Where Does ChatGPT Get Its Information? The 2026 Citation Stack Explained

Gist Answers is an AI-powered audience engagement and content discovery tool that embeds directly on your site. It allows readers to ask questions and receive clear, well-sourced answers grounded in your content. When your coverage does not address a question, Gist Answers supplements responses with information from the largest licensed content library in generative AI. Every answer includes clear citations and ProRata’s proportional attribution, so readers can see exactly which publishers contributed. Publishers maintain full control of their content. It is never used to train AI models and can be removed from the network at any time.

Early integrations show that readers who interact with Gist Answers have:

• 3x longer sessions
• 2.5 - 3x more page views
• a significantly higher likelihood of returning

Instead of leaving your site to search elsewhere, readers can ask questions directly and receive well-sourced answers grounded in your content. Customizable entry points which include Suggested Questions guide readers deeper into your coverage, turning a single visit into a longer session.

Gist Answers is free to implement and generates revenue through three distinct paths:

• Sponsored Questions: High-intent questions generated by Gist Ads are shared on a 50/50 basis with you.
• Generative Ads: Gist Ads appearing within answers generate revenue that is shared 50/50 between ProRata and the publishers whose content contributed to the response, distributed proportionally based on contribution.
• Partner Inventory: You retain 100% of the revenue from any additional advertising inventory on the answer page.

Additionally, there are no development costs or AI token costs required to run Gist Answers.

A standard citation tells your audience where information came from. ProRata's proportional attribution goes further by accurately and repeatably measuring how much each source actually shaped the answer. Every Gist Answers response reflects the relative contribution of each piece of content that informed it. This ensures that you receive fair credit and compensation whenever and wherever your content  is used to generate a response in Gist Answers. Publishers whose content contributes more to an answer receive a proportionally larger share of the ad revenue that the answer generates. By linking usage to value, proportional attribution makes fair and scalable revenue sharing possible.

No. Gist Answers uses your content only for answer generation in response to user questions. We do not train an LLM using your content. This means that, if you ever wish to end our partnership, your content can be fully removed from our systems. Your intellectual property remains yours.

Our priority is to create an answer for the user that is relevant and satisfies their curiosity. When doing this, we look at multiple signals of relevance, including timeliness, geography, and more. We look first to your content in creating the answer, and then, if needed, will supplement with additional content from the licensed Gist library. But, if neither you nor our library have the content for a great answer, we tell the user so. We won't hallucinate a response just for the sake of responding.

Integration requires minimal engineering lift. ProRata provides an embeddable widget that drops into your existing property with no custom development required. The Gist Console lets your team customize the look, feel, and behavior without touching code. Most partners are live within days.

The Gist Console provides detailed visibility into both sides of the network. You can see how many questions your widget has answered, how your content is performing across partner integrations, and insights into the actual queries being asked - revealing what topics matter most to your readers and where gaps in your coverage exist. All data is exportable. You can also use the Console to tune your integration, adjusting how the widget appears on your site.

Gist Answers is available to publishers who meet our editorial quality standards and comply with applicable laws. Our content policies prioritize accuracy, original reporting, and responsible publishing practices.

Get started in just three steps:

1. Sign a partnership agreement. Request one through our contact form, or if you're a member of the News/Media Alliance, FIPP, or the DPCMO, opt in through our umbrella agreements with these organizations.
2. Configure and integrate your widget. Gist provides an embeddable widget and full documentation to get your integration set up and customized to your brand.
3. Go live. Most partners are live within days and can begin seeing audience engagement from day one.

‍

Gist GEO is a one-of-a-kind tool that analyzes the way AI interprets your brand, product, or content and offers ways of making sure AI uses it in the best possible way.

GEO means structuring your content, distribution, and website so AI models can easily find you, understand you, and accurately cite you in their answers, similar to how SEO helps Google find and rank web pages.​

SEO ranks you and decides where you show up in search results. AEO determines how you show up in AI answers. GEO helps you show up in AI answers correctly and favorably.

Zero-click refers to the way current search engines are structured. AI answers typically appear above ranked search results, meaning that users no longer have to click through to another site to get the answer they need (this is also known as the zero-click crisis).

There are nine baseline metrics Gist GEO uses to measure your brand’s AI visibility:

• Share of Citations: How often you’re actually quoted
• Citation Rate: How often you’re part of the research
• Earned Media Score: When third-party sources mention you
• Share of Recommendations: How often you’re actively suggested
• Share of Voice: How often you’re mentioned compared to competitors
• Placement: How early in conversations your brand appears
• Sentiment: How positively or negatively your brand is mentioned
• Average Ranking in Lists: Where your brand is ranked on average
• Share of Found Links: How often your links are used to form an answer but don’t make it into the answer itself

‍

Having a comprehensive tool to measure and understand AI visibility gives your brand the ability to be understood more by AI systems. Your brand shows up in AI answers more correctly, more favorably, and at the right time.

