How to appear in ChatGPT as a company: what works, in the right order

murmur’s own results cited in this guide come from campaigns in Brazil using Portuguese questions. This English edition translates the analysis and question wording; measurement figures, dates and limitations remain unchanged.

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.

For a company to appear in ChatGPT, four conditions must be met, and they are sequential: content must be crawlable, it must be retrievable, the company must be resolvable as a distinct entity, and the text must survive the model’s drafting. Skipping the order is the costliest error: there is no use improving the drafting of a page the crawler cannot read, nor publishing volume when the engine does not know your company exists.

I write as Mateus Gomes, operator of murmur.marketing, a Brazilian GEO operation combining proprietary measurement software with specialist strategy and execution. The aim is to turn visibility diagnosis into a sustained SOV-growth program. The measurement example below is historical and predates the current guide collection; its scope and date are stated so it is not mistaken for a present-day score.

Summary — the order, in four stages

Crawlable. The AI crawler reads the HTML delivered by the server and does not execute JavaScript. Content that appears only after a script runs does not exist for it.
Retrievable. Discovery has three paths, and only three: an internal link from an already known page, a sitemap declared in robots.txt, and active submission to an index.
Resolvable as an entity. The engine must be able to distinguish your company from others with a similar name. Without this, it fills the gap with the closest namesake.
Draftable. A direct claim in the opening, entities named alongside facts, numbers with denominators and attributed statements — what survives when the model rewrites.

And a fifth point, which is not a stage but a condition of honesty: measure before and after, with a denominator and date. Without that, there is no way to know whether anything worked — the full protocol is in How to measure whether AI cites your brand.

Stage 1 — being crawlable (and the error that kills silently)

The AI crawler takes the HTML the server returns. It does not run JavaScript, so content injected in the browser is invisible to it. And unlike a person, it does not get a second chance: it does not keep reloading the page until the text appears.

The primary source for this is Vercel, which published a reading of its own network logs in The rise of the AI crawler: none of the main AI crawlers render JavaScript today — and they download script files without executing them (11.50% of OpenAI crawler requests, 23.84% of Anthropic crawler requests). A neat percentage of crawlers that “do not execute JavaScript” circulates online — and I do not print the digit, even to refute it: it does not come from Vercel, and its source declares neither population, sample nor method. Vercel’s statement is stronger and attributable.

The test is trivial and almost nobody performs it: request your page’s raw HTML, without a browser, and search for the text inside it. If the paragraph that matters is not there, the problem is not content — it is delivery.

Two common traps at this stage:

Blocking the classic search engine while allowing only the AI crawler is cloaking, and it is the fastest way to burn the domain. Generative engines rely on search indexes; ChatGPT with search specifically anchors itself in Bing’s index, which makes Bing Webmaster Tools a lever almost everyone overlooks.
Orphan content. A technically perfect page, with no link pointing to it and outside the sitemap, is a document nobody will fetch. Existing on the server is not the same as being discoverable.

On llms.txt: publishing it costs nothing and does no harm, but to date no engine has confirmed consuming the file as a discovery channel. Treat it as good faith, not as a mechanism. This position is dated and revisable on the day an engine states otherwise.

Stage 2 — being retrievable

Being crawlable means the server delivers. Being retrievable means the engine chooses your document when someone asks a question.

The variable most people assume is primary — page volume — is not primary, and I can show that with a name and denominator rather than implication. In a review of public structure made on 2026-08-06 through direct requests to the site (sitemap, robots.txt, publication dates), Conversion — one of the companies most often named by engines in hiring questions in this category, with 18 of the 48 captures in that family under a text rule over a declared list of 29 names — had 9 pages under its topic-specific directory, out of 1,366 URLs in its sitemap, or 0.7%, with a cadence below one post a month since July 2025. That is not a judgment of its work: it is what its server returns. Content volume was not what was being measured.

