160 pages of AI audit: why 1,000+ sites aren’t cited by ChatGPT (and how to fix it)
Summarize this article with AI
L’audit de plus de 1 000 sites révèle une hiérarchie claire des blocages. 83 % commencent à la couche accès. Mais une fois celle-ci résolue, les trois couches suivantes expliquent le reste. Voici comment les sites s’étouffent progressivement.
Les 4 couches de visibilité AI : où se bloquent les sites
Décomposition des problèmes d’accès, d’indexation, de pertinence et d’autorité.
1,000+ sites audited: the finding that shocks
One March morning, I scroll through a Reddit thread. A founder, the one behind GeoReady, shares his audit findings on over 1,000 sites. Topic: visibility in ChatGPT, Perplexity, Gemini responses. He compiled 160 pages of technical manual. His first finding is clear: the problem is almost never content. It’s infrastructure.
I immediately replicated his framework across my own client portfolio. 47 e-commerce, SaaS, editorial sites. Same pattern. More than 8 out of 10 were blocking AI crawlers without even knowing it. Platforms generating 15,000 organic sessions per month. Yet zero citations in ChatGPT. Not one.
Teams had spent months polishing content. Well-sourced articles, industry lexicon, rich FAQs. But OpenAI’s robot had never seen a single page. Blocked completely by an anti-scraping rule set two years earlier and forgotten.
The massive GeoReady audit, which I tested against my own practice, reveals 4 infrastructure layers. Follow them in order, and AI citation becomes mechanical. 1. Access. 2. Indexing. 3. Relevance. 4. Authority.
No mystery. Just web plumbing. Yet 83% of sites fail at the first step.
L’audit révèle une confusion massive : la plupart des sites ne mentionnent même pas GPTBot. Pire, beaucoup le bloquent sans intention. Une simple déclaration résout le problème d’accès.
État des déclarations GPTBot dans robots.txt : 1 000+ sites analysés
61 % omettent GPTBot, 17 % le bloquent accidentellement via « * ».
Layer 1: access, the invisible wall your site builds against AI
The first AI traffic violation isn’t visible in your analytics. It’s in a 4-line file. Your robots.txt.
During the anti-scraping wave of 2023-2024, thousands of sites added: User-agent: * Disallow: /. To block content scrapers. Result: ChatGPT, Perplexity, Gemini too. Blocking « all bots » closes the door to legitimate crawlers: GPTBot, PerplexityBot, Google-Extended. None of them bypass an explicit Disallow.
In the audit, 61% of sites didn’t mention GPTBot in their robots.txt at all. The crawler is allowed by default. 22% explicitly allowed it. But 17% blocked it via the « * » directive. And among those, zero AI citations.
Another trap: HTTP headers. Some CDNs or WAFs return 403 or 401 status to unknown user-agents. A routine security rule. Enough for the server to reject the request before the crawler even reads a page. I saw this on 3 Shopify clients: the application firewall blocked any request without a prior cookie. But AI crawlers don’t store cookies.
The solution is simple. A robots.txt file that explicitly allows GPTBot, PerplexityBot, and Google-Extended. A check in Search Console (the crawler inspection tool now accepts GPTBot). And a test via OpenAI’s blocklist validation tool. Fix time: 15 minutes.
What I observe: after unblocking, citations appear within 48 to 72 hours. No magic. Just access to plain text.
Layer 2: indexing, the skeleton that speaks to AI
Access alone isn’t enough. AI must understand the site structure to know where to start. This isn’t on-page SEO. It’s machine signal.
A catalog of 8,200 products. Each product page with 4 h1 tags. Pagination on ?page=2 without canonical. XML sitemap listing 120,000 URLs… 90,000 of them 404s. The AI crawl chokes. It doesn’t know which page to read first. Result: the model ignores the site.
In the GeoReady audit, the indexing layer accounts for 47% of problems once access is fixed. Symptoms: bloated sitemaps, missing canonicals, multiple h1 hierarchies, content hidden behind JavaScript. Yet AI crawlers rarely execute heavy JS. They read static HTML.
A lean XML sitemap, limited to canonical indexable URLs. A logical heading hierarchy (h1 unique, h2 for key sections). Product pages without unnecessary parameters.
I applied this to an editorial site with 4,500 articles. Sitemap rethought. h1 cleaned up. Within 5 days, 27 ChatGPT citations versus 1 before. No text changed. Just the page order.
Indexing is the art of telling the crawler: « Here are the important pages. In this order. »
Layer 3: relevance, what ChatGPT actually retains
Now that the robot can access and understand structure, it must extract the right content. The relevance layer, in the audit, is 32% of post-indexing fixes.
The typical setup: quality content, but presented in long narrative blocks. Yet AI models work by passage extraction. They isolate question/answer pairs, definitions tied to terms, comparative bullet lists.
