The assistants and AI search surfaces people ask, from Google AI Overviews to ChatGPT, Perplexity and Claude, and the parts of their answers that cite sources.
Also called: AI Overviews, AI Overview, AIO, Google AIO, Search Generative Experience, SGE
DocumentedMeasuredMechanics, models and availability per Google's own posts and Search Central documentation (authoritative about Google only); click impact from one independent study (Pew Research Center, 2025).
Google AI Overviews are the AI-generated summaries shown at the top of some Google results pages, written by a custom Gemini model and paired with links to the web pages that support them. They grew out of the Search Labs experiment known as SGE and are available in more than 200 countries and territories.
Google AI Overviews are the generative answer layer built into the classic Google results page. On queries where Google's systems judge a summary to be additive, a block above the organic results combines a synthesized answer with supporting links, placed inline in the text and in a link display beside or beneath it. Google began rolling AI Overviews out to everyone in the US on 2024-05-14, after testing them in Search Labs, and on 2025-05-20 reported availability in more than 200 countries and territories and more than 40 languages.
The model behind them has changed several times: a Gemini model customized for Search at launch, Gemini 2.0 for harder US queries from 2025-03-05 (when the sign-in requirement was dropped and teens gained access), a custom Gemini 2.5 in the US from May 2025, and Gemini 3 as the global default since 2026-01-27. The term covers the summary block on the main results page. Google's separate conversational tab is AI Mode, and a follow-up question asked from an AI Overview now continues there, on desktop and mobile worldwide since 2026-05-19.
How it works
Google states that AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources, and that they can show a wider and more diverse set of links than a classic results page. A page is eligible as a supporting link if it is indexed and eligible to appear in Search with a snippet; Google says there are no additional technical requirements and no special schema.org markup, and that its AI features build on the same core ranking and quality systems as Search. Publishers limit what is shown with nosnippet, data-nosnippet, max-snippet or noindex. Traffic from AI features is counted in the Search Console Performance report under the Web search type. On 2026-10-15 Google said it was showing more inline links, placing them higher in responses, and testing a dedicated section for official websites at the top of some answers.
Why it matters
AI Overviews sit inside Google's main results page, so they change what a ranking is worth. One independent panel study (Pew Research Center, March 2025 browsing data from 900 US adults) found users clicked a traditional result in 8% of visits when a summary appeared versus 15% without one, clicked a link inside the summary in 1% of visits, and ended the session more often (26% versus 16%). Teams therefore track whether they are cited inside the summary as well as rank and clicks, keep key pages indexed and snippet-eligible, and measure AI Overview presence query by query, because the summary does not appear on every search.
Example
A US user searches "how long does a heat pump last" and sees three short paragraphs of generated text above the blue links, with link icons after key sentences and a list of cited pages beside the answer. A manufacturer whose maintenance guide is one of those cited pages gains visibility even if few users click through, and a follow-up question opens an AI Mode conversation.
Also called: AI Mode, AI Mode in Google Search, AI Mode tab, Google AI search mode
DocumentedFeatures, models and availability per Google's announcements; individual features roll out by country, language and subscription tier at different times.
Google AI Mode is the conversational AI search experience inside Google Search, reached through its own tab, where people ask long or multi-part questions and follow up in a chat-style thread. Answers are generated by Gemini models with links to supporting pages, and since 2026-05-19 Gemini 3.5 Flash is the default model worldwide.
Google AI Mode is a separate destination within Google Search for multi-turn, reasoning-heavy questions, distinct from AI Overviews, which sit inside the classic results page. Google introduced it on 2025-03-05 as an opt-in Search Labs experiment, opened it to everyone in the US without Labs sign-up on 2025-05-20, and on 2025-10-07 reported more than 35 new languages and availability in over 200 countries and territories. The default model moved from a custom Gemini 2.0 to Gemini 2.5 (US, May 2025), Gemini 3 Flash (worldwide, 2025-12-17) and Gemini 3.5 Flash (worldwide, 2026-05-19).
Features layered on top include Deep Search, which issues hundreds of searches and returns a fully cited report (first for US Google AI Pro and Ultra subscribers, from 2025-07-16); Search Live, a voice and camera conversation offered wherever AI Mode is available since 2026-03-26; agentic booking, expanded at I/O 2026 to local services, with Google calling businesses in select categories; and opt-in Personal Intelligence, which connects apps such as Gmail and Google Photos and is expanding to nearly 200 countries and territories with no subscription required.
How it works
Like AI Overviews, AI Mode may use query fan-out: the system breaks a question into subtopics, runs many related searches across the web and Google's data sources, and composes an answer with links. Google states that AI Mode and AI Overviews may use different models and techniques, so the responses and links they show will vary for the same query. Context carries across turns, and a follow-up asked from an AI Overview continues in AI Mode. Eligibility for supporting links follows the same rule as AI Overviews: an indexed page eligible to show with a snippet, with no extra technical requirement.
Why it matters
AI Mode handles a different kind of query than a classic search: longer, comparative and conversational, with several turns in which a brand can enter or drop out of the answer. Because Google says its models and link sets differ from AI Overviews, a page cited in one surface may be missing from the other, so measurement has to sample both separately and include follow-up turns. Agentic booking and business calling also mean that structured, current business information (hours, prices, availability, booking links) can decide whether a business is offered as an option at all.
Example
A user types "compare heat pumps and gas furnaces for a 1,500 sq ft house in Ohio, including running costs" into the AI Mode tab, reads a structured answer with cited sources, then asks "which installers near Columbus offer financing?" in the same thread. Each turn triggers fresh searches and can cite a different set of sites.
Also called: Google Gemini, Gemini app, Gemini Apps, gemini.google.com, Bard
DocumentedObservedRename, models and source display per Google's posts, release notes and Gemini Apps Help; when an answer is grounded in web results and which links appear is observed behavior that Google does not document in detail.
Gemini is Google's standalone AI assistant, used at gemini.google.com and in mobile and desktop apps, and renamed from Bard on 2024-02-08. It answers from its model's knowledge and, for many questions, from web information; when it provides links, a Sources button opens a panel listing them. It is separate from Google Search and its AI features.
Gemini, in the sense used here, is Google's consumer assistant app, separate from Google Search and from the Gemini models that also power AI Overviews and AI Mode. Google renamed Bard to Gemini on 2024-02-08 and released a Gemini mobile app the same day. The models inside the app change often: Google's release notes list Gemini 3.5 Flash rolling out to all users on 2026-05-19 and a frontier model, Gemini 4 Argon, rolling out to Google AI Pro and Ultra subscribers on 2026-10-13.
Gemini Apps Help states that when sources are available, a Sources button at the bottom of a response or inline opens a side panel of related links, which can point to public websites, uploaded files or connected Workspace content, and that not every response includes one. Google's help pages also warn that Gemini can hallucinate, including when describing how it cites sources or provides fresh information. Deep Research, the app's agentic report mode launched on 2024-12-11, is covered under its own entry.
Why it matters
Gemini the app and Gemini inside Search are separate visibility targets. The app may answer from parametric memory with no links, or from web results with a Sources panel, and the two paths call for different fixes; reports that say "Gemini" without naming the surface are ambiguous. Accurate facts on a brand's own site and on widely cited third-party pages feed the grounded path, while long-standing, consistent coverage shapes what the model already believes. Sampling the app repeatedly and logging whether sources appear, and which domains they name, shows which path a given question takes.
Example
Asked "what does Acme Analytics cost per seat?", the Gemini app may give a confident figure with no Sources button, drawn from model memory that predates a price change, or a current figure with a side panel linking to Acme's pricing page. Only the second case is fixed by updating that page.
Also called: ChatGPT web search, ChatGPT with search, search in ChatGPT
DocumentedObservedCrawler rules, availability and citation display per OpenAI's documentation and help center; the search providers and index behind answers are only partly disclosed.
ChatGPT search is the web search capability built into OpenAI's ChatGPT: when a question needs current or specific information, ChatGPT searches the web, writes an answer and attaches citations and a Sources list. It launched on 2024-10-31, absorbing the SearchGPT prototype, and is available on every ChatGPT plan and to signed-out users.
