AI & Research
Claude vs ChatGPT vs Gemini vs Perplexity
Where each major AI answer system came from, how they differ in the way they retrieve and cite, and what that means for your content.
Definition
Claude, ChatGPT, Gemini, Perplexity and Google AI Overviews are the five systems that answer most AI-mediated questions about the open web. They fall into three groups. Claude and ChatGPT are general assistants built around a model, with search attached. Perplexity is an answer engine, built around retrieval with models attached. AI Overviews is not a product at all but a feature inside Google Search. The distinction matters because it decides how each one finds your page and whether it links to you.
Three different things wearing the same interface
All five present a text box and return prose, which makes them look like variants of one product. They are not. The architecture underneath decides what gets cited, how fresh the information is, and how much traffic reaches the source.
- Assistants first, search second: ChatGPT and Claude answer from the model, and reach for the web when the question needs it
- Retrieval first, model second: Perplexity searches every time and treats the model as the thing that writes up what it found
- A feature inside a search engine: AI Overviews and AI Mode sit on top of Google's existing index and ranking systems
- Gemini spans two of these: a standalone assistant, and the model powering Google's search features
ChatGPT: the product that started the category
OpenAI released ChatGPT on 30 November 2022, fine tuned from a model in the GPT-3.5 series. It reached one million users in five days and 100 million in two months, making it the fastest growing consumer application on record at the time. That is the moment the current cycle began.
The lineage runs back further: GPT-1 in 2018, GPT-2 in 2019, GPT-3 in 2020. GPT-4 was released on 14 March 2023, having finished training in August 2022, roughly seven months before it shipped. Later GPT releases have continued the pattern of a capability jump followed by a long safety and evaluation period.
For publishers the important detail is that ChatGPT answers from the model unless browsing is invoked. When it does not search, your page cannot be cited no matter how well optimised it is, because the answer never touched the live web.
| Model | Year | What changed |
|---|---|---|
| GPT-1 | 2018 | First of the series, established the pre-train then fine-tune pattern |
| GPT-2 | 2019 | Larger, generated coherent long-form text |
| GPT-3 | 2020 | Over 100 times the parameters of GPT-2 |
| GPT-3.5 | 2022 | The series ChatGPT was fine tuned from |
| GPT-4 | 14 Mar 2023 | Finished training Aug 2022, held seven months before release |
Claude: built around a stated set of principles
Anthropic published Constitutional AI in December 2022, a training method where a model critiques and revises its own outputs against an explicit written set of principles rather than relying on human labels for every harmful case. That document is the through line of the product.
Claude 2 launched in July 2023 with the public claude.ai site. Claude 2.1 followed with a 200,000 token context window. The Claude 3 family arrived on 4 March 2024 as three tiers, Haiku, Sonnet and Opus, trading speed against capability. Anthropic has continued publishing an updated constitution alongside model releases.
Claude's long context window is the practically relevant trait for research work: it can hold entire documents in a single request rather than retrieving fragments of them.
My co-founders at Anthropic and I were among the first to document and track the scaling laws of AI systems, the observation that as we add more compute and training tasks, AI systems get predictably better at essentially every cognitive skill we are able to measure.
| Release | Date | What it introduced |
|---|---|---|
| Constitutional AI paper | Dec 2022 | Training against an explicit written set of principles |
| Claude 2 | Jul 2023 | The public claude.ai site |
| Claude 2.1 | Nov 2023 | 200,000 token context window |
| Claude 3 family | 4 Mar 2024 | Haiku, Sonnet and Opus tiers |
Gemini: a model family and a rebranded assistant
Google announced Gemini in December 2023 as a natively multimodal model in three sizes, Ultra, Pro and Nano. Gemini 1.0 Ultra scored 90.04% on MMLU, which Google described as the first model to exceed human expert performance on that benchmark.
In February 2024 Google renamed its Bard assistant to Gemini, folding the product name into the model name. This is why writing from 2023 refers to Bard and writing after refers to Gemini: they are the same assistant.
Gemini matters twice over. It is a standalone assistant, and it is the model behind Google's AI Overviews and AI Mode, which is where most people encounter it without choosing to.
It was built from the ground up to be multimodal, which means it can generalize and seamlessly understand, operate across and combine different types of information including text, code, audio, image and video.
| Size | Built for | Note |
|---|---|---|
| Ultra | The most complex tasks | Scored 90.04% on MMLU, which Google called the first above human expert level |
| Pro | Scaling across a wide range of tasks | The size that first powered Bard |
| Nano | On-device tasks | Runs locally on phones |
Perplexity: the answer engine
Perplexity was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho and Andy Konwinski, and launched its search engine that December. It describes itself as an answer engine rather than a chatbot or a search engine. The distinction is architectural: it searches the live web on every query, selects sources, and synthesises an answer with numbered citations attached to specific claims.
