Yes, with a catch. Generative engine optimization is a real job: getting your pages retrieved and cited when ChatGPT, Perplexity, Gemini, Claude or Google's AI features write an answer. The claim that rewriting pages with quotations and statistics lifts AI visibility by up to 40% comes from a 2023 lab test, and it has not held up in more realistic ones.
- The 40% is lab-only. The model was handed five sources up front; the test measured quoting, not being found
- Two peer-reviewed tests disagree. C-SEO Bench (NeurIPS 2025) and SAGEO Arena (KDD 2026) found the rewrites largely ineffective once ranking is included
- Retrieval decides most. In 158,099 answers, what ChatGPT's search retrieved explained 32% of citation variation; page properties explained 2.8%
- The answers come from sellers. In our scan, 61% of the sources AI engines cited on GEO questions were companies that sell SEO, GEO or marketing services
Is generative engine optimization real?
Real problem, unproven fix. When an AI engine answers a question, it searches, picks a handful of pages, and quotes some of them. Whether your page is one of those is a real business question, and Microsoft now names the work in Bing's Webmaster Guidelines: "Generative Engine Optimization (GEO) focuses on content eligibility for grounding and reference in AI responses." The same guidelines add that "GEO does not guarantee grounding or citations in AI experiences."
Google declines the separate label. Its guide to generative AI features in Search says: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Neither company endorses the idea that a set of rewriting tricks earns citations. That idea comes from one paper, and the rest of this post is about what happened when other researchers tested it.
We sell answer engine optimization software, so we have a stake in GEO being real. That is a reason to be precise about which parts of it the evidence supports.
Where the "up to 40%" claim comes from
One experiment. Every "GEO boosts visibility by 40%" statistic traces back to "GEO: Generative Engine Optimization", by researchers at IIT Delhi and Princeton with two independent co-authors, posted in November 2023 and published at KDD 2024. It built a 10,000-query benchmark, rewrote sources nine different ways, and measured how much of each source a model used in its answer.
The setup matters more than the headline. For each query, the five top Google results were fetched and handed to a GPT-3.5 model, which wrote an answer from them. One of the five was rewritten, and the researchers measured how much of the answer drew on it, weighted by position. The best rewrite, adding quotations, raised that score by 41% over the unmodified source (the abstract rounds to "up to 40%"). Adding statistics added 31%, improving fluency 28%, and citing sources 27%. Keyword stuffing scored 8% below doing nothing. A separate check on Perplexity found quotations raised the score by 22% and statistics by 9%.
Two details rarely make it into the marketing. First, nothing was retrieved: the five sources were fixed before any rewriting, so the paper measured how a model quotes sources it already has. Second, the gains were not free. For the source already ranked first, citing sources lowered its visibility by 30.3%, quotations by 22.9% and statistics by 20.6%. The rewrites mostly moved attention from higher-ranked sources to lower-ranked ones inside a set of five.
The 40% describes a model choosing among five pages it was handed. It says nothing about whether your page gets found.
What later studies found
Shrink or reverse. Studies that added search ranking to the test found the rewriting gains shrink or reverse; studies of real AI answers found retrieval and relevance matter most. We read the abstracts and, where the numbers live in the body, the papers themselves. Peer review status is from each paper's venue listing on October 27, 2026.
C-SEO Bench: the rewrites often hurt
C-SEO Bench, accepted at NeurIPS 2025, tested published rewrite methods, including the 2023 paper's, across question answering and product recommendation on several models. Its verdict: "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." The statistics method lowered rankings in 19 of 24 settings. The authors found that "the initial ranking of documents retrieved by the search engine plays a far more dominant role," and that gains shrank as more competitors used the same methods. Their conclusion is close to Google's: traditional SEO methods are "significantly more effective," and conversational SEO "will not replace SEO, but will complement it."
SAGEO Arena: impractical in a realistic pipeline
SAGEO Arena, accepted at KDD 2026, the same conference that published the original paper, rebuilt the full pipeline a real engine runs: search, reranking, then generation. Existing approaches "remain largely impractical under realistic conditions," often degrading visibility at the retrieval and reranking stages. Structural information present in real web pages, such as markup, helped. The finding is the same as C-SEO Bench from a different angle: a rewrite that helps a model quote you can stop the search step from finding you.
Studies of real answers: retrieval, relevance, earned media
Studies that observe real engines point the same way. Chen and Koudas collected 158,099 answers with 468,037 citations and found that every page ChatGPT cited was one its own search had retrieved; which pages that search returned explained "far more of the variation than page properties (32% against 2.8%)." ChatGPT's sub-queries rarely matched Google's or Bing's results, and the findings replicated in a market of small B2B vendors. A SIGIR 2026 study of 252,000 controlled trials found "topical relevance and list position are the biggest drivers" of which source is cited first, while "formatting-only edits have little impact." A University of Toronto study found AI search favors earned, third-party media over brand-owned content far more than Google does.