ChatGPT search covers the moments when ChatGPT retrieves live web results to answer, as opposed to answering from trained knowledge alone. It launched on 2024-10-31 for Plus and Team users and the SearchGPT waitlist, then expanded to all logged-in users and, in February 2025, to people without an account in regions where ChatGPT is available. OpenAI's help center lists web search on Free, Go, Plus, Pro, Business, Enterprise and Edu, says people who are not signed in can also use it, and lets Enterprise and Edu admins turn it off.
ChatGPT decides when a search would improve an answer, and users can also invoke search manually. Answers carry inline citations, and a Sources button lists cited pages and other relevant links. Shopping results, maps and restaurant reservation availability (from providers including OpenTable, Resy and Yelp, per the help center) build on the same capability. At launch OpenAI said the search model was a fine-tuned GPT-4o drawing on third-party search providers and partner content; that statement does not establish the models or providers used today.
How it works
When search triggers, ChatGPT rewrites the user's question into one or more targeted queries, sends them to a search provider, reads the results and composes an answer with citations. OpenAI's help center says it may use an approximate location derived from the IP address and saved memories when rewriting queries, without sending the IP address or account details to search providers. Site access is governed by OpenAI's crawlers: OAI-SearchBot surfaces sites in ChatGPT's search features, and a site that disallows it will not be shown in search answers, though it can still appear as a navigational link. ChatGPT-User handles user-initiated fetches, and OpenAI says robots.txt rules may not apply to it. OpenAI says a robots.txt change takes about 24 hours to be reflected in search.
Why it matters
ChatGPT search is one of the few AI answer surfaces with a written crawler contract, so the first lever is technical: allow OAI-SearchBot in robots.txt, keep pages fast and indexable, and confirm in server logs that it visits. Beyond access, citation depends on which pages the search step retrieves for the rewritten query and how directly they answer it. Teams test the exact prompts buyers use, record which domains ChatGPT cites, and compare results over repeated runs, since answers vary. Clicks on citations arrive as referrals that analytics can attribute, though many answers are read without a click.
Example
A user asks ChatGPT "which project management tools integrate with Xero?". ChatGPT runs several searches, answers with a short comparison and numbered citations, and lists the cited review sites and vendor pages under Sources. A vendor whose integrations page blocks OAI-SearchBot can still be named through other sites, but its own page will not be cited.
DeprecatedA temporary prototype name; OpenAI folded its features into ChatGPT at the launch of ChatGPT search on 2024-10-31.
SearchGPT was the name of OpenAI's temporary AI search prototype, announced on 2024-07-25 and tested with a small group of users and publishers. Its features were folded into ChatGPT when ChatGPT search launched on 2024-10-31, so the name now survives mainly in 2024-era articles and as an informal synonym for ChatGPT search.
SearchGPT was OpenAI's prototype of a search experience that combined its models with live web information and clear links to sources. OpenAI announced it on 2024-07-25 as a temporary test for a small group of users and publishers, with a waitlist, and said it planned to integrate the best features directly into ChatGPT. In the same announcement OpenAI introduced a way for publishers to manage how they appear in search and stated that search was separate from training its foundation models, so a site could be surfaced in search results even if it opted out of training.
On 2024-10-31 OpenAI launched ChatGPT search, saying it had brought the best of the SearchGPT experience into ChatGPT, and gave SearchGPT waitlist users access. No standalone SearchGPT product exists today. The term is historical; current behavior, crawler rules and availability are described under ChatGPT search.
Why it matters
The name still appears in articles, tool menus and client questions, and some readers assume SearchGPT is a separate engine to optimize for. There is no such engine: everything said about ChatGPT search applies. The prototype's lasting legacy for publishers is the separation it introduced between search and training, which survives as the split between OAI-SearchBot (search visibility) and GPTBot (training). A site can block training and remain eligible for search citations, a choice worth reviewing in robots.txt whenever an older guide mentions SearchGPT.
Example
A 2024 blog post advising brands to "rank in SearchGPT" describes the prototype that became ChatGPT search. The practical equivalent today is to check OAI-SearchBot access in robots.txt and test citations directly in ChatGPT.
Also called: Copilot, Bing Chat, the new Bing, Bing Copilot, Copilot Search, Copilot Search in Bing
Documented
Microsoft Copilot is Microsoft's consumer AI assistant, launched as Bing Chat on 2023-02-07 and renamed Copilot in late 2023, with answers grounded in Bing search results and linked to their sources. Bing also offers Copilot Search, a results page launched on 2025-04-04 that blends classic results with a generated answer and prominently cited sources.
Microsoft Copilot here means Microsoft's consumer assistant and its search-facing surfaces in Bing and Edge, as distinct from Microsoft 365 Copilot, which grounds answers in an organization's own data. Microsoft introduced the new Bing with chat on 2023-02-07, running a next-generation OpenAI model through its own Prometheus technique and linking to the web pages behind its answers. It announced a unified Copilot brand on 2023-09-21, and a 2023-11-15 update said Bing Chat and Bing Chat Enterprise would simply become Microsoft Copilot.
On 2025-04-04 Bing launched Copilot Search, which shows a summary, a direct answer or a structured layout depending on the query, links sentences inline to their sources, reveals every source used with one click and suggests follow-up topics; Microsoft said it was available worldwide except Russia and China. In Edge, Copilot can draw on open tabs for context, and the agentic feature formerly called Copilot Actions became "Browse with Copilot" on 2026-05-13, limited to Microsoft 365 Premium subscribers in the US.
How it works
Copilot's web answers are grounded in Bing search results, so whether a page can be retrieved and cited depends on Bing crawling and indexing it. IndexNow, a protocol Bing participates in, lets a site report new or changed URLs, which are shared with all participating search engines. Copilot Search attaches citations at sentence level and repeats cited sources and related web results at the top and bottom of the page.
Why it matters
Copilot visibility rides on Bing. A page missing from Bing's index is far harder for Copilot to retrieve and cite, whatever its Google rankings, which brings Bingbot access, Bing Webmaster Tools and IndexNow back into AI search work. Teams sample Copilot answers for their key questions, check that the cited pages are the ones they want cited, and fix indexing gaps first. They also keep consumer Copilot apart from Microsoft 365 Copilot in reporting, because the enterprise product answers from tenant data that outside marketing cannot influence.
Example
A software company finds that Copilot Search answers "best invoicing app for freelancers" by citing three review sites and a competitor, never its own comparison page. A check of Bing's index shows the page was never indexed; after fixing a canonical error and submitting the URL through IndexNow, the page becomes eligible to be retrieved.
Also called: Perplexity AI, Perplexity.ai, Perplexity search, Perplexity Pro
DocumentedCrawler contract, products and plan features per Perplexity's own documentation, blog and help center; vendor descriptions of its index and ranking are not independently audited.
Perplexity is an AI-native answer engine that responds to questions with answers synthesized from live web retrieval, marking claims with numbered citations and listing its sources. First released as Ask in December 2022, it now spans free and paid search, Deep Research, the Comet browser, the Computer agent and a developer API platform.
Perplexity is both a company and an answer engine built around search from the start rather than added to an existing chatbot. It dates its first release, Ask, to December 2022, launched its iPhone app on 2023-03-28 and committed then to "citations with every answer". A standard answer runs live retrieval, synthesizes a response and shows inline numbered references with a list of sources.
Several products surround that core. Deep Research, launched on 2025-02-14, runs many searches across hundreds of sources to produce a report; Perplexity's help center says Deep Research is becoming a built-in skill of Computer, the multi-model agent it launched on 2026-02-25, and that all plans keep free usage of Deep Research. Comet is its browser, covered separately. The developer platform documents a Search API returning ranked web results as JSON and an Agent API returning finished answers with citations, alongside Embeddings, Decisions and a private-preview Orchestrator API.
How it works
Perplexity documents two agents. PerplexityBot crawls to surface and link websites in Perplexity's search results, and Perplexity says it is not used to crawl content for AI foundation models; allowing it in robots.txt is how a site becomes eligible to appear. Perplexity-User fetches a page when a user's question needs it and, because a user requested the fetch, "generally ignores robots.txt rules". Perplexity says robots.txt changes may take up to 24 hours to be reflected, and it publishes the IP ranges of both agents as JSON files so sites can verify them.