Gist GEO updates Brand Health scores on a weekly cadence. Every metric shows direction, delta from last period, and trend line.

Native: Ads match the look and feel of the AI interface


Contextual: Ads are matched in real time to the topic and content of the AI-generated answer

Large language models (LLMs) analyze AI-generated output in real time to determine which ads are relevant, with controls available to adjust how closely ads must match the context.

Gist Answers is a customizable AI search engine for websites, tuned to a site’s own content and enhanced with a licensed library of trusted publications.

It helps keep users on your website by offering interactive AI experiences, which can improve content discovery, increase time spent on the site, and boost engagement.

A conversational AI system that plans, compares, and recommends travel options, replacing multi-site research with a single interface. Implementations include ChatGPT with Expedia and Booking.com, Google AI Mode with Canvas, Perplexity with SelfBook, Kayak AI Mode, and Trip.com TripGenie.

Gist GEO turns insights from your data into Opportunities each week. Each Opportunity includes the insight, a step-by-step list of recommended actions, and tools to assign priority and owners. You can return weekly to see how your reports change as you implement, and you'll receive fresh Opportunities every week.

Enterprise plans cover all current and emerging engines.The starter plan covers ChatGPT. The Growth plan adds Claude, Perplexity, and Google AI overviews.

‍

GEO measures visibility inside AI-generated answers, not traditional search rankings. Your SEO work feeds into GEO performance, but they measure different things. You can rank #1 on Google and still be invisible in ChatGPT.

Gist GEO offers both monthly and annual billing, with a discount on the annual option. There is no charge during your trial period. Once the trial ends, you will be required to subscribe to continue using the product. You can choose monthly or annual billing at that point. Your selected cadence will be charged immediately and renew accordingly.

A dedicated support team is available by email seven days a week. An in-product support agent and a dedicated wiki are accessible directly in the product.

Brand Health scores update on a weekly cadence. Every metric shows direction (improving or declining), the delta from the prior period, and a trend line over time.

Reports use interactive charts and data visualization, and can be broken down by query, timeframe, or specific LLMs. Reports and underlying data export as .pdf, .jpg, .png, or .md files.

Clinical institutions lead. In a Gist GEO study of 8 publishers and more than 20,000 AI answers across four engines, Mayo Clinic was cited on 31.6% of health answers, Cleveland Clinic on 18.7%, and Harvard Health on 10.7%, ahead of every large commercial health publisher, which landed in the single digits for Share of citations.

Not always. AI engines frequently read a publisher's page to build an answer without linking it, a pattern tracked as Share of found links. In this study that “credit gap” was substantial and concentrated on a single engine, meaning value was extracted from publisher content with no citation and no referral traffic.

No. A mention names your brand in the answer text (tracked as Share of voice); a citation links your page as a source (tracked as Share of citations). Only the citation drives referral traffic and authority, so for publishers the citation is the metric that matters.

It varies sharply by engine, which is why blended scores mislead. In this study, Google AI Overviews and Perplexity cited publishers far more often than one chat engine did. One publisher, for example, earned 13.4% of AI Overviews answers but 1.4% on the stingiest engine. Track each engine separately.

Measure per engine and per topic, repeatedly, over time. Gist GEO samples real questions across ChatGPT, Perplexity, Google AI Overviews, and Claude, then reports nine metrics, including Share of citations and Share of found links, and prioritizes the fixes. A single spot-check is noise; a repeated mean is signal.

This study ran on ChatGPT, Perplexity, and Google AI Overviews. Gist GEO also tracks Claude, and a Gist GEO audit measures your brand across all four engines — the study's three-engine scope was a research design choice, not a tool limit.

Almost no one. Backlinko's analysis of 4 million search results found 0.63% of clicks go to page-2 results, and the #1 result earns roughly ten times the clicks of position ten. Google removed continuous scroll in June 2024 after finding that auto-loaded results didn't improve searcher satisfaction. For practical purposes, anything below the first screen of results is invisible.

Effectively, yes. The demand page 2 theoretically served is now absorbed higher up the page by AI-influenced SERP features: AI Overviews, People Also Ask, and generative product summaries answer the query before the reader reaches the lower half of page 1, and Google AI Mode replaces the results page entirely. The contest moved from which page you rank on to whether the AI answer cites you.