The one strong pro-rewriting result, AutoGEO, learned rewrite rules from what engines prefer and beat the best 2023 method by up to 50.99%, but on benchmarks where each query comes with five candidate documents already chosen, the same fixed-set design. A July 2026 survey of 45 GEO studies sums up the state of the field: the original gains "are valid within its experimental setting but conditional on a source already being present in a fixed context," and "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability."
Why the lab result did not transfer
Four mechanisms explain the gap, and each shows up in more than one study.
- Retrieval comes first. An engine can only quote what its search returned. The 2023 test skipped that step; C-SEO Bench and Chen and Koudas found it dominates. Engines also rewrite the question into their own sub-queries before searching, a step known as query fan-out, so ranking for the words a person typed is not enough.
- Citation is competitive. In the original paper, gains for lower-ranked sources came partly at the expense of the top-ranked one. C-SEO Bench found gains fall as more sites apply the same methods, and the 2026 survey notes that "competition can erode individual gains." A tactic everyone uses stops being an edge.
- Rewrites can hurt retrieval. Text tuned to be quoted is not necessarily text that matches the search step. SAGEO Arena saw existing methods degrade retrieval and reranking, and the survey reports that "citation-oriented rewrites can impair retrieval."
- Engines push back. Bing's guidelines now file content written to trigger AI citations under "Keyword Stuffing and Artificially Engineered Language." Researchers are building defenses too: GEO Defender, a September 2026 paper, cut the average success rate of seven GEO attacks from 50.32% to 6.20% by demoting rewritten documents.
What Google and Bing say about GEO
Fundamentals first. Both say the fundamentals that get a page found are the ones that make it eligible for AI answers. Google's guide (last updated July 10, 2026) says you do not need AI text files, special schema markup, or content chopped into small chunks to appear in its AI features, and that seeking inauthentic mentions is not as helpful as it might seem. Bing's guidelines, revised in February 2026, open with "SEO Fundamentals Still Apply to Grounding and AI Experiences" and add specifics that match the research: pages "focused on a primary topic are more likely to be selected for grounding results," "early clarity improves grounding visibility and reliability," and structured data "may support clearer grounding but does not guarantee visibility."
Neither guide covers ChatGPT, Perplexity or Claude, which run their own searches. That is where GEO earns a name of its own: the job is the same, but the indexes and the candidate sets are different. Our complete guide to generative engine optimization covers each engine and ranks the tactics by evidence.
Who AI engines cite when you ask about GEO
Mostly sellers. On October 26, 2026 we asked ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews and Google AI Mode five questions about GEO: what it is, how it compares with SEO, how to do it, whether it is real, and whether it works. The 29 answers that cited sources cited 232 of them. We sorted every cited website into one of six groups.
- 142 (61%) came from companies that sell SEO, GEO, marketing or content software or services: tool vendors, agencies and consultants.
- 46 (20%) came from research, reference and education sources such as arXiv, Wikipedia and university pages.
- 17 (7%) came from independent media, 15 (6%) from community and video sites, 7 (3%) from Google's and OpenAI's own documentation, and 5 (2%) from other sites.
The engines differ sharply. ChatGPT cited the fewest sources and none from sellers. Its 13 were Search Engine Land, Wikipedia, arXiv papers, Google's and OpenAI's documentation, Princeton's page for the original paper, and one unrelated CDC page. Gemini drew 33 of its 36 citations from sellers. On the two evaluation questions ("is it real" and "does it work"), the seller share was still 56%.
The research did reach some answers, unevenly. Across all seven questions we scanned, the July 2026 survey was cited in seven answers, by ChatGPT, Perplexity and Claude. The two peer-reviewed benchmarks that contradict the 40% claim, C-SEO Bench and SAGEO Arena, were cited in none of the 342 citations across all seven of the questions we scanned. If you ask an AI engine whether GEO works, you are likely to get an answer assembled largely from pages written to sell it. This is one run on one day, and AI answers change from run to run, so read it as a snapshot, not a rate.
Ask an AI engine whether GEO works and most of its sources will be companies that sell GEO. Count us among them.
What this means for a B2B marketing team
Found, relevant, mentioned. Spend on being found, being the most relevant answer, and being mentioned elsewhere, and treat rewriting tricks as unproven. In order of evidence:
- Be retrievable on every engine. Allow the AI crawlers you want, confirm key pages are indexed in Google, and verify the site in Bing Webmaster Tools, since ChatGPT and Copilot draw on Bing's index.
- One question per page. Make each key page the best answer to one buyer question, match its title and first paragraph to how buyers ask it, and consolidate near-duplicates.
- Earn third-party citations. Original data that others cite, trade-press commentary and genuine reviews feed the earned-media bias the Toronto study measured.
- Keep the editorial basics. Named sources, dated numbers and direct answers make pages more useful. Do them for readers, not for a promised lift.
- Measure, then decide. Track citation share on a fixed set of prompts over weeks, because a single snapshot is noise.
When a vendor quotes a GEO lift, ask four questions. Were the sources already in the model's context, or did the test include search? Did the lift hold over weeks, or was it one run? Did it hold across engines? What was the comparison group? The 2023 paper would answer the first question "already in context," and most vendor case studies do not answer the others at all.