Why it matters
Perplexity is the reference case of a citation-first engine, so it works as an early test of whether a page is retrievable and quotable at all. The levers are documented: allow PerplexityBot, keep pages crawlable and fast, and write passages that answer a question cleanly enough to be cited. Teams also record which third-party sources Perplexity cites for their category, since review sites, forums and documentation often shape its answers more than a brand's own pages. Referral clicks from Perplexity are attributable in analytics, which helps connect citations to visits.
Example
Asked "what is the best CRM for a 10-person agency?", Perplexity returns a short comparison with bracketed numbers after each claim and a list of sources. A review site cited after the answer's first sentence gets more attention than one that appears only further down the source list.
Also called: Claude web search, Claude search, Anthropic Claude, claude.ai
DocumentedObservedAvailability, citations and crawler roles per Anthropic's announcement and help center; the external search provider behind Claude's web search is not named in Anthropic's main documentation, and press evidence about it is observed only.
Claude is Anthropic's AI assistant. Its web search, launched on 2025-03-20 and available on all Claude plans worldwide since 2025-05-27, lets Claude look up current information during a conversation and answer with citations to the pages it used. Anthropic's help center says every web-search response includes citations so users can verify sources.
Claude web search is the capability that lets Claude query the web and cite what it finds. Anthropic launched it on 2025-03-20 as a feature preview for paid US users and updated its announcement on 2025-05-27 to say web search was available globally on all Claude plans. Anthropic's help center (updated 2026-10-07) says Claude invokes a search tool when a topic benefits from current information, that in the newer Claude experience it searches whenever that helps, and that every response includes citations with source links and, when appropriate, quotes. With search on, Claude can also fetch content from a specific URL, and image results are powered by Bing.
On Team and Enterprise plans an owner must first enable web search for the organization. Claude in Chrome, generally available on paid plans since 2026-08-26, extends Claude from reading search results to operating pages in the browser; that use is covered under computer-use and browsing agents.
How it works
Anthropic documents three agents that matter to publishers. ClaudeBot collects content that may be used for model training. Claude-User fetches pages when people ask Claude questions, and blocking it may reduce visibility in user-directed search. Claude-SearchBot crawls to improve search result quality, and blocking it prevents indexing for search. Anthropic says its bots honor robots.txt, including the non-standard Crawl-delay, and since 2026-10-30 it also recognizes a machine-readable training-preference signal based on an IETF aipref draft. How Claude's answers combine Anthropic's own crawl with an outside search provider is not fully documented: in March 2025, press evidence (a subprocessor listing, identical citations and a BraveSearchParams parameter) pointed to Brave Search.
Why it matters
Claude's crawler contract maps three agents to three decisions: training, search indexing and live user fetches. Blocking Claude-SearchBot or Claude-User to avoid training would also remove a site from Claude's search answers, so a training opt-out belongs on ClaudeBot. Because the search backend is only partly disclosed, retrieval should be verified empirically: ask Claude the questions buyers ask, record which pages it cites, and check server logs for Claude-User fetches during the test. Claude's citations make it easy to see whether a page was used.
Example
A B2B publisher blocks every Anthropic agent to opt out of training and later finds its reports missing from Claude's cited sources. Narrowing the robots.txt rule to ClaudeBot keeps the training opt-out while making the site eligible again for Claude-SearchBot and Claude-User.
DocumentedObservedDeepSearch and the search tools per xAI's announcement and API documentation; the company's 2026 rebranding is press-reported; how the consumer app selects and cites sources has limited independent corroboration.
Grok is the AI assistant built by xAI, available inside X and as a standalone app at grok.com. It can search both the open web and posts on X, and its DeepSearch mode, introduced with the Grok 3 beta on 2025-02-19, runs a research agent that weighs conflicting sources and returns a concise report.
Grok is a family of models and the consumer assistant built on them by xAI, Elon Musk's AI company. xAI introduced DeepSearch with the Grok 3 beta on 2025-02-19, calling it its first agent: a research mode built to reason about conflicting facts and opinions and distill them into a concise, comprehensive report, offered first to X Premium+ subscribers. Grok's API documentation shows how its retrieval works through two server-side tools: web_search, which lets Grok search the web in real time and browse pages, and x_search, which runs keyword and semantic searches over X posts, users and threads. Both return citations.
The corporate name has moved. SpaceX acquired xAI in February 2026, and in July 2026 the company announced on X that it was now SpaceXAI; October 2026 coverage calls it SpaceX Super Intelligence (SpaceXSI). The developer documentation still uses the xAI name, and the Grok product name is unchanged.
Why it matters
Grok's distinctive input is X. A question about a product, a launch or an incident can surface X posts next to web pages, so what a brand and its critics say on X feeds this surface directly and quickly. For teams tracking AI answers, Grok is the engine where social narrative and web content mix most visibly, which argues for an accurate, active presence on X and for monitoring Grok separately from engines that rely mainly on web indexes. Evidence on how Grok ranks or selects sources is thin, so conclusions should rest on repeated sampling.
Example
After a service outage, a user asks Grok "is Acme down right now?". The answer cites recent X posts from affected users and Acme's own status account next to a news article; a company that posts clear updates on X shapes that answer within minutes.
Also called: Meta AI assistant, meta.ai, Meta AI app, Muse Spark
DocumentedObservedSurfaces, models and features per Meta's newsroom; how Meta AI retrieves web content and whether it shows sources for it are not documented, so retrieval and citation behavior is observed only.
Meta AI is Meta's AI assistant, built into WhatsApp, Instagram, Facebook and Messenger and available as the Meta AI app and at meta.ai. Launched broadly with Llama 3 on 2024-04-18, it has run on Meta's proprietary Muse Spark model since 2026-04-08 and is adding web research, shopping and planning features.
Meta AI is the assistant Meta places across its own apps and on the web. Meta announced it broadly on 2024-04-18, built with Llama 3 technology and available on Facebook, Instagram, WhatsApp, Messenger and meta.ai. On 2026-04-08 Meta introduced Muse Spark, the first model of a proprietary Muse series from Meta Superintelligence Labs, which powers the Meta AI app and meta.ai first, starting in the US, with rollout planned into WhatsApp, Instagram, Facebook, Messenger, Threads and Meta's glasses.
Meta's posts describe a shopping mode that searches Facebook Marketplace alongside options from across the internet, answers that surface public posts and community content about places and trending topics, and, from 2026-07-24 in select markets, planning features and "research deep dives" that synthesize information from across the web. Meta says future answers will credit the creators whose posts they draw on; its posts do not describe how web sources are displayed.
Why it matters
Meta AI answers questions inside messaging and social apps, where a request about a restaurant, a product or a service is resolved without a results page in between. Its visibility mechanics are the least documented of the major assistants: Meta's posts describe no citation rules for web sources, so teams cannot rely on a crawler contract or a sources panel to diagnose absence. What Meta does say is that public posts, Reels and Marketplace listings feed answers, which makes a brand's public presence on Meta's own apps a plausible input. Measurement has to rely on direct sampling inside the apps.
Example
A WhatsApp user asks Meta AI for "a good bike repair shop near me that is open on Sunday". The answer may draw on public posts, maps data and web content, and the user may never visit a website; the shop learns about the exchange only if the customer calls or taps through.
Also called: YouChat, You.com search, You.com APIs, You.com Search API
DocumentedObservedCurrent positioning per You.com's own site; the December 2022 YouChat launch rests on contemporary press coverage because the original You.com blog post, archived only, could not be retrieved.
You.com is a company that began as an AI-assisted search engine and launched YouChat, one of the first chat assistants built into a search engine with current, cited answers, in December 2022. Its site now presents it as search and inference infrastructure for AI, selling web search, contents, answer and research APIs to developers and enterprises.
You.com started as a search engine and, in late December 2022, added YouChat, a conversational assistant placed inside its search results that could reach current information and cite sources when asked, according to coverage at the time, which compared it with ChatGPT. That made it one of the earliest products to pair a large language model with live search, weeks before Microsoft introduced the new Bing on 2023-02-07.
