How do AI assistants build answers? Search ranking picks the pages, model memory picks the brands
Search retrieval explained 32% of the variation in AI citations within a site, against 2.8% for page features such as FAQ blocks and schema. AI assistants rewrite each question into a few sub-queries, cite about half of what they retrieve, and write brands they already know into those searches.
AI assistants build an answer in four steps: decide whether to search, rewrite the question into sub-queries, filter the results, then write. Search ranking does most of the selecting. In one 158,099-answer study, ChatGPT's replayed search explained 32% of citation variation within a site, page features such as FAQ blocks 2.8% Chen · 2026. Model memory sits underneath every step, because it shapes the search queries before any page is fetched. The sections below follow the steps in order, then show where the finished answer drifts from its sources.
Whether ChatGPT searches the web depends on the prompt
ChatGPT does not search for every prompt, and the share that searches depends on the question. Graphite found evidence of search in 58% of logged-out ChatGPT responses to 100 entity-comparison prompts, each run 100 times in early August 2026 Druck · Graphite · 2026. A response counted as searched when it exposed a fan-out query or a search result. That is one vendor's test of commercial-style prompts on one engine.
Broad prompts show a lower figure. Across 66.6 million free, logged-out ChatGPT responses from August and September 2026, 14.4% contained fan-out queries Linehan · Ahrefs · 2026. Trade write-ups restated this as ChatGPT answering six in seven prompts without searching. Ahrefs corrected its own wording: the figure counts responses with visible fan-out queries, not responses that skipped search. The two studies sampled different prompt mixes, so they measure different populations rather than contradict each other.
The rate also moves within weeks. On a fixed set of entity-comparison prompts, cited logged-out responses rose from 56.6% to 88.5% between early August and late September 2026 Druck et al. · Graphite · 2026.
Query fan-out turns one question into a few searches
Query fan-out is the step in which the assistant rewrites the user's question into one or more search queries. OpenAI documents that ChatGPT Search rewrites a prompt into targeted queries and may send more specific ones after reading the first results OpenAI · 2024. Google states that AI Overviews and AI Mode run on its core Search ranking systems through retrieval-augmented generation (RAG) and query fan-out Google · 2026. RAG means pulling live web passages into the model's context before it writes.
Fan-out is smaller than its reputation. Ahrefs found that ChatGPT responses with fan-outs averaged about 1.6 queries each Linehan · Ahrefs · 2026. Peec AI counted 2.1 fan-outs per prompt on ChatGPT, 1.4 on Perplexity and 6.8 on Grok in April 2026 Rudzki · Peec AI · 2026. The data came from prompts Peec's own clients track.
Some fan-outs go straight to one website. In the Ahrefs data, 6.8% of fan-out queries used the site: operator to search a single website. Of the most common site: searches, 75.9% pointed at a brand site Linehan · Ahrefs · 2026. Factual questions about pricing, features or company details cited an official source 84.4% of the time, comparison questions 68.3%.
Ranking for the fan-out sub-queries sets the pool of citable pages
An assistant can only cite what its search returns, so ranking for the sub-queries decides the candidate pool. Across 10,000 keywords, pages ranking for both the main query and at least one fan-out sub-query collected 51% of AI Overview citations Search Engine Land · 2025. Pages ranking only for the main query collected just under 20%. The same analysis found about 68% of cited pages outside Google's top 10 for any related query. Ranking raises the odds; it does not fill every slot.
Retrieval explains far more than page design. The Chen and Koudas preprint covers 158,099 answers across ChatGPT, Gemini and Perplexity. Replayed search explained 32% of within-site citation variation, page features 2.8% Chen · 2026. FAQ blocks and schema looked positive when comparing websites and disappeared within a single site. The preprint is not peer reviewed and reports associations, not causal effects.
About half of the pages ChatGPT retrieves end up cited
Retrieval and citation are two separate filters. Across 1.4 million ChatGPT prompts, 49.98% of retrieved URLs were cited Linehan · Ahrefs · 2026. The rate depended on the channel: 88.46% for standard search results, 1.93% for Reddit and 0.51% for YouTube. Reddit pages were pulled in often and cited rarely.