Start with whether engines can find you. The free AEO audit checks crawler access and page structure among its 29 checks.
Run the free AEO audit →What we don't know yet
- Long-run causal effects. No study we found tracks a change to real pages and its effect on citations across engines over months with a control group. The 2026 survey makes the same point.
- How much is peer reviewed. Four of the ten studies in our table are peer reviewed (the 2023 paper, C-SEO Bench, SAGEO Arena and the SIGIR 2026 study). The rest are preprints and may change.
- Engine drift. Engines change models and search partners often. A result from a 2025 model may not hold for a 2026 one.
- Our scan. One run per engine and question, on one day, sorted by our own rule. A second run would differ.
How we checked
Papers. We opened each paper's arXiv abstract page on October 26 and 27, 2026, read the body where we quote numbers that are not in the abstract, and confirmed venues on arXiv, OpenReview or the ACM DOI record. The 2023 percentages are our calculations from Table 1 of the paper's third arXiv version (the KDD 2024 text), against its unmodified baseline score of 19.3; the rank-one figures are from its Table 2.
Scan. On October 26, 2026 we sent each question once to ChatGPT (with web search), Perplexity (Sonar), Gemini (with Google Search grounding), Claude (with web search), and Google AI Overviews and AI Mode (through SerpApi), and recorded every cited URL, deduplicated within each answer and excluding Google redirect links. Five questions were on topic: "what is generative engine optimization," "generative engine optimization vs SEO," "how to do generative engine optimization," "is generative engine optimization real" and "does generative engine optimization work." Two more, "what is GEO marketing" and "best generative engine optimization tools," are counted only in the 342-citation total, because the first drew geography-marketing sources and the second asks for vendors by design.
Sorting. A cited website counts as a seller when the company behind it sells SEO, GEO, marketing or content software or services, including agencies and consultants. Research and reference covers preprint servers, publishers, encyclopedias, dictionaries, universities and course platforms. Independent media covers news and trade publications. Community and video covers Reddit, LinkedIn, YouTube, GitHub, Substack, Scribd and dev.to. AI-Advisors would count as a seller under this rule.
Frequently Asked Questions
#Is generative engine optimization real?
Yes, as a job: getting your content retrieved and cited when an AI engine writes an answer. Microsoft's Bing Webmaster Guidelines define GEO as content eligibility for grounding and reference in AI responses, and Google calls the same work SEO. What has not held up is the claim that rewriting a page with quotations and statistics reliably wins citations. Two peer-reviewed tests found those rewrites largely ineffective once search ranking is part of the test.
#Does GEO actually work?
The parts that make a page findable and relevant work: being crawled and indexed, answering the exact question, and being mentioned on other sites. A 2026 study of 158,099 AI answers found that what ChatGPT's own search retrieved explained 32% of citation variation, against 2.8% for page properties. The rewriting tactics from the 2023 paper worked in a lab where every source was already in front of the model, and have not replicated in realistic tests.
#Where does the claim that GEO boosts visibility by 40% come from?
From the 2023 paper that named GEO, published at KDD 2024. Its best method, adding quotations, raised a source's visibility score by 41% over the unmodified source; the abstract rounds this to up to 40%. The test gave the model five sources up front and measured how much of each it quoted, so it shows how a model quotes sources it already has, not whether a page gets found.
#Is GEO just SEO?
For Google, yes by Google's own account: its guide says optimizing for generative AI search is still SEO. The research agrees the foundation is shared, since being retrieved depends on crawling, indexing and relevance. The difference is that ChatGPT, Copilot, Perplexity and Claude run their own searches, often through Bing's index or their own systems, and quote only a few of the pages they retrieve.
#What gets a page cited by AI engines, according to research?
Being retrieved by the engine's search in the first place, topical relevance to the exact question, and corroboration from third-party sites have the most support. A SIGIR 2026 study of 252,000 trials found topical relevance and list position were the biggest drivers of which source gets cited first, and that formatting-only edits had little impact. A recent timestamp and explicit price information also helped.
#Should I buy GEO tools or services?
Buy measurement and help with fundamentals, and be skeptical of promised citation lifts from rewriting. Ask any vendor where its claimed lift comes from, whether the sources were already in the model's context, whether it held over time and across engines, and what the comparison group was. AI-Advisors sells AEO software, so apply the same questions to us.
#Who coined the term generative engine optimization?
Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, in the paper GEO: Generative Engine Optimization, first posted to arXiv in November 2023 and published at KDD 2024. The authors were at IIT Delhi and Princeton University, with two independent researchers.
Related Reading
- What Is Generative Engine Optimization (GEO)? The Complete Guide
- AEO vs SEO vs GEO: The 2026 Three-Layer AI Search Stack
- How to Measure AI Citation Share Across All 5 Engines: A Per-Platform Methodology
- How to Get Cited by Microsoft Copilot: A 2026 Bing-Era SEO Playbook
- Are B2B SaaS Companies Blocking AI Crawlers? September 2026 Data
