No answer works for both at once, and this is measurable: in the campaign I ran on 2026-08-06, over the same twelve hiring questions, the four engines returned four different leading names on the same day — and in another question, repeated five times in each engine, one company was named 5 of 5 times by ChatGPT and 0 of 5 by Claude. Anyone looking for one tactic for “ChatGPT and Perplexity” is looking for something the measurement does not support. There is a common layer that applies to all four, and an engine-specific layer that changes.
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 campaign results below are dated observations from defined samples; the 2026-08-06 Murmur baseline predates the current guide collection and is not a present-day performance score.
This page is a spoke of How to get my brand recommended by AI, which covers four conditions in the order they fail. The slice here is narrower: what changes when the engine changes.
Summary
- The four engines disagree about the leading name. In the hiring-question family — 12 questions × 4 engines, denominator 48 — the most seen name per engine was Brasil GEO in ChatGPT (8 of 12), Brasil GEO and Criamente tied in Claude (7 each), GeoStack in Perplexity (8), and Conversion in Google AI Overview (9). Deterministic text rule over a declared list of 29 names, 2026-08-06.
- The disagreement exists within one question. Repeating [English translation of the Portuguese prompt] “what is the best company for GEO in Brazil?” five times in each engine — 20 captures — Conversion was named 5 of 5 by ChatGPT, 4 of 5 by Perplexity, 0 of 5 by Claude, and 0 of 5 by Google AI Overview: 9 of 20 in total.
- The same site can be a source in one engine and a recommendation in another. At Fly Vet, my own agency, with 100 questions and one repetition in 2026-06, ChatGPT recommended it in 30 of 100, Claude in 41 of 100, and Google never recommended it — it used the site as a source in 81 of 100.
- The engine that reads most cites least. In a client-subdomain server log, ClaudeBot was the only crawler that actually read articles — 24 requests to 20 distinct slugs — while Claude stayed at zero citations in five consecutive campaigns, each covering 394–395 questions.
- In 8 of 48 hiring captures, none of the 29 names appears. The seat is empty often enough that appearing, rather than displacing someone, is the inexpensive move.
- What all four share is the lower layer: server-delivered HTML without JavaScript dependence, declared discovery, and one page per question. That does not change by engine.
The assumption inside the question, and why it is false
“How to be cited by ChatGPT and Perplexity” assumes there is one thing called “AI” that responds one way. This is the Brazilian generative-search market’s costliest assumption, and it breaks in two independent measurements I have.
First, disagreement between engines. In the 2026-08-06 campaign, the hiring-question family — the twelve phrasings such as [English translations of the Portuguese prompts] “best GEO agency” and “who to hire for GEO” — ran in four engines, once each: denominator 48. Counting captures in which each name appears in the text, under a deterministic rule over a declared list of 29 names, the overall score was Brasil GEO 19, Conversion 18, GeoStack 18, Criamente 14, SW Agência 10, Upsend 9, Bloomin 9 and Quality SMI 7. Absolute counts, never percentages. What matters is that when opened by engine, the four leaders differ, on the same day and over the same twelve questions. In 8 of 48 captures, none of the 29 names is named.
Second, disagreement within one question. The question that originated this work ran five times in each engine: 20 captures. Conversion appeared in 5 of 5 ChatGPT answers, 4 of 5 Perplexity answers, and none of the five Claude or Google AI Overview answers. Nine of twenty. A screenshot of one question, in one engine, on one day, is not measurement; it is an anecdote with screen resolution.
This does not authorize saying one engine is better than another, or one reliable and another not. It authorizes one operational conclusion: measuring one engine and drawing a conclusion about “AI” is a methodological error. How to measure whether AI cites your brand explains how to measure without making it.
The four engines, one by one
This is the part that changes. Each of the four has different mechanics for retrieving a page, showing a source and deciding whether to name someone. I describe what I observed and what I did not observe.
1. ChatGPT — the engine anchored in a search index
ChatGPT with search does not invent its own web index: it grounds itself in the Bing classic-search index. The practical consequence is the most overlooked lever: Bing Webmaster Tools is an active-submission channel that almost nobody in the Brazilian market uses, and directly addresses discovery.
OpenAI’s agents are distinct and do distinct things: GPTBot (training), OAI-SearchBot (search index) and ChatGPT-User (search triggered by a conversation). In a client-subdomain access log, over about 3.6 days and 1,074 requests, the three combined made 184 requests, versus 327 from Googlebot and 62 from PerplexityBot. This is user-agent counting, not verified IPs — an essential caveat.
