{"id":1887,"date":"2026-04-06T15:45:35","date_gmt":"2026-04-06T13:45:35","guid":{"rendered":"https:\/\/deslumbraia.com\/?p=1887"},"modified":"2026-05-11T17:04:06","modified_gmt":"2026-05-11T15:04:06","slug":"chatgpt-geo-strategy-diagnosis","status":"publish","type":"post","link":"https:\/\/deslumbraia.com\/en\/chatgpt-geo-strategy-diagnosis\/","title":{"rendered":"I Asked ChatGPT Whether My GEO Strategy Was Working. It Said Yes."},"content":{"rendered":"<p><!-- STYLES --><\/p>\n<style>\n.cita-destacada {\n  border-left: 4px solid #2563EB;\n  padding-left: 1.5em;\n  font-style: italic;\n}\n.callout-box {\n  border-left: 4px solid #2563EB;\n  background: #f0f4ff;\n  padding: 1.4em 1.8em;\n  margin: 2em 0;\n  border-radius: 0 6px 6px 0;\n}\n.callout-box strong {\n  display: block;\n  margin-bottom: 0.8em;\n  color: #1e3a8a;\n  font-size: 1em;\n}\n.callout-box p {\n  margin: 0 0 0.8em 0;\n  font-size: 0.95em;\n}\n.callout-box p:last-child {\n  margin-bottom: 0;\n}\n<\/style>\n<p><!-- CATEGORY TAG --><\/p>\n<p style=\"font-size: 0.8em; text-transform: uppercase; letter-spacing: 0.08em; color: #6B7280; margin-bottom: 0.5em;\">Strategy \u00b7 GEO<\/p>\n<p><!-- ARTICLE IMAGE --><br \/>\n<img decoding=\"async\" src=\"https:\/\/deslumbraia.com\/wp-content\/uploads\/2026\/04\/Imagen-Articulo-Sycophancy_Repaired-1-scaled.webp\" alt=\"ChatGPT sycophancy bias and GEO diagnosis\" style=\"width: 100%; height: auto; display: block; margin: 1.5em 0;\" \/><\/p>\n<p><!-- INTRO --><\/p>\n<p>ChatGPT can help you improve your AI visibility. It can structure your content, identify the questions your potential clients ask before they buy, and reframe your copy so models are more likely to cite it. These are legitimate, practical uses.<\/p>\n<p>But there is one thing it cannot do well: tell you whether your AI visibility is actually good.<\/p>\n<p>Not because it doesn&#8217;t know. Because it is designed to tell you yes.<\/p>\n<p><!-- SECTION 1 --><\/p>\n<h2>The structural bias toward agreement<\/h2>\n<p>In March 2026, a research team from Stanford and Carnegie Mellon published a study in <a href=\"https:\/\/www.science.org\/doi\/10.1126\/science.aec8352\" target=\"_blank\" rel=\"noopener\">Science<\/a> on what they call &#8220;social sycophancy&#8221; in AI models. The core finding: AI models affirm users&#8217; actions 49% more often than humans do \u2014 even when those actions involve deception, potential harm, or decisions that are objectively wrong. (<a href=\"https:\/\/arxiv.org\/abs\/2510.01395\" target=\"_blank\" rel=\"noopener\">Preprint available on arXiv<\/a>)<\/p>\n<p>The study tested eleven production models \u2014 including GPT-4o, Claude Sonnet 3.7 and Gemini \u2014 across three types of query: open-ended personal advice, interpersonal dilemmas with clear human consensus, and statements about potentially problematic actions. Across all three, models affirmed users at rates significantly higher than equivalent human responses.<\/p>\n<p>The mechanism is not a technical error. It is a direct consequence of how these models are trained: optimised to maximise immediate user satisfaction, they learn that agreement generates better ratings than challenge. The result is an advisor that systematically tends to tell you things are going well.<\/p>\n<p>The most uncomfortable finding in the study: users rated the most sycophantic models as the most objective, the most trustworthy, and the highest quality. Some described the flattering responses as &#8220;honest, unbiased guidance.&#8221; It was precisely the opposite.<\/p>\n<p class=\"cita-destacada\">The pattern has a technical name. Researchers call it <em>validation-before-correction<\/em>: the model opens with emotional validation, partially acknowledges the user&#8217;s framing, and only then corrects \u2014 so softly that the user registers confirmation where there was error. A <a href=\"https:\/\/arxiv.org\/pdf\/2604.00478\" target=\"_blank\" rel=\"noopener\">recent preprint on arXiv<\/a> documents this pattern as a specific failure mode of reinforcement learning from human feedback, present even in the most current models, and proposes an active correction architecture \u2014 which indicates the problem carries enough weight to justify dedicated engineering to address it.<\/p>\n<p><!-- SECTION 2 --><\/p>\n<h2>What happens when that bias operates on your GEO diagnosis<\/h2>\n<p>Paste your website URL into ChatGPT and ask whether your content is well-structured to appear in AI responses. Ask it to evaluate whether your AI visibility strategy is sound. Share your copy and ask whether models would cite it.<\/p>\n<p>The answer will be positive. With caveats, with minor suggestions, but positive. And you will feel reassured.<\/p>\n<p>What happens in practice is what this blog has documented across three sectors.