--- title: "How AI Search Changed Its Sources: A Timeline from SGE to AI Mode" description: "Most advice on getting cited by AI is a timestamp disguised as a strategy." url: "https://groas.com/post/how-ai-engines-learned-to-pick-sources-a" image: "https://groas.com/media/blog/a8a5d6855027703ee9422892c9acb0336d60c3637b8727995abe2a70a6971e85.png" published: "2026-10-05T05:50:17.083Z" modified: "2026-10-05T05:50:17.841Z" --- [Google Ads News](https://groas.com/category/google-ads-news) · October 5, 2026 · 10 min read # How AI Search Changed Its Sources: A Timeline from SGE to AI Mode [DavidFounder & CEO @ groas](https://groas.com/author/david) ![Cover image for: How AI Search Changed Its Sources: A Timeline from SGE to AI Mode](https://groas.com/media/blog/a8a5d6855027703ee9422892c9acb0336d60c3637b8727995abe2a70a6971e85.png) In this article 1. [May 2023: SGE makes existing rankings the obvious starting point](#may-2023-sge-makes-existing-rankings-the-obvious-starting-point) 2. [Late 2023 to early 2024: Perplexity puts citations in the answer](#late-2023-to-early-2024-perplexity-puts-citations-in-the-answer) 3. [May 2024: AI Overviews bring the test into live US search](#may-2024-ai-overviews-bring-the-test-into-live-us-search) 4. [October 3, 2024: Ads move into AI Overviews](#october-3-2024-ads-move-into-ai-overviews) 5. [October 31, 2024: ChatGPT Search adds another citation path](#october-31-2024-chatgpt-search-adds-another-citation-path) 6. [Early 2025: AI Mode makes one prompt many searches](#early-2025-ai-mode-makes-one-prompt-many-searches) 7. [March 2026: The old citation playbooks meet the present](#march-2026-the-old-citation-playbooks-meet-the-present) Most advice on getting cited by AI is a timestamp disguised as a strategy. One consultant says to win the featured snippet; another says to build more backlinks; a third says to fix your crawler settings. Each may be describing a real problem. They are not necessarily describing the same era of search. From SGE’s limited test in 2023 to AI Mode and today’s citation patterns, the direction is hard to miss: **engines increasingly look for useful passages and identifiable facts, not just pages that rank for a keyword**. Each shift made some of the previous playbook less dependable. Here is the timeline I would use before paying anyone to follow it. ## May 2023: SGE makes existing rankings the obvious starting point Google introduced the Search Generative Experience (SGE) at Google I/O on [May 10, 2023](https://blog.google/products/search/generative-search/). It was an opt-in Search Labs test, not a replacement for ordinary search. Its AI Snapshots sat above familiar organic results, so practitioners naturally looked first at the pages Google already surfaced for a query. The early advice followed: **rank well, answer plainly, and make the answer easy to lift**. A concise definition beneath a heading looked like a sensible extension of featured-snippet SEO. If you already had a strong page, improving its structure was a cheaper bet than inventing a separate ‘AI SEO’ campaign. The mistake was turning that sensible starting point into a permanent rule. SGE was a limited test. It showed that established search visibility mattered; it did not prove that holding a top organic position would guarantee a citation in every later AI answer. A 40-word definition box could help a reader and still be a poor substitute for a page that answers the rest of their question. **Practitioner takeaway:** keep the clear answer, but do not build a citation strategy around one formatting trick observed in an early test. ## Late 2023 to early 2024: Perplexity puts citations in the answer Perplexity made a different promise visible on the screen: answers came with inline references readers could inspect. Google’s early SGE experience placed citations in an expandable interface; Perplexity made sourcing part of the answer itself. That changed what a search marketer had to notice. A cited page was not merely a high-ranking URL. It supplied something the answer needed to say. A technical documentation page, a focused comparison table, or a relevant forum discussion could be more useful to that answer than a broad pillar page with stronger conventional SEO signals. PerplexityBot and other routes to candidate pages also meant that checking Google rankings alone would not tell you whether Perplexity could find or use your material. This did not make domain reputation irrelevant. It made **direct relevance and attributable facts** harder to ignore. If the useful answer is buried below five introductory paragraphs, another source can supply it first. I have no objection to a long page when the subject needs one. I object to asking a reader, or an answer engine, to excavate the point. **Practitioner takeaway:** inspect the passage that earns the citation, not just the authority metrics of the domain hosting it. ## May 2024: AI Overviews bring the test into live US search Google launched AI Overviews for US users on May 14, 2024. The launch quickly produced widely shared bad answers, including the pizza-glue and eat-a-rock examples. Funny screenshots, expensive lesson: generating a fluent answer and selecting sound supporting material are not the same job. The visible