Yes, measurably. Pew Research Center found searchers click a traditional result on 8% of visits when an AI summary appears, versus 15% without one. Seer Interactive found brands not cited in the AI Overview saw organic CTR on those queries fall 67% over 2025, while cited brands earned 120% more organic clicks per impression. Citation, rather than mere presence, decides who keeps their traffic.

No, but its scoreboard moved. Ranking still matters because AI answers draw heavily from pages that rank; ranking alone is simply no longer sufficient. The durable strategy layers generative engine optimization, meaning extractable answers, entity authority, and citation measurement, on top of the SEO you already do. We wrote a whole post on this: Is SEO Dead?

Measure it directly. Run your priority queries through Gist GEO, which runs each query repeatedly across ChatGPT, Claude, Perplexity, and Google AI Overviews and scores nine baseline metrics, including Share of Citations and Citation Rate. The audit shows where you're cited, where you're mentioned without a link, and which competitors hold the answers your buyers see.

It is a chat-style tab inside Google Search, powered by Gemini. Ask a question and it writes a full answer with linked sources, instead of showing ranked results. It launched in March 2025, expanded worldwide through late 2025, and passed 1 billion monthly users by May 2026, per Google. Think of it as Google answering first and linking second.

AI Overviews summarize on top of the normal results page; AI Mode replaces the results page with a conversation. AI Mode's answers run several times longer, and the two surfaces share only a small fraction of their citations, so winning one gives no guarantee of winning the other. They are best treated as separate channels with separate scoreboards.

Query fan-out is AI Mode's retrieval technique: it splits your question into subtopics, runs many searches at once on your behalf, and writes one answer from everything those searches return. For marketers, it means your page can be cited for prompts you never targeted, so coverage of adjacent questions matters as much as the head term.

There are no positions to hold, only citations to earn inside the answer. You improve your odds the same way you win AI Overviews: rank retrievably, structure pages for extraction, and build brand authority on sources AI trusts. Google's guidance calls the work "still SEO." Track whether it is paying off by measuring Share of Citations and Share of Voice in Gist GEO.

Yes. Google confirmed ad tests in AI Mode in May 2025, and by mid-2026 text ads were appearing on roughly three in ten commercial AI Mode queries, most of them carrying two ads at once. Higher-value topics see ads most often. The organic citation next to those ads remains the placement no budget can buy.

Because shopping research now resolves inside the answer. AI Overviews are growing fastest on commercial queries, and when they appear on shopping searches they cite mostly video, community, and editorial sources rather than retailer pages. The research visit your product and category pages used to capture happens upstream, inside the answer itself.

Work all three layers. Keep Merchant Center feeds complete and current, since Google names feeds as a path into AI responses. Maintain Product structured data for rich-result eligibility. And earn citations where AI looks: YouTube, Reddit, review sites, editorial. Then verify with a Gist GEO audit, which shows whether AI answers cite your pages or merely mention your brand.

Yes, and now at higher rates than other channels. Adobe Analytics found AI-referred shoppers converted 31% more than other traffic during holiday 2025, with revenue per AI-referred visit up 254% year over year. That marks a reversal from mid-2025, when AI-referred visitors still converted worse. The AI-referred shopper arrives late in the journey, already persuaded by the answers they read.

Agentic commerce is AI completing the purchase on the shopper's behalf. Google's agentic checkout buys on a merchant's site when a tracked price drops, OpenAI's Instant Checkout completes purchases inside ChatGPT via the Agentic Commerce Protocol, and Perplexity offers checkout powered by PayPal. For merchants it means revenue without a session, which makes machine-readable product data the storefront.

The mechanics rhyme; the scoreboard differs. Publishers optimize mostly for citations that send readers. Merchants also compete for recommendations, meaning the products AI names on the shortlist. That is why commerce programs watch Share of Recommendations, Average Ranking in Lists, and Sentiment alongside Share of Citations in Gist GEO. The feed and entity work is retail-specific; the measurement discipline is universal.

Audit how AI engines mention your brand today, then map the queries where competitors earn the recommendation and you do not. Add your 10 highest-intent queries to Gist GEO, leading with Deep Analysis, note your baseline date, and build your gap list from the Opportunities tab. Every action after that follows the five-step plan above.

Measure your share-of-recommendations delta against a dated baseline, week over week. Supporting signals are sentiment movement on tracked queries and new AI citations earned by pages you shipped. If share of recommendations climbs across 6 consecutive weeks of Gist GEO reports, the strategy works; if a page shows zero movement after 6 weeks, rework it.

Gist GEO tells you exactly what to do next: it audits your AI mentions, ranks your content gaps by impact and effort, and hands you a prioritized opportunity list with the specific action for each. Run one report and you leave with a numbered plan you can assign the same day.