Match to the question drives the second filter. Cited page titles sat closer to the user's prompt than titles retrieved but not cited. The cosine similarity was 0.602 against 0.484 in the same Ahrefs study Linehan · Ahrefs · 2026. The Chen and Koudas preprint found the same direction: titles and paragraphs that match the question are cited more often Chen · 2026. Both rely on embedding proxies, because ChatGPT does not disclose how it ranks.
Model memory shapes the search before any page is fetched
Parametric knowledge, what the model learned in training, enters the answer before retrieval starts. In 21 of 27 ChatGPT conversations, the first search query already named brands the user never typed Mohanadasan · Snippet Digital; Keyword Insights · 2026. That study used one paid account over two days in July 2026; treat its magnitudes as provisional. The larger preprint reports the same mechanism: ChatGPT writes brands it knows into its searches and names brands without any supporting page Chen · 2026.
Search visibility and memory move together. On 3,115 entity-question pairs from logged-out ChatGPT in August 2026, search visibility rose with parametric visibility in every mention-rate bin Druck et al. · Graphite · 2026. The result is correlational and comes from one vendor, with two smaller replications.
Memory changes slowly. Models released since July 2026 had knowledge cutoff gaps averaging 4.6 months for OpenAI, 2.4 for Anthropic and 4.4 for Google Druck et al. · Graphite · 2026. Graphite treats these gaps as the minimum time before new content can reach a model's weights. The averages rest on three to five models per vendor.
The finished AI answer can drift from the pages it cites
The last step, writing the answer, is where errors enter. Across 55,393 queries, 11.0% of 98,020 atomic claims in Google AI Overviews were unsupported by the cited pages, with omission the dominant failure mode arXiv · 2026. A citation shows where the assistant looked, not that every sentence matches the source.
Attribution fails more often. The Tow Center tested eight AI search engines on 1,600 queries. More than 60% of answers got a news article's source details wrong Tow Center for Digital Journalism, Columbia · 2025. Perplexity got 37% wrong and Grok 3 got 94% wrong. The task, naming a news article from an excerpt in early 2025, was narrower than a brand question and used models since replaced.
FAQFrequently Asked Questions
Sources
Sources are tiered per our methodology & sources page.
Optimizing Your Website for Generative AI Features on Google Search (Google Search Central)
Google · 2026
Google's official guide states website owners don't need llms.txt files, content chunking, AI-specific rewrites, inauthentic mentions, or extra structured data to appear in its generative AI features. Google confirms AI Overviews and AI Mode run on its core Search ranking systems via RAG and query fan-out, and frames AEO and GEO as "still SEO." Indexing and serving remain non-guaranteed.
Methodology note
Official Google Search Central documentation, published under the new "Generative AI fundamentals" section, last updated 2026-06-15. Fetched and read directly; the five-item mythbusting list and RAG/query-fan-out explanation were confirmed verbatim from the page. Represents Google's stated position on its own systems, not independent measurement.
Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact
arXiv · 2026
Google AI Overviews appeared on 13.7% of queries overall and 64.7% of question-form queries. Politically sensitive topics saw lower rates. AI Overviews cite domains more credible on average than co-displayed first-page results, but nearly 30% of cited domains do not appear in those results at all. 11% of 98,020 atomic claims were unsupported by the cited pages, with omission the dominant failure mode. Half of cited pages carry display advertising.
Methodology note
Researchers from Washington University in St. Louis issued 55,393 trending queries across 19 topical categories over a 40-day window (March 13 to April 21, 2026), measuring AI Overview activation rates, domain credibility, claim fidelity (decomposing responses into 98,020 atomic claims), and advertising on cited pages.
AI Search Has a Citation Problem (Tow Center Report)
Tow Center for Digital Journalism, Columbia · 2025
Across eight AI search engines tested, more than 60% of news-attribution queries received incorrect answers. Perplexity got 37% wrong; Grok 3 got 94% wrong. Premium paid models were no more accurate than free ones, and often produced confidently incorrect answers without flagging uncertainty. Several engines retrieved content from publishers that had explicitly blocked their crawlers.
Methodology note
1,600 queries were run across ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, Grok-2, Grok-3, and Google Gemini. The researchers selected 10 articles from each of 20 publishers, used direct excerpts as queries, and asked each chatbot to identify the headline, publisher, publication date, and URL. Responses were manually graded against six categories.