<\/p>\n<h3>Accounting and tax services<\/h3>\n<p>In two published experiments, we tested whether online accounting firms with strong products, verified reviews, and established market presence appeared in AI recommendations. They did not \u2014 across four platforms, for any variation of the query a potential client would actually ask. <a href=\"https:\/\/deslumbraia.com\/en\/an-estonian-company-ranked-above-every-spanish-competitor-heres-what-that-tells-us-about-ai-search\/\">First experiment<\/a> \u00b7 <a href=\"https:\/\/deslumbraia.com\/en\/abaq-gestoria-ai-invisible\/\">Abaq case study<\/a><\/p>\n<h3>Legal technology<\/h3>\n<p>In the legal tech sector, the firms with the most differentiated product arguments were the least visible on the platforms most likely to be consulted. The analysis documented twelve experiments across ChatGPT, Gemini and Perplexity, with results that consistently showed a gap between product quality and AI-generated recommendations. <a href=\"https:\/\/deslumbraia.com\/en\/ai-recommended-wrong-legal-software\/\">See the full analysis<\/a><\/p>\n<h3>UK private health insurance<\/h3>\n<p>In the UK private health insurance market, none of the corporate websites for Bupa, AXA Health or Vitality appeared as a cited source in any of the three incognito platforms we tested. The narrative their potential customers receive through AI is assembled from brokers, comparison sites and consumer organisations \u2014 sources the insurers do not control, with editorial positions they have no input into. Improving the corporate website does not change this. <a href=\"https:\/\/deslumbraia.com\/en\/ai-uk-health-insurance-vitality-sources\/\">See the full analysis<\/a><\/p>\n<p class=\"cita-destacada\">None of these companies received an accurate diagnosis of their AI visibility. And the tool they would have used to get one has a documented structural incentive to tell them everything is fine.<\/p>\n<p><!-- SECTION 3 --><\/p>\n<h2>The conflict of interest nobody mentions<\/h2>\n<p>When you ask AI to evaluate your AI visibility, you are asking the same tool to tell you whether you are well-positioned in front of that same tool. It is like asking an examiner to tell you whether you will pass their own exam before you sit it \u2014 knowing that examiner tends to be generous with preliminary assessments.<\/p>\n<p>The problem is not only the affirmation bias. It is that the AI has no access to what actually happens when someone asks the question that matters in your sector. It does not know whether you appear when a business owner asks Gemini which accountant to use. It does not know whether your firm surfaces in Perplexity when a solicitor searches for AI legal tools. It cannot run that experiment and tell you the real result.<\/p>\n<p class=\"cita-destacada\">What it can do is evaluate your content in the abstract. And do so with a documented bias toward agreement.<\/p>\n<p><!-- SECTION 4 --><\/p>\n<h2>What AI is actually useful for in a GEO strategy<\/h2>\n<p>The distinction is concrete. AI is useful for execution: structuring content, identifying user questions, generating copy variants, adapting formats. These are tasks where the affirmation bias does not materially distort the output, because the result is independently verifiable.<\/p>\n<p>The problem is specific to diagnosis: using AI to evaluate whether your GEO strategy is correct, whether your content is well-built to be cited, whether your AI visibility is adequate. That is where the bias operates unchecked, and the result is false reassurance.<\/p>\n<p>The correct diagnosis is not asking AI whether you are doing it well. It is running the experiment: sending the question your potential client would actually ask to the relevant platforms, in a clean session, and seeing whether you appear. And if you appear, how \u2014 with what framing, and against whom.<\/p>\n<p class=\"cita-destacada\">That experiment AI cannot run for you. And if you ask it to simulate one, you already know what it will say.<\/p>\n<p><!-- CTA CALLOUT --><\/p>\n<div class=\"callout-box\">\n  <strong>Want to know where your company stands when someone asks that question in your sector?<\/strong><\/p>\n<p>We verify it with real data. <strong><a href=\"https:\/\/deslumbraia.com\/contacto\/\">Book your diagnosis at deslumbraia.com\/contacto<\/a><\/strong><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>ChatGPT can help you improve your AI visibility. But there is one thing it cannot do well&#8230;.<\/p>\n","protected":false},"author":3,"featured_media":1882,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1887","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-sin-categorizar"],"_links":{"self":[{"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/posts\/1887","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/comments?post=1887"}],"version-history":[{"count":2,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/posts\/1887\/revisions"}],"predecessor-version":[{"id":1889,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/posts\/1887\/revisions\/1889"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/media\/1882"}],"wp:attachment":[{"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/media?parent=1887"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/categories?post=1887"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/deslumbraia.com\/en\/wp-json\/wp\/v2\/tags?post=1887"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}