Complete the five actions in the What to Do This Week section above: pick your 10 target queries, add them to Gist GEO with Deep Analysis first, note your baseline date, pull your found-but-uncited page list, and book a recurring weekly review. Total time is under 3 hours, and it produces the baseline every later step measures against.

Yes, and increasingly on the commercial ones. BrightEdge measured AI Overviews on about 48% of tracked queries by early 2026, and Semrush found Computers and Electronics among the fastest-growing AIO categories, up 107.62%. For eCommerce queries, only 13.4% of AIO-cited sources also rank in the organic top 10, so ranking alone does not guarantee presence in the answer.

Late spring, then refresh through the season. NRF found 62% of shoppers had started by early July 2026, and Ahrefs found AI assistants cite established content, averaging around three years old and about 26% fresher than organic results. Content published in August competes for answers that were largely assembled from pages already live and cited in June.

Mostly for deals, comparisons, and list-building. PwC found 73% of families plan to use AI in back-to-school shopping, with 23% hunting deals. Tinuiti found 68% of parents using conversational AI: 33% for best deals, 32% to compare products, 28% to summarize reviews. Deloitte found the AI-using parent plans to spend $737 per child, about a third above the $557 average.

Google AI Mode is a conversational search surface, separate from AI Overviews, where shoppers research inside an ongoing AI session. For shopping it draws on the Shopping Graph's more than 60 billion product listings and adds virtual try-on, price tracking with alerts, and agentic checkout that can complete a purchase on the shopper's behalf. Retail product data reaches it through feeds, not product pages.

Build a seasonal query set from your category's real buying prompts, run it weekly across engines, and track whether AI recommends and cites you. In Gist GEO, load the prompts as Queries and watch Share of Recommendations, Average Ranking in Lists, Share of Citations, and Sentiment; the weekly refresh shows whether your season prep is moving the numbers.

An AI visibility audit is a structured measurement of how AI engines describe, cite, and recommend your brand. It runs a representative query set across engines such as ChatGPT, Claude, Perplexity, and Google AI Overviews, then scores mentions, citations, recommendations, and sentiment against competitors. A Gist GEO audit also converts each finding into a prioritized action plan.

Weekly, at minimum for the queries that matter most. SISTRIX found that ChatGPT Search swaps 74% of the domains it cites for a given prompt from one week to the next, and Google AI Mode swaps 56%, so monthly or quarterly snapshots miss most of the movement. Gist GEO refreshes its nine baseline metrics weekly so the trend, not a lucky snapshot, tells the story.

Because generative engines rebuild answers on each run, retrieval sources shift, and models update. SparkToro measured less than a 1-in-100 chance that ChatGPT or Google's AI returns the same brand list twice for the same prompt. That volatility is normal; the signal lives in your visibility rate across many runs, which is what repeated measurement captures.

Closed-loop intelligence is an operating cycle where measurement leads to action and every action is re-measured. Applied to AI visibility: audit what engines say, prioritize fixes by impact and effort, ship the work, then run the same queries again to confirm the change worked. The loop closes when the re-measurement feeds the next round of priorities.

Each finding becomes an Opportunity card with an Impact label, an Effort level, and a status you track from To Do to Completed. Cards carry two tabs, Insights and Actions, covering content work and earned-media or digital-PR outreach with named target sites. When you mark an Opportunity done, Gist GEO watches that query's metric on the next run.

Entity SEO is the practice of establishing your brand, people, and products as uniquely identifiable entities that search and AI engines recognize independent of wording. It spans a canonical entity home on your site, consistent corroborating profiles across the web, and structured data that asserts the relationships between them, so engines can confidently mention, recommend, and cite you.

A mention is your brand appearing in the text of an AI answer without a link. A citation is a linked source the engine used to construct the answer. Engines can mention you without citing you, and cite you without naming you, which is why Gist GEO tracks Share of Voice and Share of Citations as separate metrics.

Start by checking whether Google already recognizes you via the Knowledge Graph Search API. Then build corroboration: earned coverage in authoritative publications, Organization schema with sameAs links on your entity home, a Wikidata entry, and a Wikipedia article where notability supports one. Consistency matters most; every profile should state the same facts about the same entity.

Yes, though differently than assumed. Its share of visible citations varies widely by engine, and Semrush found it as low as roughly 2% of responses in Google AI Mode. But nearly all major language models train on Wikipedia, and it remains a core entity registry that grounding systems consult, so a presence there still strengthens how engines resolve your brand.