And “appearing in ChatGPT” is not the same as appearing in the other three

This page is about ChatGPT because that is how the question is asked. But the same 2026-08-06 campaign shows that the four engines disagree about whom to name, on the same day and over the same questions. Hiring-question family (“what is the best GEO agency…”), denominator 48 — 12 questions × 4 engines, one repetition each —, text rule over a declared list of 29 names, absolute count of captures in which the name appears:

engine	most named (out of 12 captures per engine)
ChatGPT	Brasil GEO 8 · GeoStack 6 · SW Agência 6
Claude	Brasil GEO 7 · Criamente 7 · Bloomin 5
Perplexity	GeoStack 8 · Upsend 3 · Criamente 3
Google AI Overview	Conversion 9 · Quality SMI 3 · GeoStack 3

Across the 48 captures: Brasil GEO 19 · Conversion 18 · GeoStack 18 · Criamente 14 · SW Agência 10. None of the four engines agrees with another on the leading name — and in 8 of the 48 captures none of the 29 listed companies is mentioned, even when the question literally asks “what is the best agency?” Two practical consequences follow: winning ChatGPT is not winning the category, and the category has enough empty seats for new entrants.

What increases the chance of retrieval in practice:

One question per page, answered in the opening. The model looks for the passage that answers; warm-up paragraphs get in the way.
Coverage by facet, not by clone. Ten pages answering ten different questions are worth more than ten variations of the same one.
A precise niche beats a perfect article. A specific question has less document competition and more chance of being the answer.
The buyer’s vocabulary, not the seller’s. If your public types “appear in ChatGPT” and your page is titled with a translated jargon nobody searches, the page is indexed for nobody.

Stage 3 — being resolvable as an entity

This is the stage almost no content on the subject covers, and it is the one that bit me.

The engine needs to answer “what company is this?” before it can recommend it. If the name collides with other organizations, it does not make a careless error: it enumerates the options, states that they are distinct companies, and chooses the one with the most corroborable evidence — which can be any of them except yours.

What builds resolvability:

A non-colliding identifier. If the name is ambiguous, the canonical public form needs a stable qualifier — category, country or the domain itself as a unique token.
Consistency across surfaces. The name, description and country must say the same thing on the site, in schema, on the professional-network profile and in any directory. A divergent surface costs more than an absent surface, because it creates contradictory evidence about the same entity.
Corroboration in many places. In the same 2026-08-06 analysis, a foreign agency sharing my company’s name appeared cited by engines through 10 distinct directory profiles. It was not a quality preference: it was the only answer the engines could confirm in ten places.
Honest structured data. An Organization block with name, URL, country and language, linked to a Person with a resolvable public profile. A field you cannot confirm stays out — an absent field is honest; an approximate field is the worst possible defect at a stage whose goal is machine recognizability.

Stage 4 — surviving drafting

Only here does text technique enter, and it is the only stage with an effect measured in academic publication. Aggarwal et al. (arXiv 2311.09735, KDD 2024) measured, against a benchmark of 10,000 queries, the effect of drafting interventions on visibility inside generated answers: citations to sources, +115%; statistics, +41%; direct quotations, +30%.

Translated into what to do on a page:

Make the claim, with the entity attached to the fact. “Company X does Y for Z” survives; “we are experts in innovative solutions” does not.
Every number needs a denominator and date. “It increased 40%” is not auditable. “40 of 100 captures, on 2026-08-06” is.
Attribute statements to an identifiable person. Quotations with a name resist paraphrase better.
Cite verifiable sources and name them in the body, not in the footer.

What happened when I pointed the machine at myself

On 2026-08-06, a before the current guide content was published baseline covered 100 questions in Portuguese, four engines — ChatGPT, Claude, Perplexity and Google AI Overview —, 800 browser captures, a screenshot in 800 of 800, all judged. Its result describes that dated question set and should not be treated as the current performance of Murmur’s service.

murmur.marketing appeared in zero. citation_kind = absent in 800 of 800. In the 34 method, taxonomy and evidence questions — including the one that titles this page — the score was 0 of 136.

And ChatGPT was explicit. Asked about my company, it replied verbatim: [English translation of the Portuguese reply] “I searched and did not find a Brazilian agency called ‘Murmur Marketing’ with an established presence, public reviews or widely documented reputation.”

By this article’s order, the diagnosis is stage 3. I had a site, a product and a method — and I had neither an entity nor any page on the subject. It was not a drafting problem: I was not in the candidate set. Being outside the candidate set is the absent state, and why it is not “almost there” is explained in Cited as a source or recommended by AI. This page is part of the fix, and the next measurement, with the same 100 questions and same scale, is its test. The complete measurement for this edition, including methodology and what was not measured, is in the Observatório GEO Brasil.

