Brands keep asking the same questions about AI search: what works, what doesn't, and what the data actually says. We answer them here, with original research grounded in peer-reviewed studies, enterprise benchmarks, and primary data from our own pipeline. Every claim links back to a source in our methodology.
Feed-integrated retrieval rose from 8% to about 65% of US ChatGPT Shopping recommendations after 10 July 2026, so product data now drives much of retrieval. Ranking has not died: list position still causally shifts which listings AI agents inspect, four to ten times more weakly than for humans.
Adding JSON-LD to 1,885 matched pages produced no measurable AI citation lift: +2.4% on Google AI Mode, statistically zero. Across two million citations, the apparent benefit of structured data vanished once domain was controlled. AI systems read the visible text that markup describes.
Six datasets put the AI citation half-life anywhere from under one day to longer than 17 weeks, depending on what is counted. Turnover is real, but survival flattens around a recurring core: in Google AI Mode the brand's own domain stayed cited all 17 weeks for 43% of brand queries.
Google's AI Overviews opt-out carries no documented ranking penalty, but it removes the site, not the answer: other sites fill the slot. Hiding AI Overviews entirely raised click-through 8.8 points in a randomized test. The opt-out reports no clicks and leaves AI training to Google-Extended.
Blocking GPTBot or ClaudeBot leaves ChatGPT and Claude search eligibility intact, but Google-Extended also controls Gemini Apps grounding. Across 3,115 ChatGPT entity questions, search visibility rose with training-data visibility, and no study has measured what a training-only block costs.
Across 18,012 ChatGPT citations, 44.2% fall in the first 30% of a page, a 1.47x lift over an even spread once band width is counted. A second study across seven verticals puts the most-cited band at 10-20% of the page, past the first 100 to 200 words the circulating advice names.
Across 52,170 chats on 555 prompts, average brand visibility differed 41% between the ChatGPT interface and the API. The sampling is unusually strong at 47 runs per prompt per surface. The configuration is undisclosed, and setting confounds do not average out the way random error does.
"GEO is just SEO" holds at the retrieval layer and fails at the citation layer. ChatGPT sources 83% of shopping-carousel products from Google Shopping's top organic results, yet only 2.7% of domains cited across five AI engines appear on all of them. Ranking decides who competes; the answer decides who is quoted.
Reliable AI visibility estimates need roughly 40 to 150 queries per platform depending on the engine, and typical 95% confidence intervals span 3 to 6 percentage points, so week-over-week movement is usually sampling noise. Four instruments measure AI search, and each one is blind to what the others catch.
51% of B2B software buyers now start research with an AI chatbot more often than Google, and 69% then chose a different vendor than originally planned. The shortlist forms inside the answer, before a buyer sees a vendor site, and a brand earns a slot through retrievable, consistent facts plus third-party proof.
Social and video sources are 5.54% of all AI citations across six engines, and the widely shared 94% figure is long-form's share of YouTube citations, not of AI citations overall. What gets a video cited is the text around it, its transcript, description, and chapters, not the footage or the channel's reach.
AI assistants recognise a named product 99.4% of the time in ChatGPT and 94.3% in Perplexity, but recommend it in discovery queries only 3.32% and 8.29%. The gap between recognition and recommendation is the diagnostic that predicts brand visibility in AI search.
The AI visibility category repeats statistics that do not survive scrutiny, often vendor blogs citing each other. Our public three-tier source system separates citable research from noise: peer-reviewed and enterprise-scale data at the top, single-data-point case studies treated as provisional.
No public first-party source from OpenAI, Anthropic, Google, Microsoft, or Perplexity discloses a browsable feed of real user prompts. Query fan-out and prompt rewriting are real and partly observable through the API. The best defensible workflow is a proxy built from public search demand, labelled as one.
Mention, citation, and absorption are three distinct outcomes, and only absorption causally shapes what an AI tells a customer. Most tracking tools report citation, the middle metric. ChatGPT cites fewer sources per answer than Perplexity, but each cited ChatGPT source carries more influence.
AI sends roughly 1% of web traffic but it converts at about 3x other channels, with +8% time on site and -23% bounce, and Pew finds AI summaries cut Google clicks from 15% to 8%. The shift is value per visit and click suppression, not volume. Treat AI as a quality channel.
Adding schema to 1,885 pages moved AI citations +2.4% on Google AI Mode and +2.2% on ChatGPT, both indistinguishable from zero. A 300,000-domain llms.txt analysis found the same null result. What drives citation is evidence density inside the answer, not the markup around it.
51% of B2B software buyers now start research with an AI chatbot more often than Google, and 69% then chose a different vendor than originally planned. The shortlist forms inside the answer, before a buyer sees a vendor site, and a brand earns a slot through retrievable, consistent facts plus third-party proof.
Social and video sources are 5.54% of all AI citations across six engines, and the widely shared 94% figure is long-form's share of YouTube citations, not of AI citations overall. What gets a video cited is the text around it, its transcript, description, and chapters, not the footage or the channel's reach.
AI assistants recognise a named product 99.4% of the time in ChatGPT and 94.3% in Perplexity, but recommend it in discovery queries only 3.32% and 8.29%. The gap between recognition and recommendation is the diagnostic that predicts brand visibility in AI search.
The AI visibility category repeats statistics that do not survive scrutiny, often vendor blogs citing each other. Our public three-tier source system separates citable research from noise: peer-reviewed and enterprise-scale data at the top, single-data-point case studies treated as provisional.
AI sends roughly 1% of web traffic but it converts at about 3x other channels, with +8% time on site and -23% bounce, and Pew finds AI summaries cut Google clicks from 15% to 8%. The shift is value per visit and click suppression, not volume. Treat AI as a quality channel.