The company has since moved toward infrastructure. Its homepage, read on 2026-11-07, leads with "Search and inference for AI" and lists a Web Search API, Contents API, Answer API, Research API, Finance Research API, an inference router, an MCP server and agent skills, and promotes no consumer chat product. A January 2026 You.com post presents these APIs as core elements of its enterprise AI platform. In this dictionary's sense, You.com matters both as an early answer engine and as a retrieval provider that other AI products can build on.
Why it matters
You.com shows how AI search splits into two layers: the assistant people talk to and the retrieval services that feed it. A brand may never see You.com in its referral reports and still be read through its APIs when a third-party agent or application uses them to ground answers in the web. Which products use which provider is rarely disclosed, so the practical response is general: keep pages crawlable, fast and clearly structured, since retrieval APIs return passages and snippets that downstream models quote directly.
Example
A developer builds a procurement agent that calls a web search API to compare vendors. The agent's answer cites a vendor's pricing page that the developer's users never visit; whether that page was retrievable and quotable decided whether the vendor appeared at all.
Also called: Brave Search, Brave Leo, Leo, Ask Brave, Brave AI Answers, Answer with AI, Brave Search API
DocumentedObservedProducts and index claims per Brave's own pages (vendor statements about its own index); which third-party AI products use the Brave Search API is mostly undisclosed and known through press reports.
Brave Search and Leo are Brave's two AI surfaces. Brave Search runs on its own independent web index, with AI Answers that summarize results with cited sources and Ask Brave, a chat mode with follow-ups and Deep Research. Leo, the assistant inside the Brave browser, launched on desktop on 2023-11-02. The Brave Search API sells access to the index.
Brave's place in AI search has three parts. Brave Search runs on an index Brave builds itself, which Brave described in September 2025 as covering over 35 billion webpages; its "Answer with AI" feature, announced on 2024-04-17 and now called AI Answers, gives a concise summary with references to sources for question-like queries. Ask Brave, introduced on 2025-09-29, offers longer answers, chat follow-ups and a Deep Research mode that runs dozens of queries for cases where thoroughness matters more than speed. Leo, launched on desktop in Brave 1.60 on 2023-11-02, is the browser assistant: it summarizes pages and videos, answers questions about content, translates and rewrites, offers a choice of models and commits to not retaining chats.
The Brave Search API opens the same index to developers and AI companies for grounding model answers in web results. Brave states that grounding answers in search results does not guarantee they are true.
How it works
Because Brave runs its own crawler and index, a page's presence in Brave Search depends on Brave having discovered and indexed it, independently of Google or Bing. AI Answers and Ask Brave synthesize from those results and link the sources. The index can also reach other AI products through the API: in March 2025 TechCrunch reported evidence (a subprocessor listing, identical citations and a BraveSearchParams parameter) that Claude's web search used Brave.
Why it matters
An independent index is a separate gate. A page that ranks in Google can be missing from Brave's index, and then it is missing from Brave's AI Answers and from any AI product that grounds its answers through the Brave Search API. Since those downstream customers are rarely named, Brave coverage is worth checking directly: search for key pages on search.brave.com, confirm they appear with current titles and descriptions, and keep pages open to well-behaved crawlers in general instead of allowlisting only the largest bots.
Example
A startup searches its own name on search.brave.com and finds its newest product page missing while an outdated press release ranks first. Because that index may also feed third-party assistants, fixing internal links and the sitemap improves discovery for more than one surface.
Brave, Brave Leo, the AI browser assistant, is now available in Brave for desktop, published 2023-11-02, read 2026-11-07: https://brave.com/blog/leo-launch/
Brave, Brave unveils new privacy-focused AI answer engine (Answer with AI; updated 2024-08-23), published 2024-04-17, read 2026-11-07: https://brave.com/blog/answer-with-ai/
Also called: Le Chat, Mistral Vibe, Vibe, Vibe for work, chat.mistral.ai
DocumentedLaunch, features and the rename per Mistral's own pages; the late-May 2026 timing of the rename comes from press coverage of Mistral's AI Now Summit, because Mistral's product page gives no date.
Mistral Le Chat is the AI assistant of the European model maker Mistral AI, launched in beta on 2024-02-26 and offering web search with detailed citations by 2025-02-06. In May 2026 Mistral renamed it Vibe, merging the chat assistant with work-automation and coding agents; existing conversations and plans carried over.
Le Chat launched on 2024-02-26 as a conversational entry point to Mistral's models (Mistral Large, Mistral Small and a prototype called Mistral Next) and, at launch, could not access the internet. Mistral's relaunch on 2025-02-06 described answers that balance pre-trained knowledge with web search, journalism and other sources, with evidence-based answers and detailed citations. A Deep Research preview followed on 2025-07-17, producing structured, reference-backed reports.
At its AI Now Summit in late May 2026, Mistral renamed Le Chat to Vibe. Its product page says "Le Chat is now Vibe" and describes a chat assistant and agent for work and code, with web search on the free plan and sourced analysis across the web, internal documents and connected tools. Mistral's documentation labels the original Deep Research a legacy Chat feature, replaced by a Deep Research skill in Work that still produces inline citations and a sources list. Enterprise customers can deploy Vibe on premises or in a private cloud.
Why it matters
Mistral is a European model maker, and its assistant is a distinct answer surface for people and companies that choose a European provider, including enterprises running it on their own infrastructure. Citation checks limited to US engines can miss how questions in French, German or Italian resolve there. The rename matters for monitoring too: prompts, tools and reports that still say "Le Chat" refer to the same product, now called Vibe. Evidence on how it selects sources is limited, so findings should come from direct sampling in the languages that matter to the business.
Example
A French software vendor asks Vibe, in French, which invoicing tools meet France's e-invoicing requirements, checks whether its documentation is among the cited sources, then repeats the test in ChatGPT and Gemini to compare which French-language sites each engine relies on.
Also called: DeepSeek app, DeepSeek Chat, chat.deepseek.com, DeepSeek-R1, DeepSeek-V4
DocumentedObservedModels, app features and the search API per DeepSeek's own announcements and documentation; no DeepSeek crawler or robots.txt documentation was found, so its retrieval behavior toward publishers is unknown.
DeepSeek is a Chinese AI lab whose free chat assistant, on the web at chat.deepseek.com and in mobile apps launched on 2025-01-15, offers a web search option and a reasoning mode. Its open-weight models, from R1 (2025-01-20) to the V4 family (preview 2026-04-24), power the assistant, and an in-house search engine backs its Search API.
DeepSeek is both a model developer and a consumer assistant. Its app, launched on 2025-01-15 and powered then by DeepSeek-V3, listed web search and a Deep-Think reasoning mode among its features; the web interface offers the same controls. DeepSeek-R1, released on 2025-01-20 under the MIT license, brought a reasoning model to the web and API. The current generation is the V4 family: V4-Pro and V4-Flash went live as an open-weight preview on 2026-04-24 with a one-million-token context, and V4-Pro reached general availability on the app, web and API, under an "Expert Mode", on 2026-08-13.
DeepSeek also offers a Search/Fetch API that it says is powered by its in-house search engine, returning ranked results with titles, URLs, snippets, summaries and an authority tier. What DeepSeek does not publish, in the English documentation reviewed for this entry, is a crawler name, a robots.txt policy or a description of how its assistant displays sources.
Why it matters
DeepSeek's assistant is a distinct answer surface, but publishers have little to work with: without a documented crawler there is no robots.txt lever, and without documented citation rules the only way to learn how a brand appears is to sample the assistant with search switched on and off. An in-house search engine also means its retrieval may not mirror Google or Bing results. Teams whose buyers use DeepSeek should include it in sampling and keep facts on their own pages explicit, current and easy to quote.
Example
A hardware brand asks the DeepSeek assistant to compare two of its routers, once with search off and once with it on. Without search the answer repeats specifications of an older model; with search it cites a recent review, which shows which path needs correcting.