No. Structured data is one implementation layer of entity SEO: schema markup asserts your entity's facts in machine-readable form. Entity SEO is the broader architecture, choosing the entity home, reconciling profiles across the web, earning the corroboration engines trust, and monitoring how they describe you. Our guide to Structured Data for AI Search covers the markup layer in depth.

No. GEO adds a citation-focused layer on top of the platform strategy you already run. What works on each platform still works. What changes is that you now have a second thing you are also optimizing for, and a second way of measuring whether content did its job.

Two reasons. Google and Reddit have a data licensing agreement, and Reddit's own scale means hundreds of millions of weekly users generating billions of posts across active communities. Citation share moves month to month, so treat any single figure as a snapshot rather than a stable number.

Rarely. As of BrightEdge's September 2025 measurement, YouTube took 29.5% of citations in Google AI Overviews against 0.2% in ChatGPT. Video is a Google-surface play right now. If your audience is asking ChatGPT, the citations you care about are coming from somewhere else, which is why the answer depends on the engine as much as the platform.

A mention names your brand inside an AI answer. A citation links your content as the source of it. That distinction applies to a YouTube video or a Reddit thread the same way it applies to a blog post, with one wrinkle: the cited page is not always yours. A citation of your video sends someone to your channel. A citation of a thread about you sends them to a community you do not control.

No. AI Mode is its own answer surface rather than a feature inside a results page, and citation patterns differ between the two. As of September 2025 BrightEdge had YouTube taking 29.5% of citations in AI Overviews and 16.6% in AI Mode. Google's own answer surface covers how it works.

Ask each engine the questions your buyers actually ask, not just your brand name, and repeat on a schedule. Cover direct description, pricing, product status, leadership, and comparisons against competitors. Because engines cite different sources and rotate them week to week, a single check proves very little. Gist GEO runs this continuously across major AI engines.

Yes, but expect no reply. Use the thumbs-down control beneath the overview, then "Report a problem," and pick the category that fits. Google says the feedback helps improve AI Overviews but publishes no service commitment, case tracking or correction process. Treat it as one input, and fix the underlying source in parallel, because that is what actually changes the answer.

Longer than you would like, and it varies by engine. Google's own documentation says recrawling a page can take anywhere from several days to several months. Even after the source is corrected, engines rotate which sources they cite week to week, so the old claim can resurface before it disappears for good. Plan on repeated verification rather than a single confirmation.

Not directly. Wikipedia strongly discourages paid advocates from editing articles about their own organization, and the Wikimedia Terms of Use require disclosing your employer, client and affiliation. The compliant route is to disclose, then propose the change on the article's talk page using the request edit template, citing a reliable independent source. Uncontroversial fixes such as clear vandalism or spelling may be handled directly.

It has been tried, with limited success so far. US courts have ruled for AI providers in early defamation cases, in part crediting the accuracy disclaimers shown alongside outputs, while European regulators have rejected that reasoning under data protection law. One documented case saw false claims still appearing months into active litigation. Treat legal action as a backstop for real commercial harm rather than a correction workflow.

It retrieves sources that discuss the category, including roundups, comparison pages, review sites and community threads, then synthesizes a recommendation from what those sources say rather than evaluating the products itself. On some engines more than a dozen sources go into a single answer. That is why the composition of third-party coverage about your category shapes the recommendation more than your own product page does.

Publish it for the buyers who reach your site, and do not expect it to carry your AI visibility. In professional services, third-party listicles received 80.9% of listicle citations against 19.1% for self-promotional lists. You also have to avoid presenting your own list as independent, because the FTC's Consumer Reviews and Testimonials Rule prohibits misrepresenting that an entity provides independent reviews.

A little, and far less than commonly claimed. G2's own analysis of 30,000 AI citations across 500 software categories found that categories with 10% more reviews saw roughly 2% more citations, and that review volume explained under 2% of the variance in AI visibility. Reviews are worth maintaining, and they clear inclusion thresholds like G2's 10-review Grid minimum, but they are not the lever.

Comparison-shaped content. Across 75,000 AI answers, listicles took 40.9% of citations on commercial-intent queries, while informational questions skewed toward standard articles. An April 2025 study of 768,000 citations found that product-related content, a bucket including best-of articles, vendor comparisons and product pages, made up over 70% of citations on decision-stage queries, with blogs at 3 to 6%.

Frequently enough that presence needs monitoring rather than a one-time check. One measurement of listicle citations found ChatGPT's dropped about 30% in a single month between December 2025 and January 2026. Engines also differ sharply in what they read, with top-100 citation overlap between them ranging from 16% to 59%, so a placement that works on one engine may be invisible on another.