ChatGPT, Gemini, Perplexity, Copilot: why each engine sees your brand differently

Marketing teams often check their brand in one AI assistant and treat the answer as the verdict of "AI" in general. That habit hides the central fact about how AI engines see your brand: ChatGPT, Gemini, Perplexity and Microsoft Copilot do not look at the same web. Each reads pages through its own crawler, draws on its own index or a partner's, decides by its own rule whether to search and shows sources in its own way, so a brand cited by one engine can be absent from another for reasons unrelated to content quality. An engine's retrieval architecture, in the sense used here, is the chain of four decisions between a page and a cited answer: which agent may read the page, which index holds it, whether the engine consults that index for a given question, and how the answer attributes its sources.

The comparison below relies only on the operators' own documentation, which we re-examined in October 2026, and on independent research, and it says so wherever the internals are not public. Its conclusion is practical: an optimisation made for one engine does not transfer automatically to another, so visibility has to be checked and prioritised engine by engine.

What is a retrieval architecture, and why does it decide visibility?

Every cited answer passes through four layers. Access is whether your robots.txt, firewall and CDN admit the user-agent that tries to read the page; index is where the page is then stored, in the engine's own index, a partner's, or both. Trigger is whether the engine searches for a given question or answers from what the model learned in training, and attribution is how the pages it used are shown to the reader. Location, language and personalisation act on top of all four. The layers multiply rather than add, because a page blocked at access cannot be rescued by good writing and an indexed page that is never consulted gains nothing from being fresh; this is why the principles in our guide to generative engine optimization must be translated into engine-specific checks.

ChatGPT: three OpenAI agents with three different jobs

Besides OAI-AdsBot, which only visits pages submitted as ads, OpenAI documents three agents that matter to publishers. OAI-SearchBot surfaces sites in ChatGPT's search features; sites opted out of it are not shown in ChatGPT search answers, although they can still appear as navigational links, and robots.txt changes take about 24 hours to apply. GPTBot collects content that may be used to train OpenAI's foundation models, and disallowing it signals that the site should not be used for that training. ChatGPT-User fetches pages on a user's request, which is why OpenAI warns that robots.txt rules may not apply; the policy side of these switches is covered in our article on whether to block or allow AI crawlers.

For the index, OpenAI's help centre says ChatGPT search sometimes partners with third-party search providers and refers to Microsoft's privacy policy, and its Enterprise and Edu documentation names Bing explicitly. Eligibility also requires the site's host or CDN to accept traffic from OpenAI's published IP ranges, and even then placement is not guaranteed. ChatGPT may search automatically when current information would help; answers that use web search may include citations, and a Sources view, when available, lists the cited sources and other relevant links.

Perplexity: an answer engine built on its own index

PerplexityBot surfaces and links sites in Perplexity's results and, according to the company, is not used to collect training data for foundation models, whereas Perplexity-User handles actions a user requests and therefore generally ignores robots.txt. Unusually, Perplexity has described its infrastructure: a research post from September 2025 reports an index tracking more than 200 billion unique URLs, continual refreshing of high-traffic documents, and documents split into self-contained spans that are retrieved and ranked individually. Answers carry numbered footnotes to their sources. Span-level retrieval rewards passages that state their claim completely, a point developed in our article on how to show up in Perplexity.

Google: AI Overviews and Gemini are two different doors

Google-Extended needs careful reading. Google's crawler documentation, updated in October 2026, describes it as a robots.txt product token rather than a crawler: existing Google agents do the crawling, and the token governs whether content may be used to train future Gemini models and for grounding in the Gemini apps and on Vertex AI, without affecting inclusion in Google Search. AI Overviews and AI Mode, by contrast, are Search features: a page must be indexed and eligible for a snippet, no additional technical requirements apply, and what appears is limited with nosnippet, data-nosnippet, max-snippet and noindex. Blocking Google-Extended therefore does not remove a page from AI Overviews. Both features may also use "query fan-out", running related searches across subtopics, which helps explain why an overview can cite a page that does not rank for the literal query; see our piece on AI Overviews and zero-click search.

The Gemini app is a separate product. Its privacy hub says that some responses are grounded on Search results, and its help pages note that not every answer includes sources; when one does, a Sources button opens them in a side panel. In the Gemini API, the model itself decides whether to search, and the grounding tool documents no location or region parameter.

Microsoft Copilot: the Bing index is the entry point

For Copilot, the publisher's entry point is Bing. Bingbot is Bing's crawler, and Microsoft's documentation for the work editions, Microsoft Copilot and Copilot Chat, says the assistant may fetch information from the Bing search service with a short generated query stripped of user and tenant identifiers; the consumer transparency note describes grounding in web results and citations below the answer without naming a provider. The AI Performance report in Bing Webmaster Tools, released in public preview in February 2026, shows how a site is cited across Copilot, AI summaries in Bing and selected partners, together with a sample of the grounding queries involved. Bing has documented since 2023 that NOARCHIVE keeps pages out of its chat answers and NOCACHE limits them to URL, title and snippet, and Search Engine Journal reported that its February 2026 guidelines extend these directives to Copilot.

Engine by engine: crawler, index, sources and what to check

The table condenses this documentation; its last column names the first place to look when you are absent from an engine's answers, before any debate about content quality begins.

An AI engine never sees the whole web, only the part its crawler may read, its index holds and its trigger decides to consult, which is why visibility has to be earned and verified one engine at a time.

