--- title: "Will optimizing content for ChatGPT hurt our Google rankings?" description: "Not if you protect pages already earning Google clicks. In one field test, a guarded rewrite lifted ChatGPT referrals with no Google loss beyond the site trend." canonical: "https://underneath.agency/resources/chatgpt-optimization-google-rankings" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # Will optimizing content for ChatGPT hurt our Google rankings? Not necessarily: the one field test published so far found no measurable Google harm when pages already earning Google clicks were left alone. The evidence is thin, comes from a single website, and does not prove the two channels never conflict. What it does show is how to run the work so you would notice if they did. ## The short version 1. In a five-month field test on one website, Google clicks to rewritten pages fell about 25% while clicks across the whole site fell about 20%, a decline the authors read as the general trend ([Watanabe and Nakayashiki](https://arxiv.org/abs/2606.04362)). 2. The same site’s ChatGPT referrals grew 5.7 times overall, but pages nobody touched grew 3.5 times, so most of the growth was ChatGPT itself, not the optimization (Watanabe and Nakayashiki). 3. ChatGPT and Google pick different pages: only 8.3% of the pages ChatGPT cited ranked in Google’s top 10 for the question ([our AI citations and Google rankings study](https://underneath.agency/research/ai-citations-google-rankings-study)). 4. In a lab test of 252,000 trials, formatting-only edits had little effect on which source AI assistants cited, while topic match, price, dates and list position mattered most ([Vishwakarma and colleagues](https://arxiv.org/abs/2605.25517)). ## Did optimizing for ChatGPT hurt Google traffic in the field test? No measurable harm was found, but the test protected the pages that mattered most to Google. [Watanabe and Nakayashiki](https://arxiv.org/abs/2606.04362) work at Glasp and studied their own site, glasp.co, so this is a self-reported case. In January 2026 they rewrote titles as questions and turned page summaries into short standalone answers across hundreds of thousands of pages. The safeguard is the important part. They called it an “SEO Guard”: any page earning meaningful Google clicks was locked and never rewritten. Pages with neither Google nor AI interest were taken down, and only the rest were edited. How to read that AI interest from your own records is covered in [what AI bots request in server logs](https://underneath.agency/resources/ai-bot-server-logs-content-demand). Google clicks to the rewritten section fell about 25% from the second half of 2025 to 2026. Site-wide, Google clicks fell about 20% over the same window, and the rewritten pages’ impressions stayed in their normal range. The authors read the decline as the general trend, not damage from the rewrite, and conclude that “AEO and SEO need not be in tension” (AEO, answer engine optimization, is their term for this work). ## Why does a ChatGPT gain look bigger than it really is? Because ChatGPT’s own growth lifts every site, so raw before-and-after numbers overstate what the optimization did. On the same site, total ChatGPT referrals grew 5.7 times on monthly figures. Pages that received no optimization at all grew 3.5 times over the same months. The rewritten pages grew 6.1 times. Comparing them with the untouched pages, the authors put the real effect at about 1.8 to 2.3 times, and call even that “suggestive” rather than proven. A stricter check could not rule out that a jump of that size happened by chance. This matters for the Google question too. If you only watch one channel, a rising ChatGPT line can hide a falling Google line, or the reverse. Compare optimized pages with similar pages you left alone, in both channels. ## Do ChatGPT and Google even reward the same pages? Mostly not, which is why work for one does not automatically help or hurt the other. In [our study of 80 US buyer questions](https://underneath.agency/research/ai-citations-google-rankings-study), only 8.3% of the pages ChatGPT cited ranked in Google’s top 10 for the question. Claude was the most Google-like assistant, at 25.7%. The gap is wide. For ChatGPT, 68.5% of cited pages were not in Google’s top 100 for the question or for any of the searches it ran behind the scenes. Google rankings describe less than a third of what ChatGPT cites. The reverse question is covered in [whether ChatGPT citations boost Google rankings](https://underneath.agency/resources/do-chatgpt-citations-boost-google-rankings). Google’s own AI surfaces differ from each other as well. On searches that showed both, [the AI Overview, the AI summary at the top of Google’s results, cited 29.7% of top-10 pages and AI Mode cited 16.3%](https://underneath.agency/research/ai-mode-vs-ai-overviews-study). So “optimizing for AI” and “ranking on Google” overlap only in part. ## Which changes help AI assistants without touching what Google rewards? Changes to substance, not layout, are where the lab evidence points. [Vishwakarma and colleagues](https://arxiv.org/abs/2605.25517), researchers at the software vendor Sprinklr, ran 252,000 trials in which six AI assistants chose between two