotations, +30%.
Two honest caveats belong with these numbers. First: they are effects against an academic benchmark, for a particular set of engines and queries — not a promise of commercial results. Second: they describe drafting, the difference form makes after a page has already been retrieved. A document the engine cannot reach is not improved by any citation. How to structure content to be cited by AI covers how this becomes page structure.
What category measurement says about where there is room
All figures below come from the same 2026-08-06 campaign, under a deterministic text rule over a declared list of 29 names — a limitation to state first, not last. Across 100 (question, engine) pairs, each run five times, counting pairs where the name appeared in most repetitions:
| brand | pairs won (out of 100) |
|---|---|
| Conversion | 19 |
| GeoStack | 15 |
| Brasil GEO | 14 |
| Criamente | 9 |
| Profound | 8 |
| no brand with a majority | 51 |
Two caveats travel in the same sentence as this table. The first three are separated by less than 1.3 binomial standard errors over 100 pairs — not a stable ranking, so reading it as a podium goes beyond the data. And the name with 19 pairs is absent in the other 81. This page declares nobody the market owner.
The historical comparison shows two different routes to visibility. Conversion’s long-standing off-site authority corresponded to 19 of 100 pairs; GeoStack’s concentrated content footprint corresponded to 15. The survey did not isolate causation, and its 2026-08-06 Murmur baseline predates the current guide collection. The practical takeaway is to diagnose whether a brand needs stronger entity recognition, more retrievable owned content, or both before choosing an intervention.
One last context line: in 216 of 400 captures from the one-repetition run, none of the 29 names appeared, and the median number of distinct brands per capture was zero.
Why I do not promise a timeframe
At Fly Med, publication was 2026-05-22 and the first citation of two /geo/ pages in ChatGPT was 2026-05-27 — T+5 days, with ?utm_source=chatgpt.com in the URL. In the same probe and on the same day, its sister company Fly Vet had zero /geo/ URLs cited. Same methodology, same date, opposite results. In a third client universe I cannot name, the first guide citation came in Perplexity at T+28 and in ChatGPT at T+30.
T+5 in one case, zero that same day in another, T+28 in a third. Anyone promising timing is guessing, even when they get it right. What can be done is measure the real window afterward rather than estimate it in advance; see How long does AI take to start citing my site?.
The twelve questions in this family
This page answers the broad question. Each item below is a closed question with its own page and points back here.
| question | what it resolves that this page does not |
|---|---|
| How to appear in ChatGPT as a company | the operating sequence in one engine, in the buyer’s literal wording |
| How to be cited by ChatGPT and Perplexity | all four engines side by side — they disagree |
| What to do when AI recommends my competitor | displacement when a name already occupies the seat |
| How to structure content to be cited by AI | page form: what survives machine rewriting |
| What to do when AI Overview caused my organic traffic to fall | separating click loss from citation loss |
| How many pages do I need to publish for AI to recommend me? | the volume ceiling and why the most cited name has nine pages |
| How long does AI take to start citing my site? | the measured window, with both sides of the data |
| How to appear in Perplexity | the engine with the most explicit citation mechanics |
| Own content or a third-party mention | allocation between conditions 3 and 4 |
llms.txt: does it work for AI citation? | the file that became folklore, and what is known about it |
| How to do local GEO for a neighborhood business | the geographic slice, where the question changes nature |
| My boss asked why ChatGPT does not recommend us | the same question in the voice of the person answering the board |
Frequently asked questions
What is the difference between being cited and recommended by AI?
Being cited means the engine links or names the page; being recommended means the engine presents the brand as a choice in the answer text. They are different states with different remedies: a brand that appears as a source but not as the choice has a drafting problem, not a discovery problem. One “mention” metric merges the two and returns the same number for opposite problems, so the four-state classification — recommended, source, mentioned, absent — is the usable minimum.
Does being recommended require being the biggest name in the category?
No. Across the 100 question-engine pairs run five times each on 2026-08-06, the most seen name under the declared 29-brand rule wins 19 and is absent in the other 81. The first three are separated by less than 1.3 binomial standard errors, preventing a podium reading. Size is not one of the four conditions and substitutes for none of them.
Which of the four conditions is most expensive to unlock?
Entity, by far, and it is the only one that does not yield to a text budget. Retrieval can be checked in minutes and repaired on your domain. Entity is the engine resolving your company as a distinct thing, built on surfaces that are not yours. In 33 of the 40 captures of questions containing my name, the engine cited a domain similar to mine and only 5 were mine. No amount of published pages changes that number.
How do I find which condition is blocking my brand?
Eliminate from the top down because each condition leaves a different signature. Check discovery in server logs: if no AI crawler seeks your URLs, nothing below matters yet. Check retrieval by asking the engine something only your page answers. Recognize drafting when the page is cited as a source but the brand is not in the body. Entity is the residual diagnosis and most likely when absence is uniform across all four engines: there was no lost comparison, there was failure to enter the candidate set.
Must the four conditions be solved in order, or can they be worked in parallel?
The order is dependency, not scheduling. Discovery and retrieval are preconditions: a document the crawler cannot reach is not improved by drafting. Entity breaks the sequence because it needs verifiable existence outside your site; nothing published on your domain substitutes for it. At Fly Vet, 324 articles were counted on disk on 2026-07-02 and the cluster was live, but 2026-06-27/28 measurement returned absence in 8 category questions with the competitor listed first. Money was going to the wrong condition.
Who wrote this, and the declaration of interest
Mateus Gomes operates murmur.marketing, a Brazilian GEO operation combining proprietary measurement software with specialist strategy and execution to grow SOV. Historical measurements in this guide predate the current guide collection and are not current visibility claims.
The historical 2026-08-06 baseline is part of the method record: 88 string matches in 800 captures all fell within the ten questions that already contained the brand name. That is why brand-naming prompts must be reported separately from unprompted discovery; the baseline predates the current guide collection and is not a present-day visibility score.
Conclusion
Getting a brand recommended by AI means satisfying four conditions in order: the crawler must reach server-delivered HTML, the engine must retrieve the right page for the right wording, the model must know the company exists as a distinct thing, and the text must survive rewriting. The order decides spending: drafting technique for an entity problem is money in the wrong place. The category has measured space — 51 of 100 pairs without a majority — and neither of the two paths that work exceeds one fifth of the field. There is no timeframe to promise; there is a window to measure. To discuss measurement for your case, contact Mateus Gomes on LinkedIn.