Also called: Rufus, Alexa for Shopping, Amazon AI shopping assistant, Amazon Rufus
DocumentedFeatures, data sources, rename and ad placements per Amazon's own pages; how the assistant ranks products against each other is not documented.
Amazon Rufus was Amazon's generative AI shopping assistant, announced in beta on 2024-02-01 and later offered to all US customers in the Amazon Shopping app and on desktop. On 2026-05-13 Amazon renamed it Alexa for Shopping, merging Rufus with Alexa+. It answers product questions from Amazon's catalog, reviews, community Q&A and the web.
Rufus is the canonical vertical answer engine: an assistant that lives inside a retailer's own search and product pages. Amazon announced it on 2024-02-01 as a beta for a small subset of US customers in its mobile app and later said it was available to all US customers in the Amazon Shopping app and on desktop. Amazon describes it as trained on its product catalog, customer reviews, community Q&As and information from across the web; on a product page it answers from the listing details, reviews and Q&A.
On 2026-05-13 Amazon renamed Rufus Alexa for Shopping, combining it with Alexa+ into one assistant. Amazon's advertising pages describe questions asked from the main search bar, personalized shopping guides, comparisons, up to a year of price history and automated deal-finding and cart-building. Sponsored products and brands can appear inside the conversation, and existing sponsored ads campaigns are automatically eligible, so paid and organic placements now share the same assistant.
Why it matters
For brands that sell on Amazon, this assistant is often the last answer before a purchase, and it draws on material the brand partly controls: listing details, plus the volume and substance of reviews and community Q&A. Gaps in a listing become gaps in the assistant's answer. Because sponsored placements now appear inside conversations, teams should report paid and organic presence separately. External content still counts, since Amazon says the assistant also uses information from across the web, which makes independent reviews and guides part of the same effort.
Example
A shopper on a tent's product page asks "is this good for winter camping?". The assistant answers from the listing's specifications and from reviews that mention snow and cold; a listing that omits its temperature rating leaves the answer to reviewers' impressions.
Amazon, Amazon announces Rufus, a new generative AI-powered conversational shopping experience (with later updates), published 2024-02-01, read 2026-11-07: https://www.aboutamazon.com/news/retail/amazon-rufus
Also called: Apple Intelligence, Siri, Siri AI, Apple Foundation Models
DocumentedFeatures, models, availability and publisher controls per Apple's newsroom and Applebot documentation; Apple does not detail when Siri answers display source links.
Apple Intelligence and Siri form Apple's AI system for iPhone, iPad and Mac, first released on 2024-10-28. Its rebuilt assistant, Siri AI, announced on 2026-06-08 and rolling out in beta from 2026-09-14, answers questions using information from across the web, and Apple says such answers may include links to sources.
Apple Intelligence arrived with iOS 18.1, iPadOS 18.1 and macOS Sequoia 15.1 on 2024-10-28, initially in US English, running on device and on Apple's Private Cloud Compute servers; optional ChatGPT access in Siri and Writing Tools was announced for December 2024. At WWDC on 2026-06-08 Apple introduced Siri AI, which can go to the web for up-to-date information and is built into Spotlight on iPad and Mac. It began rolling out in beta in English on 2026-09-14 across iOS 27 and Apple's other platforms, with French, Japanese, Korean, Portuguese and Spanish planned next. Apple says it runs on the next generation of Apple Foundation Models, "custom-built in collaboration with Google and its Gemini models".
Siri AI is not initially available in the EU on iOS, iPadOS or watchOS, which Apple attributes to the Digital Markets Act, nor in China. Apple's Applebot page (published 2026-09-04) says crawled data may provide context for answers to broad world-knowledge questions in Siri and Search that may include links to sources.
Why it matters
Siri AI puts a web-grounded answer engine into the operating system, so a question asked by voice or in Spotlight can be answered without opening a browser or a search engine. For publishers, Apple's crawler documentation is the control surface: Applebot crawls for search features and answers, Applebot-Extended governs whether crawled data may train Apple's models, and the nosnippet rule keeps a page from being used as context for AI answers while it stays discoverable. Apple's documentation describes no reporting on Siri answers, so impact shows up only indirectly, through referral traffic and direct sampling.
Example
An iPhone user asks Siri "what time does the Louvre close on Wednesdays?". Siri AI answers from web information and may link to the museum's site; a page marked nosnippet would not be used as context for that answer, though Apple says it would remain discoverable through Spotlight, Siri and Safari.
Also called: citation UI, sources list, Sources button, source cards, inline citations, link panel
DocumentedObservedEach engine documents its own display; the comparison across engines and the consequences for clicks are observed, and layouts change often.
Sources panel, or citation UI, names the part of an AI answer interface that shows which web pages the answer relied on, through inline citation markers, numbered references, link cards or a side panel listing cited and related pages. It is where a citation becomes visible and clickable, and every engine designs it differently.
The sources panel, or citation UI, covers every element an answer engine uses to attribute its answer: markers or links attached to sentences, numbered references, favicon chips, cards above or below the answer, and expandable lists or side panels. Engines document their own versions. ChatGPT shows citations in the response and a Sources button that lists cited sources and other relevant links. Gemini Apps show a Sources button, at the bottom or inline, that opens a side panel. Copilot Search links sentences inline and reveals every source with one click. Google added a right-hand link display to AI Overviews on desktop in August 2024 and has since added more inline links, hover previews and labels such as "Preferred" and "Subscribed". Claude's help center says every web-search response includes citations, and Perplexity has promised citations with every answer.
The term describes the display layer only. Which pages are retrieved and chosen for citation is decided earlier, in retrieval and citation selection.
Why it matters
Being retrieved, being cited and being clicked are three different outcomes, and the citation UI sits between the last two. A source attached to the first sentence of an answer, with its name visible, is easier to notice than the seventh item in a collapsed list, so the same "citation" can carry very different value. Teams should define what counts as cited for each engine in reports and contracts, record position and prominence along with presence, and re-check regularly, because engines redesign these elements often.
Example
In one engine a brand's guide is the inline citation after the answer's opening sentence; in another it is source 7 of 9 in a panel the user has to open. A raw tally counts both as citations, and a report that weights prominence separates them.
Also called: AI Overviews carousel, link carousel, article carousel, news carousel in AI answers, link cards
DocumentedObserved"Link carousel" is an industry label; Google documents carousels of links for some news and developing-topic answers but publishes no single named component or triggering rules.
AI Overview link carousel is the common name for a horizontally scrolling row of link cards attached to a Google AI Overview or AI Mode answer, separate from the links placed inline in the text and from the link list shown beside the answer on desktop. Google documents article carousels for some news and developing-topic answers, highlighting Preferred Sources.
The term names one of several ways Google displays supporting links in its AI answers. Google has described a right-hand link display for AI Overviews on desktop, reached on mobile by tapping site icons (2024-08-15); links placed directly in the text, with hover previews on desktop (2026-05-06); and links at the end of some responses pointing to in-depth articles, which gained image thumbnails in October 2026. Carousels are the scrollable variant. On 2026-05-27 Google described a prominent carousel of links, including Preferred Sources, for some questions about developing news, plus a coming carousel of firsthand perspectives from forums and social media; on 2026-10-15 it said article carousels were rolling out to more locations and languages.
Because Google names no single "carousel" component, the label is descriptive. Layouts vary by query type, market and device, and change through continuing tests.
Why it matters
Carousel cards carry an image, a title and a source name, which makes them more visible than a small inline link, and for news queries they can highlight sources the user marked as preferred. That points to concrete levers: a clear, current headline, a strong lead image, and readers who add the site as a Preferred Source. Google has not published how cards are chosen, so treat these as plausible inputs. Reports should also record which link format a citation appeared in, since a carousel card and a link deep in a side list are not equivalent.
Example
For "latest on the port strike", an AI Overview shows a two-sentence summary followed by a carousel of four news articles with thumbnails, one labeled Preferred because the user follows that outlet; a regional paper with a clear headline and image appears as the second card.
Also called: featured snippets, position zero, answer box, direct answer box
Documented
Featured snippet is the name of a Google search result shown in a special box where the usual order is reversed: a passage extracted from one page, answering the query, appears first, followed by that page's title and link. Google's systems choose the page automatically. It is the single-source ancestor of today's multi-source AI answers.