Measured behavior: at my agency, with 100 questions and one repetition, ChatGPT recommended Fly Vet in 30 of 100. In the category campaign, for hiring questions, it was the engine that named Brasil GEO most — 8 of 12.
2. Perplexity — the engine with the most explicit citation
Perplexity hides its mechanics least: it shows numbered sources beside text, making it easiest to audit and easiest to celebrate misleadingly. Appearing in the sources list is not being recommended in the answer body — Cited as a source or recommended by AI explains the four possible states, and the distinction is the difference between fixing drafting and fixing discovery.
PerplexityBot appears in the log with 62 requests over the same 3.6 days. A case changed how I read every crawler counter: on another subdomain, 2,256 log lines in 24 hours showed 6 PerplexityBot requests by raw user-agent count. Cross-checking user-agent against real IP left zero. The six were one address sending fake PerplexityBot, Googlebot, bingbot, YandexBot, DeepSeekBot and CCBot user agents in the same second, requesting /.env, /.git/config and /service-account.json. It was not a crawler; it was credential scanning dressed as a bot. Anyone counting AI crawlers by user-agent is counting an intruder.
Measured behavior: in the hiring family, Perplexity was the only one of four engines where GeoStack led — 8 of 12.
3. Claude — the engine that reads most and cites least
This is my most counterintuitive finding. In a 24-hour log with verified IPs, ClaudeBot was the crawler whose traffic survived cleanup: 88 raw requests became 79 legitimate, with Anthropic IPs. What it did was read: 24 article requests, 20 distinct slugs, once each, plus 17 to the content sitemap, 17 to the sitemap index and 15 to robots.txt. That is ordered reading, not scanning.
Yet Claude had zero citations in five consecutive campaigns, with universes of 394 to 395 questions each. The bot that read the site most cited it least. Anyone selling “optimize for the crawler” is selling half the equation — crawling is not citation, and the distance between them is covered by How long does AI take to start citing my site?.
In the opposite direction, Claude was the engine most generous to my agency in a different universe: 41 recommendations in 100 questions, against ChatGPT’s 30. It was also the engine in which Conversion did not appear once in five repetitions of the category question. It is not “worse”; it is different.
4. Google AI Overview — the engine that cites and does not recommend
The google engine in this measurement is AI Overview, the generated block Google shows above results. The Gemini app is outside this campaign: my pipeline has no harvester for it, so nothing here speaks for it. Declaring what was not measured is method, not a footnote.
AI Overview behaved most distinctly at my agency: in 100 questions it never recommended Fly Vet, and used it as a source in 81. Eighty-one source citations and zero endorsement. This is the exact portrait of a drafting problem, not discovery: the page is reached, retrieved and linked — then ignored when the machine writes the sentence that decides.
In the hiring family, AI Overview was the only engine in which Conversion led — 9 of 12. Alongside that, its public structure observed on the same day records over fifteen years of domain history and a presence beyond its own site. These are two observations from the same date, placed side by side: observed structure, not a causal verdict — nothing here authorizes explaining why someone wins.
The table: engine × what it did × what that authorizes
| engine | leader in hiring family (n=48, rule over declared list of 29 names) | at my agency, 100 questions × 1 repetition, 2026-06 | crawler in log (3.6 days, by user-agent) | what this data does NOT authorize |
|---|---|---|---|---|
| ChatGPT | Brasil GEO, 8 of 12 | recommended in 30 of 100 | OpenAI ≈184 (GPTBot + OAI-SearchBot + ChatGPT-User) | saying Bing “solves” citation — grounding is discovery, not endorsement |
| Perplexity | GeoStack, 8 of 12 | not measured in this campaign | PerplexityBot 62 (and 0 after IP cross-check in another log) | treating presence in the sources list as recommendation |
| Claude | Brasil GEO 7 · Criamente 7 | recommended in 41 of 100 | ClaudeBot 9 in this log; 79 IP-verified in another | concluding reading leads to citation — five consecutive zero campaigns say the opposite |
| Google AI Overview | Conversion, 9 of 12 | 0 recommendations · 81 of 100 as source | Googlebot 327 · bingbot 1 | speaking about the Gemini app: it is not part of this measurement |
Two caveats travel with this table; without them, it lies. First: its leader column uses a deterministic text rule over a declared list of 29 names, blind to everything outside it — the judge that read those same 800 captures extracted 447 distinct brands, and the most cited one absent from my list appeared 49 times in 800. I did not know it existed. Second: the crawler column counts user agents, not verified IPs, except where noted.