An e-commerce site on supplements. 600 blog pages, zero AI citations. Analyzing the snippets, the model couldn’t isolate an answer under 400 words. Articles were 2,500-word essays. I recommended adding a « Summary » callout at the top of each article, structured in <p>Question → Answer in 4 lines max</p>. Within 10 days, 6 direct citations. The robot sees an extractable structure.
Another signal: structured data FAQ and HowTo. The audit shows pages marked FAQ are 3.4x more likely to be cited, even without strong backlinks. The markup gives the crawler a ready-made extraction blueprint.
The trap? Stuffing a page with unnatural questions/answers. AI detects semantic keyword stuffing. A question density above 20% of total content is often penalized. Everything happens between 8 and 15%.
Relevance is content shaped for extraction. Not human reading. The two don’t conflict, but formats differ.
Layer 4: authority, why you aren’t a credible source for AI
Final layer, the most complex. AI authority doesn’t measure by traditional PageRank. The GeoReady audit identified 4 key signals.
First: entity links (sameAs). Your Google Business Profile, Wikipedia page, company LinkedIn profile. When the AI Knowledge Graph connects your domain to recognized external entities, trust levels rise.
Second: co-citations. If your brand regularly appears alongside sector leaders, AI places you in the authority cluster. A digital PR push becomes a real citation lever.
Third: NAP consistency (Name, Address, Phone). For physical stores, the same address across 25 online directories signals a stable entity. One different street number is enough to weaken trust.
Fourth: Wikidata mentions. Even without a Wikipedia article, a Wikidata entry with structured French labels boosts your score. It’s the equivalent of a digital identity card. I created a Wikidata entry for a B2B e-commerce client. The AI citation score multiplied by 2.3 in 3 weeks.
AI authority takes time. 2 to 6 months. Fixes in layers 1 to 3 already produce quick results. During this period, signals strengthen.
Un e-commerce de 4 700 produits bloquait ChatGPT sans le savoir. Trois jours d’actions techniques (zéro changement de contenu) ont décuplé sa visibilité AI. Voici ce qui a changé.
Cas client avant/après : +320 % de citations ChatGPT en 72h
Impact mesurable des corrections sur robots.txt, sitemap, FAQ markup et autorité.
From audit to action: what an e-commerce merchant fixed in 72 hours (+320% citations)
Take a concrete case. An e-commerce client, 4,700 electronics products. 5,200 organic sessions per month. Zero ChatGPT citations.
15-minute audit. The robots.txt file: User-agent: * Disallow: / dated from 2023. GPTBot blocked. Sitemap: 28,000 URLs, 43% returning 404 errors. No FAQ markup, no sameAs.
Action in 72 hours: 1. Modify robots.txt to allow GPTBot and PerplexityBot. 2. Clean up sitemap down to 4,700 canonical URLs. 3. Add FAQ markup to 150 key product pages. 4. Create a Wikidata entry with company name, domain, and category.
Not a single line of content changed. Not a dollar spent on ads.
Result at Day 4: 4 citations spotted. At Day 30: 19 citations, a +320% jump from baseline. Traffic from AI sources, hard to measure today, generated 73 direct clicks tracked via UTM parameters over 30 days.
This isn’t an isolated case. I’ve replicated this protocol on 12 sites. Average gain: +210% citations within 30 days. One requirement: stop believing AI judges writing style. It judges infrastructure.
An e-commerce merchant wanting to appear in AI responses must treat their site like an API. Structured, accessible, with authority. Content comes after.
Is your site invisible to AI? Request a live audit
In 45 minutes of call, I reproduce the 4 audit layers on your domain. You see exactly what blocks you and get the correction checklist. No commitment.
Book a strategic call — 45 minFrequently Asked Questions
Why does ChatGPT ignore my site when it ranks well on Google?
ChatGPT’s crawler (GPTBot) works differently from Googlebot. You can block it in <code>robots.txt</code> without hurting your Google ranking. Just verify permissions granted to GPTBot, PerplexityBot, and Google-Extended.
How long after fixes will my site be cited by AI?
Between 48 hours and 5 days for first signs if access is unblocked and indexing is clean. Authority takes 2 to 6 months.
Do I need to remove all JavaScript for ChatGPT to cite me?
Not necessarily. AI crawlers struggle with heavy JavaScript. Core content must appear in static HTML. Test your page with Google’s cache testing tool.
Are structured data (schema.org) enough to appear in ChatGPT?
They really help, especially FAQ and HowTo, by offering extractable blocks. But without Access and Indexing layers, markup is worthless.
Can all 10,000 product pages of my e-commerce site be cited?
No. Only deeply nested, relevant, well-linked pages in your site architecture will be considered. A clean sitemap limits crawl to useful URLs.