footprint then shrank. [BrightEdge Generative Parser data](https://searchengineland.com/google-ai-overviews-visibility-drops-15-queries-442878) reported AI presence falling from 84% of test queries in Search Labs to less than 15% in live search by early June 2024. Those figures describe different environments, so they are not a clean before-and-after measurement of one fixed set of queries. They do show why a tactic tested against Labs appearances could mislead a team buying work for live search. The launch also exposed a limit of the ‘rank first, get cited’ theory. A conventional ranking still gives a page visibility, but an AI Overview must select material it can use in an answer. A page that ranks for the broad query may say little that resolves a particular follow-up. A lower-ranking, more precise passage may do that job better. I would not tell a client to abandon SEO fundamentals on that basis. I would tell them to stop treating a top-three ranking as a citation receipt. **Practitioner takeaway:** measure the answers that actually appear in live search, then check whether your page supplies a clear claim those answers can use. ## October 3, 2024: Ads move into AI Overviews On [October 3, 2024](https://www.searchenginejournal.com/google-officially-launches-ads-in-ai-overviews/528994/), Google launched ads in AI Overviews for US mobile queries. Existing Shopping and Search ads from Performance Max and automated Search campaigns could appear in or below the generated answer. Google did not need a separate ‘AI ads’ playbook to put a paid placement near an AI response. That matters to a PPC operator because the surrounding answer changes the context of the ad. A person can read a generated recommendation, see its cited sources, and encounter a sponsored product on the same screen. Paid placement and earned mention are different things, but the reader experiences them together. Treating the ad report and the citation report as unrelated paperwork misses that interaction. I would still resist the neat story that an organic citation automatically validates the ad beside it, or that a competitor mention imposes a measurable conversion penalty. The useful question is narrower: **what does the answer say when your ad appears?** If it identifies a different product as the fit, your bid has to work in that context. If it names your business, you still need to know whether the ad adds value or merely buys space beside a mention you earned. **Practitioner takeaway:** review commercial AI answers alongside paid placements before deciding what the auction data means. ## October 31, 2024: ChatGPT Search adds another citation path OpenAI [launched ChatGPT Search on October 31, 2024](https://openai.com/index/introducing-chatgpt-search/), bringing cited web answers into a product people already used for conversational research. A brand could no longer assume that a strong Google result described its visibility across the places where a customer might ask for a recommendation. This is where technical access stopped being a detail to leave entirely to the SEO checklist. If your site blocks OAI-SearchBot in `robots.txt`, that can limit its access through that route; [crawler guidance](https://rankmath.com/kb/how-to-get-your-site-indexed-in-chatgpt-search/) is worth checking before you declare yourself invisible or safely indexed. Bing crawl errors and stale sitemaps deserve attention too. None of that means opening a bot guarantees a ChatGPT citation. It means a preventable access problem should not decide whether your useful page is considered. The writing problem is separate. A page built to rank for a short keyword may not answer a conversational request about price, fit, implementation, or alternatives. If the answer is present but scattered across vague copy, the engine has less usable material to cite. Fixing access without fixing the page is like opening the shop and leaving the shelves unlabelled. **Practitioner takeaway:** check crawl access across engines, then read your page as an answer to an actual question rather than as a container for target keywords. ## Early 2025: AI Mode makes one prompt many searches With AI Mode and the broader move toward conversational search, the key change for content planning was [_query fan-out_](https://ahrefs.com/blog/query-fan-out/). A complex prompt can be broken into related searches before the system assembles an answer. The user asks one question; the retrieval work can pursue several parts of it. ![One search query entering a central node and branching into several sub-queries](https://groas.com/media/blog/bcb62d937de0255ca3cb71c7296dbd9cba580a12c376443f5ff64382ed38b600.png) That weakens exact-match keyword targeting for a straightforward reason. A page titled ‘best enterprise CRM software’ may match the words in the first prompt yet fail to address the questions underneath it: implementation, integrations, or fit for a particular organisation. Repeating the original phrase in six headings does not answer any of those. Useful sections, explicit comparisons, and concrete attributes have a better chance of serving the different parts of the task. Where those answers sit matters as well. [CXL’s analysis of AI Overviews](https://cxl.com/blog/where-google-ai-overviews-pull-answers/) found