OpenAI documents that ChatGPT Search rewrites a user query into one or more targeted queries and may send additional, more specific queries after reviewing initial results. This is first-party proof that query fan-out behaviour is real inside a production chatbot search system.
Methodology note
Official OpenAI Help Center article describing how ChatGPT Search functions, including the prompt-rewriting and follow-up query behaviour. Content verified by fetch on 2026-05-27 (HTTP 200 confirmed; full body accessible in a browser session). Article documents product behaviour without disclosing the rewriting algorithm, query-volume statistics or model-side reasoning.
Parametric Knowledge Optimization (Graphite)
Graphite · Gregory Druck et al. · 2026
Graphite reports that models released since July 2026 had knowledge cutoff gaps of 4.6 months for OpenAI (n=5), 2.4 for Anthropic (n=4) and 4.4 for Google (n=3), which the authors treat as the minimum time before new content can reach a model's weights. In 3,115 entity-question pairs on logged-out ChatGPT, search visibility rose with parametric visibility in every mention-rate bin; correlational.
Methodology note
Vendor research paper by Gregory Druck, Ethan Smith and Bevin Benson of Graphite, published 16 October 2026 and read directly. Gaps use vendor documentation and model cards (30.44 days a month; Anthropic training-data cutoff); the visibility experiment is correlational, logged-out ChatGPT, collected August 2026, with two replications. The authors call the PKO tactics exploratory.
How ChatGPT Does Query Fan-Out (15 Million Queries Studied) (Ahrefs)
Ahrefs · Louise Linehan · 2026
Across 66.6 million logged-out GPT-5.6 Luna responses and 15 million fan-out queries (August to September 2026), only 14.4% of responses contained fan-out queries and 6.8% of fan-outs used the site: operator. Among the roughly 160 most-searched sites, 75.9% of site: searches targeted brand sites. Reddit appeared in about 1 in 5 US answers in March but about 1 in 43 by late September.
Methodology note
Single-vendor Ahrefs study by Louise Linehan, reviewed by Ryan Law, published 12 October 2026 and fetched directly. Uses 66.6 million free logged-out responses, Brand Radar US citations from March to September 2026 (2.4 to 3.8 million answers a month), a 20,000-prompt official-source test and about one million site: queries. The article limits findings to one free-tier model snapshot.
What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO) (arXiv 2610.15033)
Mahe Chen, Nick Koudas · 2026
Across 158,099 answers and 468,037 citations in two 90-day collections, FAQ blocks, summary boxes, authors and schema markup raised citations when comparing websites (FAQ blocks +18.1%) but not within one site (FAQ blocks -3.4%). For ChatGPT, replayed search retrieval explained 32% of within-site citation variation against 2.8% for page features. Findings are associations, not causal effects.
Methodology note
arXiv preprint 2610.15033 (cs.IR) by Mahe Chen and Nick Koudas, submitted 10 October 2026, not peer reviewed. Main dataset: 651 US fiber-internet questions on ChatGPT, Gemini and Perplexity, 24 June to 21 September 2026, plus a 55-question B2B software set; within-site comparison, API search replay and name-masking tests. Abstract and full HTML read directly; about a third of cited URLs could not be analysed.
Do Not Force Search in Prompt Tracking (Graphite)
Graphite · Gregory Druck · 2026
Across 100 entity-comparison prompts with 100 logged-out ChatGPT responses per prompt in each setting (31 July to 5 August 2026), about 58% of unforced responses showed evidence of search. Forcing search changed answers, with the largest visibility differences on prompts that never trigger search naturally. In one test, prompts with no search in 600 default responses were classed as searched in all 18 forced responses.
Methodology note
Vendor research paper by Gregory Druck, PhD, Graphite, dated 20 August 2026 and read directly. Search use was inferred from fan-out queries or a Sources panel, not citations alone. Logged-out ChatGPT only, with a 10,872-response logged-in comparison from April 2026. Rates change often; the Similarweb 7% citation figure is cited secondhand.
Patterns we see in ChatGPT query fanouts (Peec AI)
Peec AI · Tomek Rudzki · 2026
Peec AI analysed 5 million query fan-outs collected 1-21 April 2026 across ChatGPT, Perplexity and Grok. Fan-outs per prompt: Perplexity 1.4, ChatGPT 2.1, Grok 6.8. Best is the most-injected word absent from the original prompt, appearing in 24.3% of fan-outs for advice-style prompts; Grok scopes with site: in 18.3% of chats.