versions of a product review. The test was simulated: each assistant saw exactly two sources, with brand names removed. Topic match, a stated price, a recent date and being listed first won citations across all six assistants. Formatting-only edits, such as breaking dense text into sections, “had no impact.” That is useful for the Google question: the edits that worked are content edits you would usually want on the page anyway. A second lab study points the same way. [Wu and colleagues](https://arxiv.org/abs/2510.11438) at Carnegie Mellon built a tool that rewrites pages to suit AI engines and reported an average improvement of 35.99% in their visibility measures. They found this did not lower the quality of the AI engines’ answers. Neither study measured Google rankings. Dates are a low-risk change worth testing. In [our study of 3,096 top-10 pages](https://underneath.agency/research/ai-overview-cited-pages-study), a machine-readable date was the one page feature linked to AI Overview citation that held up, at 7.9 points. And [pages published in the last 90 days](https://underneath.agency/research/ai-source-freshness-study) made up 17.4% to 22.6% of each AI assistant’s dated citations, against 6.9% of Google’s top 10. Keeping pages current is a change AI assistants appear to notice more than Google’s top 10 does. ## What should you do about it? Treat Google-earning pages as protected and run AI-focused changes as a measured experiment. In practice: 1. List the pages that earn meaningful Google clicks today. Exclude them from rewrites, or change them only after a small test. 2. Start AI-focused work on pages that earn little from Google: they carry the least risk. 3. Prefer substance over layout: a clear direct answer, stated prices and specifications, a visible date, and claims backed by evidence. 4. Keep a comparison group of similar pages you do not change. Judge both ChatGPT referrals and Google clicks against that group, not against last quarter. 5. Watch Google clicks and impressions for the edited pages monthly. A drop that outpaces the rest of the site is your signal to stop and review. Deciding [who owns AI search visibility](https://underneath.agency/resources/who-should-own-ai-search-visibility) makes that review someone’s job. If you want help setting up that kind of guarded program, see [our generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? The research has not tested the Google risk properly, so treat today’s answer as provisional. - The only field evidence is one website, one AI engine and about five months. The authors also say their changes were a bundle, so no single tactic can be credited or blamed. - The safeguard itself is the main reason for the result. Nobody has published what happens when pages that earn Google clicks are rewritten for AI. - The lab studies measured AI citations only. They say nothing about Google rankings. - Our studies compare which pages each system cites or ranks on given days. They do not show what happens to rankings when a page changes. ## Frequently asked questions ### Can SEO and GEO work against each other? They can in principle, but the one field test found no measurable conflict. On the site studied, Google clicks to rewritten pages fell about 25% against about 20% site-wide, which the authors treat as the general trend. ### Should we rewrite our best-ranking pages for ChatGPT? Not as a first step. The only safe result published so far came from locking pages that earned meaningful Google clicks and rewriting everything else. ### Does formatting content for AI assistants hurt SEO? The evidence does not say, but formatting alone barely moved AI citations in a lab test. In 252,000 simulated trials, structure-only edits had little effect, while topic match, prices and dates mattered. ### Why is our ChatGPT traffic growing when we have not changed anything? ChatGPT’s own growth lifts most sites. On one site, pages with no optimization at all grew 3.5 times in ChatGPT referrals over about five months. ## Sources - Watanabe and Nakayashiki (2026), [Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic](https://arxiv.org/abs/2606.04362), arXiv:2606.04362. - Vishwakarma, Kumar and Jamidar (2026), [What Gets Cited: Competitive GEO in AI Answer Engines](https://arxiv.org/abs/2605.25517), arXiv:2605.25517. - Wu, Zhong, Kim and Xiong (2025), [What Generative Search Engines Like and How to Optimize Web Content Cooperatively](https://arxiv.org/abs/2510.11438), arXiv:2510.11438. - Underneath (2026), [Do ChatGPT, Gemini, Perplexity and Claude cite pages that rank?](https://underneath.agency/research/ai-citations-google-rankings-study) - Underneath (2026), [AI Mode vs AI Overviews: how different are the sources?](https://underneath.agency/research/ai-mode-vs-ai-overviews-study) - Underneath (2026), [What pages cited by AI Overviews have in common: 3,096 pages](https://underneath.agency/research/ai-overview-cited-pages-study) - Underneath (2026), [How fresh are the pages AI engines cite?](https://underneath.agency/research/ai-source-freshness-study) --- This is the Markdown twin of https://underneath.agency/resources/chatgpt-optimization-google-rankings. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.