It routes queries across several frontier models rather than owning one, so the model is a component and retrieval is the product. Of the five systems, it is the most consistent at attributing individual sentences to a source.
For a publisher this makes Perplexity the most legible of the five. If your page is retrievable and answers the question directly, it is the system most likely to cite you by name and link to you.
| Stage | What the system does | Why it matters to you |
|---|---|---|
| Interpret | Reads the question and decides what is being asked | Ambiguous pages match fewer questions |
| Search | Queries the live web every time | Your page must be crawlable now, not at training time |
| Select | Picks the sources it will rely on | Being retrievable is necessary but not sufficient |
| Synthesise | Writes the answer from those sources | Clear standalone passages survive summarisation |
| Cite | Attaches numbered citations to claims | This is where a link back to you appears |
AI Overviews and AI Mode: not a product, a surface
Google began rolling AI Overviews out to everyone in the United States in May 2024, after a period in Search Labs under the name Search Generative Experience. It expanded past the US that August, past 100 countries in October 2024 and past 200 countries and 40 languages by May 2025. Google reported more than a billion monthly users by late 2024.
AI Mode, a dedicated conversational search surface, followed in 2025. Both are built on Google's existing crawl and index rather than a separate corpus.
This is the one that changes the most for site owners, because there is no separate system to optimise for. Google's own position is that the fundamentals still apply and there are no additional requirements to appear in AI Overviews or AI Mode.
| When | Milestone | Scale |
|---|---|---|
| May 2024 | General rollout in the United States | Hundreds of millions of users that week |
| Aug 2024 | First expansion beyond the US | Additional countries |
| Oct 2024 | More than 100 countries | Over 1 billion monthly users |
| May 2025 | More than 200 countries and territories | Over 40 languages |
| 2025 | AI Mode rolls out in the US | A dedicated conversational surface |
How they differ where it counts
The table below compares them on the properties that decide whether your content appears and whether a reader can follow the link back.
| System | Type | Searches the live web | Citation style | Where the traffic goes |
|---|---|---|---|---|
| ChatGPT | Assistant | Only when browsing is used | Inline links when it searched | Some click through, most answers end in the chat |
| Claude | Assistant | Only when search is enabled | Cites sources it retrieved | Similar, with long documents read in full |
| Gemini | Assistant and model | Yes, integrated with Google | Links alongside the answer | Feeds both the assistant and Search surfaces |
| Perplexity | Answer engine | Every query | Numbered citations per claim | Highest link visibility of the five |
| AI Overviews | Search feature | Uses Google's index | Links within the overview | Sits above results, often answers without a click |
What this means if you publish
The systems disagree about almost everything except one requirement: they can only use content they can actually read. Several of the crawlers that gather material for these systems do not execute JavaScript, so content that only appears after rendering is invisible to them even when it ranks normally in Google.
- Put primary content in the HTML response, not behind client side rendering
- Answer the question near the top of the page, in a passage that stands alone
- Use real headings in order, since passages are selected, not whole documents
- Date and attribute the page, because trust signals feed the selection
- Track citation and mention separately from ranking, since they move independently
Which one to reach for
For research where sources must be checkable, Perplexity's per claim citations make verification fastest. For reasoning over a long document you already have, Claude's context window means the whole thing fits without chunking. For general assistance with the widest tool ecosystem, ChatGPT has the largest surface area. For anything tied into Google's own data, Gemini is the one with the integration. AI Overviews is not chosen at all, which is exactly why it matters most for visibility.
Best practices for SEO remain relevant for AI features, and there are no additional requirements to appear in AI Overviews or AI Mode.
References
- OpenAI: GPT-4 Technical Report (arXiv)
- MIT Technology Review: ChatGPT is everywhere, here's where it came from
- MIT Technology Review: ChatGPT is OpenAI's latest fix for GPT-3 (30 Nov 2022)
- Nature: ChatGPT one year on, who is using it, how and why?
- Wikipedia: ChatGPT
- Wikipedia: GPT-4
- Anthropic: Constitutional AI, harmlessness from AI feedback
- Anthropic: Introducing the next generation of Claude
- Anthropic: Claude 2
- Anthropic: Claude's Constitution
- Google: Introducing Gemini, our most capable AI model
- Google: Bard becomes Gemini
- Wikipedia: Perplexity AI
- Google: New generative AI experiences in Search, May 2024
- Google: Expanding AI Overviews and introducing AI Mode
- Google: AI Overviews in over 200 countries and territories
- Google Search Central: AI features and your website
- Google Search Central: Guide to optimizing for generative AI features