Featured snippets are a long-standing Google Search feature that lifts a passage from one web page and displays it prominently, as a paragraph, list or table, with the source's title and URL. Google states that its systems determine whether a page would make a good featured snippet for a given search, and there is no markup to request one. Site owners can prevent them with the nosnippet rule, which also removes regular snippets, or make them less likely with a lower max-snippet value, which Google says does not guarantee removal; text marked with data-nosnippet is excluded from both. A featured snippet also appears when a user expands a question in the People Also Ask group.
The informal name "position zero" reflects its place above the first organic result. As a bridge term, it links snippet-era answer engine optimization, which aimed to win this box, to AI search, where answers draw on several sources at once.
Why it matters
The featured snippet established the behavior AI answers extend: the user gets the answer on the results page and may never click. Tactics that won snippets (a direct answer under a question heading, concise definitions, clean lists and tables) carry over partly to AI search, where engines also extract self-contained passages. The difference is that an AI answer blends several sources, so being one of the cited passages replaces being the single winner. Google applies the same snippet controls to its AI features, so restricting snippets also restricts how a page can appear in AI Overviews and AI Mode.
Example
Searching "what is BM25" may show a boxed two-sentence definition taken from one tutorial, with its link underneath. The same tutorial might be one of five cited sources in an AI Overview answering a longer question about ranking functions.
Also called: Google knowledge panel, knowledge box, entity panel, Knowledge Graph panel
DocumentedDefinition and claiming process per Google's help center; the link between Knowledge Graph presence and AI answers is an inference that Google does not document for its AI features.
Knowledge panel is the name of the information box Google shows when someone searches for an entity in its Knowledge Graph, such as a person, place, organization or thing. Google's systems generate it automatically from sources across the web and from data partners, and the entity's official representatives can claim it and suggest changes.
Knowledge panels are Google's structured summaries of entities, shown alongside search results. Google says they appear for entities in the Knowledge Graph, that their information comes from various sources across the web and, for some topics, from data partners, and that they update automatically as information on the web changes. A person or organization that is the subject of a panel, or its official representative, can claim it through Google's verification process and suggest edits; anyone else can send feedback through the link on the panel.
A knowledge panel presents stored facts about a recognized entity, whereas an AI Overview generates an answer. As a bridge term, it connects classic entity SEO to AI search: both depend on a search engine resolving a name to the right entity and holding consistent facts about it.
Why it matters
Entity resolution comes before citation. If a search engine confuses a brand with a namesake, or holds outdated facts about it, generated answers can inherit the error. A claimed, accurate knowledge panel is a visible sign that Google recognizes the entity and its key facts. Google documents no direct link between the panel and AI Overviews, so treat it as a health check. Teams keep official facts consistent across their own site, structured data, Wikidata and major profiles, claim their panel, and correct errors through Google's feedback channel.
Example
A software company named Atlas shares its name with several others, and its knowledge panel shows the wrong founding year and logo. After claiming the panel and aligning its About page, Organization markup and Wikidata entry, the company has one consistent record, which lowers the risk that AI answers mix in a namesake's facts.
Also called: PAA, related questions, related questions group, People also ask box
DocumentedObservedThe feature is documented in Google's visual elements gallery; its use as a map of AI fan-out questions is a practitioner inference.
People Also Ask is Google's expandable group of questions related to a search, shown on the results page; expanding a question reveals a short answer, shown as a featured snippet from one page. Google's documentation calls it a related questions group. For AI search work it is a free, visible sample of the questions people ask around a topic.
People Also Ask, which Google's documentation calls a related questions group, is a cluster of questions related to what the user searched for. When the user expands a question, Google shows a featured snippet that answers it, extracted from one web page and linked to it. The questions reflect what Google's systems consider related needs around the original query.
As a bridge term it links classic SEO to AI search in one practical way: the questions in the group approximate the subtopics that an AI system exploring the same query might cover. That link is an inference. Google documents query fan-out for AI Overviews and AI Mode, describing related searches across subtopics, but does not say that fan-out queries match People Also Ask questions.
Why it matters
AI answers are assembled from sub-questions, and People Also Ask offers a free, public sample of the questions around a topic. Teams use it to find gaps in their coverage, then answer each relevant question clearly on the right page, which serves both the classic box and AI engines looking for self-contained answers. It is a starting map and an incomplete one, because AI systems generate their own sub-queries, so it works best combined with direct sampling of AI answers for the same topic.
Example
Under "generative engine optimization", the group lists questions such as "How is GEO different from SEO?" and "Does GEO work for B2B?". A team that answers each one in a short, clearly headed section covers likely fan-out sub-questions and the box itself.
Also called: Atlas, Atlas browser, OpenAI Atlas, ChatGPT browser
DeprecatedOpenAI announced the deprecation on 2026-07-09 and Atlas stopped working on 2026-08-09; its browser-agent functions moved into ChatGPT and Codex. OpenAI's help center now shows the retirement without its dates; the dates come from press coverage of OpenAI's announcement.
ChatGPT Atlas was OpenAI's AI-native web browser, launched on macOS on 2025-10-21 with a ChatGPT sidebar that could read the open page and an agent mode that could browse on the user's behalf. OpenAI announced its deprecation on 2026-07-09 and switched it off on 2026-08-09, moving those capabilities into ChatGPT and Codex.
Atlas was a standalone browser from OpenAI built around ChatGPT. It launched worldwide on macOS on 2025-10-21 for Free, Plus, Pro and Go users, with Windows, iOS and Android versions announced as coming soon; none of them shipped. Its Ask ChatGPT sidebar answered questions about the page in view, and agent mode, a preview for Plus, Pro and Business users, let ChatGPT navigate and act inside the browser.
On 2026-07-09 OpenAI announced it was deprecating Atlas and moving browser-based agentic capabilities into ChatGPT and Codex. Atlas stopped working on 2026-08-09, and bookmarks, tabs and history did not migrate automatically. The successors are the ChatGPT desktop app, which brings Chat, Work and Codex together with a built-in browser, and a cloud browser that runs ChatGPT Work's web tasks on a separate machine. The term now refers to a retired product, though it remains common in 2025 and 2026 articles about AI browsers.
Why it matters
Atlas showed a direction that outlived it: an agent that reads and acts on a page on a person's behalf. The same functions continue in ChatGPT's built-in and cloud browsers and in competing products such as Comet, Gemini in Chrome and Copilot in Edge. For site owners, the lessons carry over: pages should be readable without fragile scripts, key facts should sit in plain text, and forms and checkouts should work for an automated visitor acting with the user's permission. Advice in older guides that mentions Atlas should be applied to these successors.
Example
A 2025 guide recommends testing a pricing page "in Atlas agent mode". The equivalent test now is to ask ChatGPT, through the desktop app's built-in browser or ChatGPT Work's cloud browser, to compare that page with a competitor's and check whether it reads the plans correctly.
Also called: Comet, Comet browser, Comet Assistant
DocumentedLaunch, availability and plan limits per Perplexity's blog, changelog and help center; usage share is out of scope here.
Perplexity Comet is the web browser made by Perplexity, launched on 2025-07-09 for Max subscribers, free to everyone since 2025-10-02 and available on Windows, Mac, Android and iOS since March 2026. Its built-in Comet Assistant explains and compares pages and can act for the user, for example by booking a meeting or sending an email.
Comet puts Perplexity's answer engine and an assistant into the browser itself. At launch on 2025-07-09 Perplexity described an assistant that can act on the user's behalf, such as booking a meeting or sending an email based on something the user saw, comparing the current page with something read earlier, finding other sellers of an item and explaining highlighted text. Access began with Perplexity Max subscribers and a waitlist. On 2025-10-02 Perplexity released Comet free to everyone, adding background assistants that work asynchronously, and a changelog dated 2026-03-26 says Comet is available on iOS, Android, Mac and Windows.
The browser and its agent are priced differently: Perplexity's plan comparison lists browser-agent (Comet Assistant) queries as unavailable on the free plan and subject to monthly limits on Pro and Max. The term covers the browser and its assistant; Perplexity's answer engine is described under its own entry.