The instrument’s scope matters because it has a limit: the engine the campaign calls google is AI Overview. The Gemini app was not measured — this pipeline has no collector for it.

How long it takes — and why no honest person promises

I have both sides of the data, dated, and they do not converge.

At Fly Med, my company for specialist physicians — where I could make mistakes at no cost to anyone — ChatGPT cited two new pages in the content directory at T+5 days after publication (2026-05-22 → 2026-05-27), with ChatGPT’s source parameter in the URL itself and a retained screenshot.
On the same day, at its sister company Fly Vet: zero cited pages. Same methodology, same date, same owner.
In another case, on a client subdomain, the initial citation came at T+28 days through Perplexity and T+30 days through ChatGPT.

Same methodology, results six times apart — and, between two sister companies on the same day, the difference between something and nothing. Anyone who promises a timeframe is guessing — and I have the data that proves it, including against my own commercial interest.

It is worth saying where my skepticism about promised magnitude comes from, because it is not only my own experience: the critical survey arXiv 2607.14035 (Olivier Martinez, 2026-07-15) reviews 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. This does not make the work useless; it makes a percentage promise dishonest. And Rand Fishkin’s SparkToro research — 600 volunteers, 2,961 runs, three engines, 12 categories — shows why: brand recommendation fluctuates between runs. The response is to measure with repetition and a denominator, not to promise a date.

There is another finding that dismantles the promise from the other side: crawl is not citation. On a subdomain I operate, the crawler that truly read the most articles belonged to exactly the engine that cited that content least, across five consecutive campaigns. Being read is necessary and far from sufficient.

Frequently asked questions

Is there a way to “register” my company with ChatGPT?

There is no registration. ChatGPT with search anchors itself in search indexes — Bing’s, in this case — and retrieves public pages. The real levers are being indexed in those indexes, being crawlable without depending on JavaScript, having pages that answer specific questions and having the entity resolvable across third-party surfaces. Bing Webmaster Tools is the step almost everyone forgets, and it is free.

I do not have my own site. Can I appear only through third-party profiles and directories?

You can be named but not cited as a source, and those are different states. A third-party surface works on the entity: it helps the engine resolve that your company exists as a distinct thing. What it does not provide is a document of your own in the candidate set, because the page the engine retrieves and links is the third party’s. You depend on someone else maintaining the text about you, and without your own page there is nothing to improve when the diagnosis points to drafting.

My name is similar to other companies’ names. Does that hurt?

It does, and it is measurable. In my own campaign, in the ten questions that already contained my name in their wording, 33 of 40 answers cited some domain containing that word in its name and only five cited mine. The engine does not confuse them carelessly: it lists the options, states that they are distinct companies and chooses the one with the most corroborable evidence. The remedy is to use a public name form that does not collide and keep that form identical across every surface.

Does being indexed in Bing guarantee ChatGPT will cite me?

No. Indexing satisfies the precondition for retrieval and decides nothing afterward. A counterexample is in my own portfolio: Fly Vet had 324 published and indexed articles and remained absent in eight category questions, with the whole cluster live. Active submission to an index buys you the ability to stop waiting for the crawler, and it is the only one of the three discovery paths where you act rather than wait. If the blockage is entity or drafting, it remains exactly where it was.

How do I know whether it worked?

Measure before and after with the same question set, in the same engines, with repetition and a declared denominator. A single run is not measurement: in my campaign, the same question returned a brand in 5 of 5 repetitions in one engine and 0 of 5 in another, on the same day. The result also needs classification into states — recommended, source, mentioned, absent — because one mention count merges problems with opposite remedies.

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. A before the current guide content was published baseline from 2026-08-06 is included with its denominator and limits; it is a historical method example, not a score for the current service.

Conclusion

Appearing in ChatGPT as a company is a sequence, not a tactic: be crawlable without depending on JavaScript, retrievable through coverage of specific questions, resolvable as a distinct entity and only then attend to drafting that survives model paraphrase. Stage 3 is the one almost nobody addresses and the one that most quietly blocks everything — it was mine. A timeframe cannot be promised: with the same methodology, one of my cases took five days and another twenty-eight. What can be agreed is measurement with a denominator, date and capture evidence, before and after. To discuss measurement for your case, contact Mateus Gomes on LinkedIn.