What all four share — and where to start
Everything that changes is above. The lower layer does not change and is short:
- Server-delivered HTML, not HTML assembled in the browser. None of the main AI crawlers renders JavaScript. Vercel’s own network-log reading in The rise of the AI crawler gives the detail rarely repeated: 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 exists only after a script runs does not exist for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, CCBot or Google-Extended.
- Declared discovery through only three paths: an internal link from an already known page,
sitemap.xmldeclared inrobots.txt, and active submission to an index.llms.txtis good faith: no engine has confirmed consuming it as a crawling route. Publishing costs nothing; relying on it does. - One page per question. Facets, not clones: leadership is won with several pages, each owning a distinct question, not one excellent page trying to cover them all. How to structure content to be cited by AI covers the page form.
The third layer has academic evidence: Aggarwal et al., Generative Engine Optimization, presented at KDD 2024 and published as arXiv 2311.09735, built GEO-bench with 10,000 queries and measured citations to sources +115%, statistics +41%, direct quotations +30%. These are effects against an academic benchmark, for a particular set of engines and queries, and describe drafting: the form difference after a page has already been retrieved.
Where others win, and where I do not
Across 100 (question, engine) pairs, each run five times, Conversion appears in a majority in 19, GeoStack in 15, Brasil GEO in 14, Criamente in 9 and Profound in 8 — deterministic rule over a declared 29-name list, 2026-08-06. Two mandatory caveats: the first three are separated by less than 1.3 binomial standard errors, so this is not a stable ranking; and the 19-pair name is absent in the other 81. In 51 of 100 pairs, no listed name reaches a majority.
The dated comparison shows two routes associated with visibility: Conversion’s long-established off-site authority corresponded to 19 of 100 pairs, while GeoStack’s concentrated content footprint corresponded to 15. The survey did not isolate causation. Its search pass did not confirm third-party coverage for GeoStack, which is not proof that none exists. Murmur’s baseline in that same campaign predates the current guide collection and should not be read as a current result.
Frequently asked questions
Is there one tactic that works for ChatGPT and Perplexity at the same time?
Only at the lower layer. Server-delivered HTML without JavaScript dependence, discovery declared through sitemap and internal links, and one page per question apply to all four engines. Above that, it splits: in the 2026-08-06 measurement over the same twelve hiring questions, each of four engines returned a different leading name, and in a question repeated five times per engine a company appeared 5 of 5 in ChatGPT and 0 of 5 in Claude.
Does appearing in Perplexity’s sources list mean I was recommended?
No. They are different states within the same answer. My clearest case is Google AI Overview at my own agency: in 100 questions it used the site as a source in 81 and recommended it in zero. Eighty-one source citations and no endorsement describe a page-drafting problem, not discovery — treating both with one metric returns the same number for opposite problems.
If the crawler is reading my site, does citation come afterward?
Not necessarily. In a server log with verified IPs, ClaudeBot was the only crawler that truly read articles — 24 requests to 20 distinct slugs, once each — while Claude had zero citations in five consecutive campaigns, each with 394 to 395 questions. Crawling is a condition, not a consequence.
Why not publish every brand’s percentage share by engine?
Because a percentage depends entirely on how many names the detection list sees, and mine sees 29. The judge that read the same 800 captures extracted 447 distinct brands, and the most cited unlisted one appeared 49 times in 800 without my knowing it existed. Any normalized share would be inflated, so I publish absolute counts with the denominator in the same sentence.
If I only have the capacity to work one engine at a time, which should I start with?
Measurement does not select one; the closest answer is the cut by state. At my own agency, with 100 questions and one repetition, Claude recommended in 41, ChatGPT in 30 and Google AI Overview in zero — while using the site as a source in 81 of those 100. An engine that cites but does not endorse has a drafting problem; an engine that does not cite has a discovery or entity problem, and those are different expenditures. The useful choice is therefore not between engines but states: which state is blocked in the engine where your buyer actually asks. The lower layer — served HTML, declared discovery and one page per question — applies to all four and needs no prioritization.
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. The 2026-08-06 campaign is a historical, pre-publication baseline; its stated method limitations and results apply only to that dated sample. It should not be treated as a current performance score.
Conclusion
The question requests one answer for two engines; measurement returns four behaviors. The common layer is short and inexpensive: HTML that exists without JavaScript, discovery declared through sitemap and internal link, and one page per question. The engine layer is where money is lost when someone generalizes — the same wording returned Brasil GEO in ChatGPT, GeoStack in Perplexity, Criamente and Brasil GEO tied in Claude, and Conversion in Google AI Overview, on the same day. The operational conclusion is not to choose one engine: measure all four, with repetition, and stop treating “AI” as one subject. To discuss measurement for your case, contact Mateus Gomes on LinkedIn.