Why it matters
Together with agentic features in Chrome, Edge and ChatGPT, Comet moves the answer engine from a destination website into the browser frame. When an assistant summarizes or acts on a page, it reads content the user may never scroll through, so what survives that mediation depends on clear structure, plain text for key facts and pages that do not break for an automated visitor. Teams should test their important flows (pricing, sign-up, booking) by asking the assistant to complete them, and fix what it misreads.
Example
A user reading a hotel's booking page in Comet asks the assistant to compare its cancellation policy with two hotels viewed earlier and then reserve the cheapest refundable room. A hotel whose policy appears only inside an image is described inaccurately or skipped.
DocumentedObservedLaunch, SDK, directory rules and the move to plugins per OpenAI's announcements and help center; how ChatGPT suggests or ranks apps inside conversations is not documented.
Apps in ChatGPT are third-party applications that users can call inside a ChatGPT conversation, with their own interface shown in the chat. OpenAI launched them on 2025-10-06 with the Apps SDK, built on the Model Context Protocol, opened an app directory to developer submissions in December 2025 and folded it into a Plugin directory on 2026-07-09.
Apps in ChatGPT let a service such as a travel site, a design tool or a music app run inside the conversation: the user names the app or ChatGPT suggests it, the app returns results and interface elements in the chat, and the user continues without leaving ChatGPT. OpenAI introduced apps on 2025-10-06 for logged-in Free, Go, Plus and Pro users outside the EEA, Switzerland and the UK, with pilot partners including Booking.com, Canva, Coursera, Figma, Expedia, Spotify and Zillow, and released the Apps SDK in preview as an open standard built on the Model Context Protocol. In December 2025 it opened submissions from verified developers to an in-product app directory, stating that it does not guarantee any placement or ranking there.
On 2026-07-09 OpenAI replaced the app directory with a Plugin directory. Apps remain the integrations that connect to outside data and actions, while plugins bundle them with skills and templates; existing app connections were unaffected.
Why it matters
Apps offer a second way to be present in an AI assistant. A cited source earns a mention inside someone else's answer; an app puts the brand's own service, data and interface into the conversation, often at the moment of action, such as booking or ordering. The trade-off is effort and control: an app requires building on the SDK and passing review, and OpenAI promises no placement. Teams in categories where users complete tasks (travel, commerce, scheduling, software) should weigh an app alongside citation work.
Example
A user planning a trip asks ChatGPT for hotels in Lisbon under $200 a night. With a travel company's app connected, ChatGPT shows that company's live listings in an interactive card, and the user books there; a competitor without an app appears only if its pages are cited in a text answer.
Also called: deep research, Deep Research mode, Deep Search, research agents, agentic research mode
DocumentedObservedLaunches and stated behavior per each vendor's announcements; how many pages a run reads and which ones it cites are vendor-described or observed, and vary by product.
Deep Research names a category of agentic research modes in AI assistants that plan a multi-step investigation, run many searches, read large numbers of pages and return a long, cited report after several minutes. Google's Gemini Deep Research (2024-12-11) and OpenAI's deep research (2025-02-02) were early examples, and most major assistants now offer one.
A deep research mode trades speed for coverage. Instead of one retrieval round, the system drafts a research plan (often shown to the user for approval), searches iteratively, reads results, refines its queries and writes a structured report with citations. Google launched Deep Research in Gemini Advanced on 2024-12-11, describing a multi-step plan the user can revise and a report with links to sources. OpenAI introduced deep research in ChatGPT on 2025-02-02, saying it finds, analyzes and synthesizes hundreds of online sources. Perplexity's Deep Research (2025-02-14), xAI's DeepSearch (2025-02-19), Google's Deep Search in AI Mode (2025-07-16) and Mistral's Deep Research preview (2025-07-17) followed.
The category keeps shifting. OpenAI removed its legacy deep research mode on 2026-03-26 while keeping the current one; Perplexity is turning Deep Research into a skill of its Computer agent; Mistral's original mode is now legacy, replaced by a skill in Vibe's Work mode; and Ask Brave includes a Deep Research mode.
How it works
A typical run restates the question as a plan of sub-questions, issues dozens or hundreds of searches, opens and reads pages (and, where connected, files or email), decides when coverage is sufficient, then writes a report with a summary, sections and citations. Vendors describe scale differently: Google's overview page speaks of browsing hundreds of websites, OpenAI of hundreds of online sources, Perplexity of numerous searches across hundreds of sources in two to four minutes at launch, and Brave of dozens of queries over thousands of pages. Runs take minutes, and some products let users edit the plan or steer a run while it works.
Why it matters
One deep research run can open far more pages than a one-shot answer, so it multiplies the moments when a site can be read and cited, and it rewards depth: comparison pages, documentation, original data and clear methodology are what these reports are built from. The fetches typically go through each vendor's user-triggered agents, so blocking AI fetchers also removes a site from these reports. Because the final report cites only part of what was read, teams should track both fetches in server logs during tests and the citations that survive into the report.
Example
A buyer asks a deep research mode for "a shortlist of SOC 2 compliance automation vendors with pricing". Over several minutes the agent reads vendor pricing pages, documentation, review sites and forum threads, then writes a ten-page report citing a dozen of them; vendors without public pricing are described from third-party estimates.
Also called: Preferred Sources, preferred source, Top Stories source preferences, Add as a preferred source on Google
DocumentedFeature, availability and placement per Google's announcements; the click figure is Google's own, and the effect of preferences on any one site's visibility is not independently measured.
Google Preferred Sources is a Search setting that lets people pick news sites and other publishers they want to see more often. Launched on 2025-08-12 in the US and India for Top Stories, it reached all supported languages from 2026-04-30 and, since 2026-05-27, marks links from chosen sites as "Preferred" in AI Overviews and AI Mode.
Preferred Sources lets a user select outlets, through an icon beside Top Stories or in Search personalization settings, so that Google shows more of their fresh, relevant articles. Google launched it on 2025-08-12 for all users in the US and India, with no cap on the number of sources and with content from other sites still shown. It expanded to English-language users worldwide on 2025-12-16 and, on 2026-04-30, began rolling out globally in all supported languages. On 2026-05-27 Google said links from preferred sites would carry a "Preferred" label in AI Overviews and AI Mode, that any website publishing fresh content can be added, and that people are twice as likely to click through to a Preferred Source, a figure Google reports itself. On 2026-10-15 it said the news carousels in its AI answers highlight Preferred Sources.
Publishers can point readers to the setting with an "Add as a preferred source on Google" button.
Why it matters
Preferred Sources is the clearest case of a user-declared brand preference feeding a major search surface, now including Google's AI answers. A publisher cannot buy a place in a reader's list; it earns one by being a source people want to see again, then asking them to add it. The practical steps are simple: place Google's button or link where loyal readers will see it, publish fresh content regularly, and track referrals from Top Stories and AI answers. Google has not published how much a preference changes what is shown, so treat the effect as plausible and unmeasured.
Example
A regional business newspaper adds an "Add as a preferred source on Google" button to its newsletter. Readers who tap it see its articles more often in Top Stories, and its links carry a Preferred label when an AI Overview cites it.
Also called: browser agents, browsing agents, computer-use agents, CUA, agentic browsers, auto browse
DocumentedEmergingIndividual products are documented by their vendors; the category's names, the ways agents identify themselves and the rules that govern them are still unsettled.
Computer-use and browsing agents are AI systems that operate a browser or a whole computer for a user: they view the screen, click, type, fill forms and move between pages to finish a task. Examples include Claude's computer use and Claude in Chrome, Gemini in Chrome's auto browse, Copilot in Edge, Comet's assistant and ChatGPT's cloud browser.