When does an engine search, and when does it answer from memory?

Each engine combines knowledge frozen at a training cutoff with information retrieved at question time, and the documentation treats retrieval as a decision: OpenAI says ChatGPT may search when current information would help, Microsoft that a determination is made whether a request needs grounding data from the web, and Google that the Gemini API model decides whether a search would improve the answer. Answers from training reflect how the brand was described before the cutoff, largely on third-party pages a new product page cannot change, whereas answers from search reflect what the crawler can read now. A training opt-out such as GPTBot affects only future training, search agents such as OAI-SearchBot and PerplexityBot affect retrieval, and Google-Extended covers both Gemini training and grounding in the Gemini apps; teams that confuse these block the wrong agent or expect a page update to change an answer that retrieved nothing.

Why do two users asking the same question see different answers?

Context shapes answers even within one engine. ChatGPT and the Gemini app approximate location from the IP address, ChatGPT may use saved memories when it rewrites a question into search queries, and Perplexity, Gemini and Copilot can personalise answers from memory or past conversations; Google has also brought Personal Intelligence to AI Mode: introduced in January 2026 as an opt-in Labs experiment for US subscribers, it was opened in May 2026 to users in nearly 200 countries and territories without a subscription, and it lets people choose to connect apps such as Gmail and Google Photos. A screenshot from a colleague's account is therefore evidence about that account, not about the market.

Language adds a further layer. In a September 2025 arXiv preprint, Chen, Wang, Chen and Koudas compared API versions of several engines, including Perplexity's sonar-pro, Gemini 2.5 Flash with Google Search grounding and GPT-4o search preview, with Google's top ten results, and found significant differences in domain diversity, freshness, cross-language stability and sensitivity to phrasing; when English queries were compared with versions in five other languages, the overlap of cited domains for the GPT configuration was near zero. Developer APIs are not consumer apps and the study did not test Turkish or Russian, so the figures do not transfer directly, but they justify treating a Turkish and an English version of one question as two separate questions in measurement, as our articles on multilingual GEO and on why AI answers change every time explain.

Why does an optimisation for one engine not transfer to another?

A common foundation exists, since crawlable pages that state facts plainly and are corroborated elsewhere help in every engine; this is the overlap described in our comparison of AEO and GEO. The controls that decide eligibility, however, are engine-specific. A firewall that challenges unfamiliar bots can block OpenAI's IP ranges while admitting Googlebot, a restrictive max-snippet limits what Google's AI features may show although Perplexity's crawler documentation does not mention the directive, and a leftover NOARCHIVE directive, which Bing applies to its AI answers, can remove pages from Copilot. Consider a hypothetical company whose rewritten comparison pages began to appear in AI Overviews while a security plugin quietly rejected OAI-SearchBot; only an engine-level check would expose the gap. Source preferences differ too: the preprint above found that AI engines lean on earned media far more than Google does, and Ahrefs' Brand Radar data for June 2025 placed YouTube among the most-cited domains in Perplexity and AI Overviews but not in ChatGPT's top ten, a pattern we examine in YouTube and video content in AI answers.

How should visibility be measured per engine and prioritised?

Measurement should mirror the architecture: one set of buyer questions, a fixed market and language for each, every relevant engine, and separate records of whether the brand is named and whether its pages are cited, because the difference between a mention and a citation grows when engines display sources so differently. Repeat the runs, neutralise personalisation where possible, and diagnose a gap in layer order: access, index, trigger, attribution.

Prioritise by your own audience rather than by global market-share headlines, which change quickly and rarely describe a given sector or country. Referral sessions from chatgpt.com, perplexity.ai, gemini.google.com or copilot.microsoft.com, answers to a "how did you hear about us" question and what buyers tell your sales team are better evidence; our article on AI referral traffic shows how to tie these visits to leads. Clicks from AI Overviews and AI Mode, however, arrive as ordinary Google traffic, so the largest surface may be the least visible one in analytics.

Rocketly GEO Suite follows the same distinction. Its standard engine is Google AI Overview, measured weekly for the questions you track in each market and language, while ChatGPT and Perplexity measurement is in a controlled rollout and not yet open to every account. Gemini is not measured, because its grounding tool accepts no location and a market-specific measurement is therefore not possible, and Copilot is not a measured engine: it appears only as an AI referrer domain in CRM lead attribution. To see engine-by-engine visibility for your own questions, you can create a Rocketly account and add the GEO Suite.

Frequently asked questions

Does blocking Google-Extended remove my site from AI Overviews?

No. Google documents Google-Extended as a token governing Gemini training and grounding in the Gemini apps and Vertex AI, and states that it does not affect inclusion in Google Search. AI Overviews and AI Mode are Search features, governed by indexing and snippet controls such as noindex, nosnippet and max-snippet.

If Perplexity cites my page, will ChatGPT cite it too?

Not necessarily. Perplexity retrieves from its own index through PerplexityBot, whereas ChatGPT search depends on OAI-SearchBot access and third-party search providers, and the two decide differently when to search. A citation in one engine shows that your content is usable, not that the other engine can reach it.

Does ChatGPT use Bing?

OpenAI's general help article mentions third-party search providers and refers to Microsoft's privacy policy, and its Enterprise and Edu documentation names Bing explicitly. Inclusion in ChatGPT search also requires allowing OAI-SearchBot, so being indexed by Bing is useful but not sufficient on its own.