Six AI Search Products: Modes, Sources, and Citations

CH.01

Compare the same task and mode before comparing products

A source link, a brand mention, and a citation to your page are different observations. A fair comparison must record the product and interface, model or mode when shown, search availability, date, language, region, and account permissions. A percentage of answers with any source cannot establish how often the same article or brand was cited.

These products can combine model responses, web search, uploaded files, and connected work data in different ways. Public documentation describes available tools and controls; it does not reveal a complete retrieval pipeline or a universal ranking formula. Do not divide whole products into fixed “retrieval,” “index,” and “memory” camps: a mode or configuration can change what information the answer uses.

When comparing answers, keep the question and test conditions consistent, save the answer and source URLs, and identify whether a cited source is public web content or a supplied document. An answer without links does not prove that it came only from training data. An answer with links does not prove that your brand was recommended or that the linked page supports every claim.

CH.02

Perplexity Search and Google AI Overviews

Perplexity offers search-oriented products and configurable source choices; AI Overviews is a feature within Google Search. Compare the interfaces your audience actually uses. Neither product documentation nor the historical figures below establish which channel will return your investment fastest.

Perplexity: record the selected mode and sources

Perplexity Pro Search documents model and source choices, including web, academic material, finance, and files. Enterprise workflows can use web content or selected internal sources. These controls make “every prompt searches the whole web” too broad. Its Agent API is a separate developer interface with configurable search tools; an API result does not establish what a consumer interface will show.

• →Put the main answer where a reader can find it, with clear headings and enough context. This is a readability recommendation, not a verified 80-word extraction rule.
• →Support important claims with relevant sources. For original research, disclose the sample, period, method, and limits; a chart does not guarantee a citation.
• →Check intended access for PerplexityBot and distinguish it from user-triggered fetching. Repeat relevant questions in the same selected mode and record actual links, without promising a fixed response time after publication.

Google AI Overviews: Search eligibility is not a citation guarantee

Google says its AI Search features use the existing Search foundations; pages must be indexed and eligible to appear with a snippet to qualify as supporting links. There is no required special AI markup or top-ten ranking prerequisite. Semrush reported AI Overviews on 6.49% of sampled keywords in January 2025 and 13.14% in March 2025. That historical keyword incidence is not a current user reach estimate or a six-product growth comparison.

• →Use Search Console and the page’s actual crawl, indexing, and snippet status to investigate eligibility. An organic ranking does not guarantee an AI Overviews citation.
• →Use relevant structured data that matches visible content. Google does not require a special schema package for AI Search; FAQ rich results stopped appearing on May 7, 2026.
• →Answer the reader’s question clearly and verify any supporting link in the actual AI Overview. People also ask and featured snippets can inform reader questions; they are not proven shortcuts to citations.

CH.03

Copilot and Gemini: separate work data, apps, and APIs

“Copilot” and “Gemini” name product families, not one fixed search pipeline. Account type, connected sources, permissions, and web access affect what can be used. Do not infer a public website’s citation prospects from a work-document answer or an API demo.

Microsoft Copilot: identify the product and permitted sources

Microsoft 365 Copilot can use permitted work data, attached files, and web information, depending on the product and controls. Administrators can manage public-web access; web search uses generated queries sent to Bing. This does not make consumer Copilot, Microsoft 365 apps, and all agents the same interface. The documentation does not establish higher purchase intent, lower competition, or a guaranteed Bing optimization return.

• →If Bing Search is relevant to your audience, inspect Bing Webmaster Tools for actual crawl or indexing problems. Treat this as a diagnostic task, not a highest-ROI hour.
• →Publish useful security, compliance, integration, or deployment documentation when readers need it. Do not label these formats as having a proven Copilot citation weight.
• →Record the tested Copilot product, work/web access, language, region, and source links. A prompt typed into Bing alone is not a complete test of Copilot’s source selection.

Gemini: app source links and API grounding are different

Gemini Apps may show links to public websites, uploaded files, or connected Workspace material; not every response includes sources. The Gemini API separately supports a Google Search tool that the model can use when needed. Neither document establishes that a Knowledge Graph entry outweighs page quality or that a QID is required for citation. AI Overviews in Search is also a different interface.

• →Check important brand identity facts and correct actual contradictions in sources you control. Accurate identity information helps readers verify the organization; it is not a guaranteed Gemini inclusion rule.
• →Use Organization and sameAs only for true, supported identity information. Do not invent relationships or fill optional fields to simulate authority.
• →Use existing independent references when they are relevant and reliable. Wikipedia or Wikidata contributions must satisfy their own eligibility and community rules; they do not guarantee an AI recommendation.

CH.04

ChatGPT and Claude: measure the source mode you used

Bain’s February 2025 report says 42% of LLM users used them for shopping recommendations. Its December 2024 survey had 1,117 respondents in total; the LLM-user subset size is not stated here. This describes reported use, not a ChatGPT-versus-Claude traffic ranking, citation rate, or effect of training memory.

ChatGPT: distinguish Search, connected sources, and API tools

ChatGPT can search the web and may show citations or related links. Availability and behavior depend on the account, interface, workspace controls, and request. The API has separately configured web-search tools and source annotations. An answer without a visible web-search citation does not reveal exactly what training material was used, and repeated third-party mentions do not prove future training inclusion.

• →Separate training permission from search access: GPTBot is for training, while OAI-SearchBot supports search. Opting out of training does not require opting out of Search.
• →Use genuine, independent references where relevant. Do not create mentions on the assumption that volume will make a future model remember or prefer your brand.
• →Keep time-sensitive facts and substantive update dates accurate. Record whether Search ran and inspect its cited pages; freshness or structured data does not reserve a citation slot.

Claude: citations are evidence to inspect, not a quality badge

Claude supports web search, subject to product availability and workspace controls. Its API offers a separate web-search tool with citations; other citation features can refer to supplied documents. These capabilities contradict a fixed “training-memory-only” classification. A Claude citation is not a certification of content quality, nor evidence that other products already cited the page.

• →Test Claude when it is relevant to your audience. Do not use an unverified citation-rate ranking to make it a final examination or the last channel to address.
• →Make research traceable: name the responsible author or organization and disclose the relevant source, method, and dates. Inspect whether the cited material supports the answer.
• →Distinguish ClaudeBot training access, Claude-SearchBot search access, and Claude-User user-triggered retrieval. Choose permissions deliberately; opening the training bot is not a prerequisite for web-search citations.

CH.05

One useful content base, with product-specific checks

Clear writing and traceable claims help people evaluate the content across interfaces. They are practical editorial choices, not a proven common ranking formula for all six products. Keep one coherent source of truth and test the interfaces that matter to your readers.

Prioritize demonstrated access, factual, and reader comprehension problems. Tenten recommends a baseline, a focused improvement, and a repeat measurement under recorded conditions. A rewrite schedule is a planning choice; it does not mean a product will cite a changed page within a set number of weeks or that a model training update will deliver a return.