that 55% of citations came from the first 30% of a source page. Research by Kevin Indig across 18,012 ChatGPT citations found that 44.2% came from the top 30% of documents. Those findings are not permission to delete every introduction. They are a strong reason to put the point before the ceremonial throat-clearing. Our [passage-first guide](https://groas.com/post/how-to-get-your-business-cited-in-ai-ove) goes deeper on that structure. **Practitioner takeaway:** organise a page around the questions a prompt implies, and put the most useful answer where a reader encounters it early. ## March 2026: The old citation playbooks meet the present The clearest sign that a top ranking is no longer a complete citation plan comes from an [Ahrefs study of 863,000 search results and four million AI Overview URLs](https://ahrefs.com/blog/ai-overviews-study/). By March 2026, it found that 37.9% of URLs cited in Google AI Overviews ranked in the top ten organic positions, down from roughly 76% in mid-2025. Another 36.7% of cited URLs did not rank in the top 100 organic results. I would not read those percentages as ‘rankings no longer matter’. I would read them as **rankings no longer explain citations on their own**. Across this timeline, engines have made it more useful to be the source of a specific, intelligible passage about a recognisable subject. That is a different job from placing a broad page near the top of a results list. ### Three habits I would retire - **The mandatory 40-word box beneath every heading.** A short, direct answer is useful when the question calls for one. A formulaic definition pasted everywhere does not supply the detail a more complex answer needs. - **Google-only visibility checks.** Google rankings cannot tell you whether ChatGPT Search or Perplexity can access, interpret, or cite your material. Check the relevant crawlers and look at the answers themselves. - **Exact-match headings as a complete plan.** When an engine explores related questions, repeating the starting phrase does less work than answering those questions clearly. None of this calls for deleting every existing page and starting over. It calls for editing with a more demanding test: can someone extract a specific answer, understand which product or business it concerns, and see what supports the claim? A page that says a platform is ‘the most effective solution’ gives an answer engine little to work with. A page that states its relevant attributes plainly at least gives the claim something to stand on. Off-site references can help establish context that a company cannot create merely by praising itself. ### What I would do this quarter 1. **Move the useful answer up.** Review important pages for a clear opening claim, concrete details, and comparisons where they help the reader. The top-of-page citation findings make this a sensible place to start, not a guarantee of placement. 2. **Check access beyond Google.** Inspect bot access in server logs, look for CDN blocks affecting OAI-SearchBot or PerplexityBot, and check how Bing Webmaster Tools reads your sitemaps. Fix the avoidable obstacles before commissioning more copy. 3. **Read paid and earned visibility together.** For commercial queries, note whether the AI answer names your business, a competitor, or neither when ads appear. Use that context when reviewing bids rather than treating a citation and an ad impression as reports from separate planets. This is also why I am more interested in execution than another monitoring dashboard. A tool that alerts you to a lost citation has identified work; it has not done it. The work may be a clearer passage, a corrected crawl block, or a page that finally states what the product does. [groas](https://groas.com/) is built around that paid-and-organic search work rather than asking a team to collect reports from two silos and call the hand-off a strategy. The line from SGE to AI Mode points toward a simple, uncomfortable conclusion. You cannot freeze your citation advice at the last interface you noticed. The engines keep changing how they gather an answer; your job is to make the answer worth gathering, then check whether it shows up. Last year’s formatting trick is not an operating model. ## Frequently Asked Questions ### Did ranking well on Google guarantee citations when SGE launched? No. SGE was a limited opt-in test launched in May 2023, and it showed that established search visibility mattered, not that holding a top organic position would guarantee a citation in every later AI answer. Early advice to rank well and make answers easy to lift was a sensible starting point, not a permanent rule. ### What changed when Perplexity started showing citations inline? Citations became part of the visible answer itself, so a cited page was judged on supplying something the answer needed to say, such as technical documentation, a focused comparison table, or a forum discussion. Domain reputation stayed relevant, but direct relevance and attributable facts became harder to ignore. ### What went wrong when Google launched AI Overviews in May 2024? The US launch produced widely shared bad answers, such as the pizza-glue and eat-a-rock examples, showing that generating a fluent answer and