These agents use the web the way a person does, through a rendered browser, rather than fetching raw HTML for an index. Anthropic opened computer use in public beta on 2024-10-22, letting Claude look at a screen, move a cursor, click and type, and made Claude in Chrome generally available on paid plans on 2026-08-26. OpenAI launched Operator on 2025-01-23, folded it into ChatGPT agent in July 2025, and has since retired ChatGPT agent in favor of ChatGPT Work, a cloud browser and the desktop app's built-in browser. Google added auto browse to Gemini in Chrome on 2026-01-28 for US Google AI Pro and Ultra subscribers. Microsoft's Copilot Actions in Edge became "Browse with Copilot" on 2026-05-13, when Copilot Mode was retired. Perplexity's Comet and Brave's opt-in browser agent belong to the same family.
The term excludes crawlers that index pages and simple fetchers that retrieve one URL for a chat answer, though the boundaries blur.
How it works
The agent receives a goal, observes the page (as a screenshot, an accessibility tree or both), chooses an action, performs it and repeats until the task is done or it needs the user. Vendors add guardrails: Google says auto browse pauses for confirmation before purchases or social posts; OpenAI's cloud browser asks before opening new sites by default and requires confirmation for payments and bookings; Anthropic checks each action against the user's request with a classifier and scans page content for prompt injection. Identification is improving: OpenAI's cloud browser signs its requests with Web Bot Auth (HTTP Message Signatures, RFC 9421) and a Signature-Agent header naming https://chatgpt.com, which CDNs such as Cloudflare, Akamai and Vercel can verify. Sites can also expose actions directly: ChatGPT's site tools use WebMCP, a proposed web standard, inside the desktop app's browser.
Why it matters
These agents fetch like users: they run JavaScript, use the user's sessions and can complete purchases, so robots.txt rules written for crawlers do not clearly govern them. Some now identify themselves cryptographically, while others are hard to tell apart from human visits, which complicates bot policy and analytics alike. In practice, teams should verify signed agents instead of blocking all automation, keep critical flows (pricing, sign-up, checkout) usable without fragile scripts or challenges designed only for humans, and test those flows by asking an agent to complete them.
Example
A user asks an agent to renew a domain. The agent navigates the registrar's site, dismisses a cookie banner and stalls at a CAPTCHA built for humans; a registrar that verifies signed agent traffic and offers a clean renewal flow keeps the renewal.
Also called: Gemini in Chrome, Chrome auto browse, auto browse, Chrome Gemini side panel
DocumentedFeatures and availability per Google's announcements; how auto browse behaves on third-party sites is not documented in detail, and usage figures are unpublished.
Gemini in Chrome is Google's AI assistant built into the Chrome browser, opening in a side panel that can read the current page and other open tabs. Its auto browse feature, launched in preview on 2026-01-28 for US AI Pro and Ultra subscribers, lets Gemini carry out multi-step tasks on websites, pausing for confirmation before purchases.
Gemini in Chrome places Google's Gemini model in a persistent side panel of the Chrome browser. The assistant answers questions about the page in view, compares information across open tabs, and connects to Google apps such as Gmail, Calendar, Maps and Google Shopping when the user turns them on. Google expanded the side panel to India, New Zealand and Canada on 2026-03-11, in more than 50 additional languages.
Auto browse is the agentic part: given a goal, Gemini navigates sites, fills forms and adds items to carts on the user's behalf. At launch it was a US preview for Google AI Pro and AI Ultra subscribers on desktop; Google brought it to Android for the same US subscribers on 2026-08-18. Google states that auto browse pauses to ask for explicit confirmation before sensitive actions such as purchases or social posts, and that Chrome will support the Universal Commerce Protocol. This entry covers Google's browser assistant; standalone AI browsers such as ChatGPT Atlas and Perplexity Comet have their own entries.
Why it matters
Building an agent into a mainstream browser moves agent-mediated browsing beyond dedicated AI browsers. A site visited through auto browse is read and operated by a model: unclear buttons, interstitials, forced logins and JavaScript-only flows can stop the task, and the user may never see the page itself. Teams can test key journeys (search, configure, add to cart, book) with an agent, keep labels and form fields explicit, and watch logs and analytics for automated sessions. Google has not published usage figures for auto browse, so its traffic weight is unknown.
Example
A subscriber asks Gemini in Chrome to reorder last month's coffee beans and book parking near a stadium for Saturday. Auto browse opens the retailer and the parking site, fills in the forms, and stops at each payment step for the user to confirm.
Also called: shopping in ChatGPT, ChatGPT product results, ChatGPT merchant program, Instant Checkout (scaled back)
DocumentedObservedMerchant program and feed options per OpenAI's merchant page and commerce documentation; the March 2026 shift of Instant Checkout to merchant apps is press-reported with an OpenAI statement, and OpenAI's merchant page confirms it is moving away from in-chat checkout.
ChatGPT shopping is the product-discovery experience inside ChatGPT: when a request shows buying intent, ChatGPT can show product results with images, prices and merchant links, drawn from merchant feeds, partner catalogs and crawled web content. Purchases complete on the merchant's own site or app since OpenAI moved away from in-chat Instant Checkout in 2026.
ChatGPT shopping covers the product cards, comparisons and merchant lists ChatGPT shows when a user researches a purchase. OpenAI's merchant page states that shoppers see rich results with images, pricing and key details, that ChatGPT matches products using shoppers' preferences and past interactions, and that purchases are completed on merchant-owned websites or apps with no fee on purchases that start in ChatGPT. Shopping is live for US users, with more regions planned.
Merchants can supply a structured product feed in OpenAI's published format, by SFTP, API or a commerce platform; Shopify and Etsy catalogs are already integrated. A feed is optional because ChatGPT can crawl a site, but OpenAI presents it as the way to control the accuracy and freshness of price and availability. On 2025-09-29 OpenAI and Stripe launched Instant Checkout so US users could buy from Etsy sellers inside the chat. In March 2026 OpenAI told Digital Commerce 360 that Instant Checkout was moving to merchants' ChatGPT apps and that it would prioritize search and product discovery.
Why it matters
Buyers increasingly compare products inside ChatGPT before visiting any store, so a merchant missing from its feeds or catalogs, or showing stale prices there, risks being absent or misdescribed at the moment of choice. Teams act by applying to the merchant program or confirming their Shopify or Etsy catalog is current, keeping product data and reviews consistent with the site, and tracking ChatGPT referrals to product pages. Because checkout went back to merchant sites, the landing page after the click still decides the sale, and it should match what the product card promised.
Example
A user asks ChatGPT for a quiet dishwasher under $900. ChatGPT shows four product cards with prices and images; tapping one lists retailers that carry it, and each link opens the retailer's own product page where the purchase happens.
Also called: multimodal search, search by image, camera search, Google Lens, Circle to Search
DocumentedVolumes are Google's own figures; how visual queries select and cite sources in AI answers is not documented.
Visual search is searching with an image instead of, or alongside, typed words: pointing a phone camera, uploading a photo or circling something on screen, then getting results or an AI answer about what is shown. Google reported on 2026-10-28 more than 50 billion visual searches per month, and one in five AI Mode searches use voice, images or documents.
Visual search covers queries whose main input is an image or a live camera view. On Google the main entry points are Google Lens and Circle to Search, and AI Mode accepts images and documents as part of a conversation. The answer can be a set of visually similar results, a product match with shopping links, or a generated explanation with cited sources.
In Alphabet's Q3 2026 remarks, Sundar Pichai said Google had crossed 50 billion visual searches per month and that one in five AI Mode searches globally use voice, images or documents. Search Console's generative AI performance report separates 'Web: multimodal' impressions, which Google defines as image-based searches such as Lens and Circle to Search, from typed queries. The term excludes classic image search, where a typed query returns pictures; that is a different retrieval problem.
Why it matters
A visual query starts from an object, a label, a screenshot or a scene, so the pages that win are those a system can tie to that object: clear product images, descriptive file names and alt text, structured product data, and text that names the item precisely. For retailers, local businesses and anyone selling physical goods, it is a discovery path that skips typed keywords entirely. Teams can now check multimodal impressions separately in Search Console's generative AI report, audit image quality and markup on key pages, and test how assistants identify and describe their products from a photo.
Example
A runner photographs a friend's trail shoe with Google Lens and asks in AI Mode whether it suits wide feet. The answer names the model, summarizes fit feedback from reviews, and links two retailer pages that carry it.