Feed-integrated retrieval rose from 8% to about 65% of US ChatGPT Shopping recommendations after 10 July 2026, so product data now drives much of retrieval. Ranking has not died: list position still causally shifts which listings AI agents inspect, four to ten times more weakly than for humans.
Six datasets put the AI citation half-life anywhere from under one day to longer than 17 weeks, depending on what is counted. Turnover is real, but survival flattens around a recurring core: in Google AI Mode the brand's own domain stayed cited all 17 weeks for 43% of brand queries.
Across 52,170 chats on 555 prompts, average brand visibility differed 41% between the ChatGPT interface and the API. The sampling is unusually strong at 47 runs per prompt per surface. The configuration is undisclosed, and setting confounds do not average out the way random error does.
"GEO is just SEO" holds at the retrieval layer and fails at the citation layer. ChatGPT sources 83% of shopping-carousel products from Google Shopping's top organic results, yet only 2.7% of domains cited across five AI engines appear on all of them. Ranking decides who competes; the answer decides who is quoted.
Reliable AI visibility estimates need roughly 40 to 150 queries per platform depending on the engine, and typical 95% confidence intervals span 3 to 6 percentage points, so week-over-week movement is usually sampling noise. Four instruments measure AI search, and each one is blind to what the others catch.
No public first-party source from OpenAI, Anthropic, Google, Microsoft, or Perplexity discloses a browsable feed of real user prompts. Query fan-out and prompt rewriting are real and partly observable through the API. The best defensible workflow is a proxy built from public search demand, labelled as one.
Mention, citation, and absorption are three distinct outcomes, and only absorption causally shapes what an AI tells a customer. Most tracking tools report citation, the middle metric. ChatGPT cites fewer sources per answer than Perplexity, but each cited ChatGPT source carries more influence.
Adding JSON-LD to 1,885 matched pages produced no measurable AI citation lift: +2.4% on Google AI Mode, statistically zero. Across two million citations, the apparent benefit of structured data vanished once domain was controlled. AI systems read the visible text that markup describes.
Across 18,012 ChatGPT citations, 44.2% fall in the first 30% of a page, a 1.47x lift over an even spread once band width is counted. A second study across seven verticals puts the most-cited band at 10-20% of the page, past the first 100 to 200 words the circulating advice names.
Adding schema to 1,885 pages moved AI citations +2.4% on Google AI Mode and +2.2% on ChatGPT, both indistinguishable from zero. A 300,000-domain llms.txt analysis found the same null result. What drives citation is evidence density inside the answer, not the markup around it.
Google's AI Overviews opt-out carries no documented ranking penalty, but it removes the site, not the answer: other sites fill the slot. Hiding AI Overviews entirely raised click-through 8.8 points in a randomized test. The opt-out reports no clicks and leaves AI training to Google-Extended.
Blocking GPTBot or ClaudeBot leaves ChatGPT and Claude search eligibility intact, but Google-Extended also controls Gemini Apps grounding. Across 3,115 ChatGPT entity questions, search visibility rose with training-data visibility, and no study has measured what a training-only block costs.
The terms used across the info.link research library, defined once.
The practice of optimising content to surface in AI-generated answers from systems like ChatGPT, Perplexity, and Google AI Overviews.
Coined in a 2023 research paper from Princeton, Georgia Tech, and the Allen Institute for AI, which showed that adding citations, quotations, and statistics to a page can raise its visibility in AI answers by up to 41%. Used most widely in 2025–2026 trade and analyst coverage.
Source: Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024.
Optimising content to surface in answer engines, including AI assistants and pre-LLM answer surfaces such as featured snippets.
Used in SEO industry coverage since around 2019, originally describing optimisation for Bing’s answer features and Google’s featured snippets. Now used interchangeably with GEO in some vendor contexts, though AEO is technically broader because it covers pre-LLM answer surfaces.
A catch-all synonym for GEO and AEO, used primarily in vendor and trade contexts to describe optimisation for large language model–powered surfaces.
Emerged in 2024–2025 as the category proliferated synonyms. Functionally identical to GEO in most usage. Where GEO emphasises the generative nature of the system and AEO emphasises the answer surface, LLMO emphasises the model underneath. The label matters less than the work.
The practice of optimising content to rank in traditional search engine results pages such as Google and Bing, the “ten blue links” a user clicks through to.
The discipline GEO, AEO, and LLMO all evolved from, dating to the mid-1990s. The techniques overlap heavily, covering structured data, authoritative sourcing, and crawlable HTML, but the objective differs: SEO optimises for a ranked link the user clicks, while GEO and AEO optimise for inclusion in a synthesised answer the user may never click through. “GEO vs SEO” is now the defining comparison of the field.
The user-facing term for searches conducted through or mediated by AI assistants such as ChatGPT, Gemini, Perplexity, Claude, and Microsoft Copilot, rather than through a traditional search engine results page.
Used by McKinsey, Bain, and major analysts since 2024 to describe the consumer-side shift. McKinsey projects AI-powered search will mediate around $750 billion in US consumer revenue by 2028. Bain finds roughly 80% of consumers now rely on zero-click or AI-summarised answers for at least 40% of their search needs.
See also: AI Visibility, SEO
Source: McKinsey & Company, “New Front Door to the Internet: Winning in the Age of AI Search,” 2025.
The measurable presence of a brand, product, or piece of content in AI-generated answers, typically measured as mention rate, citation rate, or absorption rate across AI assistants.
Three distinct metrics, in increasing order of strategic value: a mention names the brand without a source link; a citation explicitly attributes information to a URL; absorption measures how much a cited source actually shaped the answer’s language and evidence. Tracking tools that report only citations measure the middle.
Definitions last reviewed: . Maintained by info.link Research.