Methodology note
Single-vendor study on the Peec AI blog, authored by Tomek Rudzki and published 5 May 2026. Fan-outs were collected from Peec's own prompt-tracking pipeline over a 21-day window; most came from ChatGPT, so unattributed figures describe ChatGPT. Page fetched directly. Sample selection reflects Peec's tracked client prompts, not a random prompt population.
Why ChatGPT Cites One Page Over Another (Ahrefs, 1.4M Prompts)
Ahrefs · Louise Linehan, Xibeijia Guan · 2026
Across 1.4 million ChatGPT prompts, 49.98% of retrieved URLs were cited, but rates varied by channel: search 88.46%, news 12.01%, Reddit 1.93%, YouTube 0.51%. Reddit supplied 67.8% of non-cited URLs. Titles semantically closer to ChatGPT's fan-out sub-questions were cited more (cosine 0.656 versus 0.484), and natural-language URL slugs were cited 89.78% versus 81.11% of the time.
Methodology note
Single-vendor Ahrefs study by Louise Linehan with Xibeijia Guan, published 2026-04-15 and fetched directly. Results are split by ChatGPT's internal ref_type field to avoid composition artifacts, and relevance is approximated with open-source embeddings. Ahrefs retracts a first-pass metadata finding as a Reddit artifact and notes the non-cited search pool (about 3 million) is far smaller than the cited one (about 23 million).
AI Overview Fan-Out Rankings Boost Citation Odds by 161% (Surfer SEO study, 10K keywords)
Search Engine Land · 2025
Pages ranking for both the main query and at least one fan-out sub-query collected 51% of AI Overview citations. Pages ranking only for the main query collected just under 20%. Ranking for fan-out queries makes citation 161% more likely than ranking only for the head term. Around 68% of cited pages did not rank in Google's top 10 for any related query.
Methodology note
Search Engine Land coverage, December 2025, of a Surfer SEO analysis of 10,000 keywords and 33,000 fan-out queries extracted with Gemini. Surfer measured the share of AI Overview citations going to pages ranking on the head query, on fan-outs, on both, or on neither, and reported a Spearman correlation of 0.77 between fan-out coverage and citation rate.
ChatGPT Already Knows Who's In The Running Before It Searches
Snippet Digital; Keyword Insights · Suganthan Mohanadasan · 2026
In 21 of 27 ChatGPT conversations, the model's first search query already named brands the user never typed, before any page was fetched. Brands named in ChatGPT's own query reached the final answer 68.9% of the time against 2.1% for brands whose pages were fetched but never named. Only 110 of 3,554 retrieved pages were cited (3.1%), falling from 5.2% to 0.3% by position within a domain group.
Methodology note
Single-researcher study by Suganthan Mohanadasan, published 10 August 2026 and updated 17 August, read from PDF on 2026-08-24. One ChatGPT Plus account in Dubai, 24-25 July 2026: 57 conversations for citation figures and 27 for the first-query test, weighted toward software and AI tools. The author calls percentages directional and flags personalisation and tokenisation errors; the mechanism is reproducible, the magnitudes are not.
About the author Max Ackermann
Max Ackermann is founder and Managing Director of info.link, the product data platform that makes brands visible in AI search and connects every physical product to the web through GS1 Digital Link. He writes about AI search and generative engine optimization (GEO), AI-powered commerce, and how brands can structure product data for ChatGPT, Gemini, Perplexity, and retailer AI assistants like Amazon Rufus. For the past two years he has built the pipelines that put structured product data into AI answers, and run the experiments that test what actually moves AI citations.
Max has 20+ years of experience building digital products and businesses. He previously led McKinsey's Corporate Venture and Design teams across Europe, and as Managing Director of a leading US digital agency he built platforms with Nike, Google, Meta, and Airbnb. He founded the UX Design program at Central Saint Martins College, University of the Arts London, and is a Fellow of the UK's Higher Education Academy. Based in Hamburg, he works closely with GS1 on Digital Link adoption; info.link is headquartered in Hamburg and Berlin and counts GS1 Germany among its investors.
Follow Max on LinkedIn.