selecting sound supporting material are not the same job. BrightEdge data also reported AI presence falling from 84% of Search Labs test queries to less than 15% in live search by early June 2024, so tactics tested in Labs could mislead teams working for live search. ### How should I treat ads that appear inside AI Overviews? Review the AI answer alongside the ad, because the reader experiences both on the same screen. If the answer identifies a different product as the fit, your bid has to work in that context; if it names your business, you still need to know whether the ad adds value or merely buys space beside an earned mention. ### Why does blocking OAI-SearchBot matter for ChatGPT Search? If a site blocks OAI-SearchBot in robots.txt, that can limit access through that route, so a preventable access problem may decide whether a useful page is considered at all. Blocking a bot does not guarantee a citation either way, and the page still needs to answer a conversational question rather than hold scattered keyword-targeted copy. ### What is query fan-out and how does it affect my content? Query fan-out means a complex prompt can be broken into related searches before the system assembles an answer, so a page matching the original phrase may miss the underlying questions about implementation, integrations, or fit. Useful sections, explicit comparisons, and concrete attributes serve the parts of the task better, and putting the point early helps since studies found roughly 55% of AI Overview citations and 44.2% of ChatGPT citations came from the top 30% of a page. ### Do top rankings still get cited in AI Overviews? Less reliably than before. By March 2026, an Ahrefs study of 863,000 search results found 37.9% of URLs cited in Google AI Overviews ranked in the top ten, down from roughly 76% in mid-2025, and 36.7% of cited URLs did not rank in the top 100. Rankings still matter, but they no longer explain citations on their own. ### What should I actually do to improve my chances of being cited by AI search? Move the useful answer up by adding a clear opening claim and concrete details to important pages, check crawl access beyond Google by inspecting server logs for CDN blocks affecting OAI-SearchBot or PerplexityBot and reviewing Bing Webmaster Tools, and review commercial AI answers alongside paid placements before deciding what auction data means. ## Related Posts - [![A foggy crystal ball beside a clipboard of green-ticked tasks: the agency sells the verifiable checklist, not a prediction of what AI will say.](https://groas.com/media/blog/36d26400cbf355ec679093c89d21061bee66c8a853abb9a2bf29d3f6433cdb55.png)](https://groas.com/post/the-aeo-retainer-swipe-file-copy-ready-t) ### [The AEO Retainer Swipe File: Tiers, Scope Clauses and Client Scripts](https://groas.com/post/the-aeo-retainer-swipe-file-copy-ready-t) [October 6, 2026 • 10 min read](https://groas.com/post/the-aeo-retainer-swipe-file-copy-ready-t) [Written by Alexander Perelman](https://groas.com/post/the-aeo-retainer-swipe-file-copy-ready-t) - [![A speech-bubble-shaped bookshelf where gold 'Sponsored' gift boxes squeeze out the books, leaving two leaning sources and one fallen to the floor.](https://groas.com/media/blog/de47ce6e28adc50d217461e0f1214122bab80aded11ff6108eadf834719b9b8f.png)](https://groas.com/post/ads-are-moving-into-ai-overviews-five-pr) ### [AI Overview Ads Are Crowding the Answer: Five Citation Predictions for 2027](https://groas.com/post/ads-are-moving-into-ai-overviews-five-pr) [October 6, 2026 • 10 min read](https://groas.com/post/ads-are-moving-into-ai-overviews-five-pr) [Written by David](https://groas.com/post/ads-are-moving-into-ai-overviews-five-pr) - [![Cartoon of a businessman in a sinking rowboat proudly admiring a gold-framed chart of the rising water while his bailing bucket sits unused.](https://groas.com/media/blog/267ab77791d4bf1bfac9d34e77020fcd7536c01d8309aa711f3ad08d6e7c8c1b.png)](https://groas.com/post/introducing-citeboost-platinum-watch-you) ### [Introducing CiteBoost Platinum: Beautiful Charts of Every AI Citation You Didn’t Earn](https://groas.com/post/introducing-citeboost-platinum-watch-you) [October 6, 2026 • 6 min read](https://groas.com/post/introducing-citeboost-platinum-watch-you) [Written by David](https://groas.com/post/introducing-citeboost-platinum-watch-you) ## The Machines Already Run Search, You Should Probably Own One [get started free](https://groas.typeform.com/to/xC1bQNUT) ## Structured data ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#webpage","url":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a","name":"How AI Search Changed Its Sources: A Timeline from SGE to AI Mode","description":"Most advice on getting cited by AI is a timestamp disguised as a strategy.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://groas.com/media/blog/a8a5d6855027703ee9422892c9acb0336d60c3637b8727995abe2a70a6971e85.png"},"breadcrumb":{"@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#breadcrumb"},"dateModified":"2026-10-05T05:50:17.841Z","mainEntity":{"@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"Google Ads News","item":"https://groas.com/category/google-ads-news"},{"@type":"ListItem","position":4,"name":"How AI Search Changed Its Sources: A Timeline from SGE to AI Mode","item":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#article","url":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a","mainEntityOfPage":{"@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"How AI Search Changed Its Sources: A Timeline from SGE to AI Mode","description":"Most advice on getting cited by AI is a timestamp disguised as a strategy. One consultant says to win the featured snippet; another says to build more backlinks; a third says to fix your crawler settings. Each may be describing a real pr…","datePublished":"2026-10-05T05:50:17.083Z","dateModified":"2026-10-05T05:50:17.841Z","image":{"@type":"ImageObject","url":"https://groas.com/media/blog/a8a5d6855027703ee9422892c9acb0336d60c3637b8727995abe2a70a6971e85.png"},"author":{"@id":"https://groas.com/author/david#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"articleSection":"Google Ads News","keywords":"Google Ads, Google Ads News","wordCount":2013,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/david#person","name":"David","jobTitle":"Founder & CEO @ groas","image":"https://groas.com/media/blog/363be9ad5255816654dac065d01bf1c6063110b400bb38cc199570267232027c.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"FAQPage","@id":"https://groas.com/post/how-ai-engines-learned-to-pick-sources-a#faq","mainEntity":[{"@type":"Question","name":"Did ranking well on Google guarantee citations when SGE launched?","acceptedAnswer":{"@type":"Answer","text":"No. SGE was a limited opt-in test launched in May 2023, and it showed that established search visibility mattered, not that holding a top organic position would guarantee a citation in every later AI answer. Early advice to rank well and make answers easy to lift was a sensible starting point, not a permanent rule."}},{"@type":"Question","name":"What changed when Perplexity started showing citations inline?","acceptedAnswer":{"@type":"Answer","text":"Citations became part of the visible answer itself, so a cited page was judged on supplying something the answer needed to say, such as technical documentation, a focused comparison table, or a forum discussion. Domain reputation stayed relevant, but direct relevance and attributable facts became harder to ignore."}},{"@type":"Question","name":"What went wrong when Google launched AI Overviews in May 2024?","acceptedAnswer":{"@type":"Answer","text":"The US launch produced widely shared bad answers, such as the pizza-glue and eat-a-rock examples, showing that generating a fluent answer and selecting sound supporting material are not the same job. BrightEdge data also reported AI presence falling from 84% of Search Labs test queries to less than 15% in live search by early June 2024, so tactics tested in Labs could mislead teams working for live search."}},{"@type":"Question","name":"How should I treat ads that appear inside AI Overviews?","acceptedAnswer":{"@type":"Answer","text":"Review the AI answer alongside the ad, because the reader experiences both on the same screen. If the answer identifies a different product as the fit, your bid has to work in that context; if it names your business, you still need to know whether the ad adds value or merely buys space beside an earned mention."}},{"@type":"Question","name":"Why does blocking OAI-SearchBot matter for ChatGPT Search?","acceptedAnswer":{"@type":"Answer","text":"If a site blocks OAI-SearchBot in robots.txt, that can limit access through that route, so a preventable access problem may decide whether a useful page is considered at all. Blocking a bot does not guarantee a citation either way, and the page still needs to answer a conversational question rather than hold scattered keyword-targeted copy."}},{"@type":"Question","name":"What is query fan-out and how does it affect my content?","acceptedAnswer":{"@type":"Answer","text":"Query fan-out means a complex prompt can be broken into related searches before the system assembles an answer, so a page matching the original phrase may miss the underlying questions about implementation, integrations, or fit. Useful sections, explicit comparisons, and concrete attributes serve the parts of the task better, and putting the point early helps since studies found roughly 55% of AI Overview citations and 44.2% of ChatGPT citations came from the top 30% of a page."}},{"@type":"Question","name":"Do top rankings still get cited in AI Overviews?","acceptedAnswer":{"@type":"Answer","text":"Less reliably than before. By March 2026, an Ahrefs study of 863,000 search results found 37.9% of URLs cited in Google AI Overviews ranked in the top ten, down from roughly 76% in mid-2025, and 36.7% of cited URLs did not rank in the top 100. Rankings still matter, but they no longer explain citations on their own."}},{"@type":"Question","name":"What should I actually do to improve my chances of being cited by AI search?","acceptedAnswer":{"@type":"Answer","text":"Move the useful answer up by adding a clear opening claim and concrete details to important pages, check crawl access beyond Google by inspecting server logs for CDN blocks affecting OAI-SearchBot or PerplexityBot and reviewing Bing Webmaster Tools, and review commercial AI answers alongside paid placements before deciding what auction data means."}}]}]} ```