Technical SEO: The Essential Foundation to Get Cited by AI Search

2,300 sessions. 8,000 references. Zero AI citations.

A Tuesday morning, a client calls me. Industrial spare parts e-commerce. Catalog of 8,000 products. Annual SEO budget: €28,000. Result: 2,300 organic sessions per month.

He thought he was missing content. 65 blog articles written. Optimized pages. But here’s the thing. His pages weren’t appearing in Perplexity answers, ChatGPT, or Google AI Overviews.

Yet the search queries existed. « How to test a hydraulic valve » generated 1,700 monthly searches. The content was there. Radio silence. No citations.

I opened Search Console. And there, the diagnosis took 17 seconds. The indexed page rate stagnated at 42%. Crawl budget was being wasted on 14,000 useless URLs—sorting parameters, duplicate pages, facets indexed in loops. The sitemap declared 26,000 URLs, but only 8,100 were actually relevant. No schema.org on product sheets. No structured breadcrumbs. Lighthouse at 36/100. Core Web Vitals LCP exploded at 5.8 seconds.

The problem was architecture, not content.

And that’s exactly what LLMs can’t handle.

What LLMs actually look at (and no one tells you)

Why does shaky technical structure keep you out of AI answers? Let’s go back to mechanics.

Large language models like ChatGPT or Gemini don’t « know » anything. They’re probabilistic engines. They predict the most plausible next words. To anchor their answers in reality, they use RAG—Retrieval-Augmented Generation. Concretely, they query a search index, retrieve the documents deemed most relevant, then generate an answer from those sources.

That’s what Dan Taylor reminds us in his article on Search Engine Journal in April 2025: « To answer a query using RAG, an AI search engine needs a high-quality data pipeline. It needs an organized, easy-to-navigate source with authority—something only semantic HTML, logical site hierarchy, and clean indexation make possible. »

In other words, if your site is a mess of non-canonical URLs, duplicate content, 8-second load times, LLMs find nothing. Or worse, they find noise. So they pick a competitor that’s better organized.

Google says search queries hit an all-time high in 2025. But traffic to many sites keeps shrinking. Why? Google’s interface holds users longer, displaying synthetic answers—AI Overviews, rich snippets, direct citations. If your technical foundation isn’t solid, your pages won’t appear in these premium spaces. Result: net loss of visibility.

The worst part? Publishers anticipate a 50% drop in organic traffic over the next three years, per the same article. Half. Not in 10 years. By 2028.

This situation goes beyond a simple trend. It’s a structural emergency.

Voici le roadmap qui a permis de passer de 2 300 sessions à des citations ChatGPT en 5 mois.

The technical overhaul that cut crawl by 58% in 3 months

My client, the 8,000-piece industrial catalog. Here’s the roadmap we followed.

Step 1: clean up indexation. Intensive log analysis over 60 days. We identified 11,200 URLs being crawled when they should never have been. URL parameters, printable versions, infinite filter pages. We rolled out rules in robots.txt. We removed 4,500 pointless directives. We deployed strict canonical tags. Result: in 6 weeks, unnecessary URL crawls dropped 58%.

Step 2: sitemap overhaul. We went from 26,000 declared URLs to 8,100 unique URLs matching the real catalog exactly. Every URL returns a 200 status. No redirects, no errors. The sitemap became a clean, usable plan for Googlebot…and for LLM RAG systems.

Step 3: the Core Web Vitals wall. With 5.8-second LCP, no chance. We compressed images to WebP, migrated DNS, deployed measured lazy-loading. Lighthouse hit 92/100. LCP dropped to 1.6 seconds. Bounce rate fell 29%. But more importantly, the unlocked crawl budget let Google explore 4x more useful pages per day.

Step 4: structured data injection. Product schema with availability, price, brand, image. FAQ schema on 320 critical pages. ItemList schema for catégories. Entities linked themselves. Engines now understand that a « hydraulic valve » is a product, not floating text.

The result didn’t take long. In three months, the indexation rate went from 42% to 97%. Pages in positions 1-3 tripled. Technical signals finally spoke.

Les chiffres parlent d’eux-mêmes : une refonte technique ciblée transforme la visibilité dans les moteurs et les IA génératives.

47 pages cited by AI: the signal Google can’t ignore

It was around month 5 that the AI effect kicked in. No additional action taken.

I opened Search Console. Brand queries were climbing, but more importantly, new traffic sources appeared. Sessions from chat.openai.com. From perplexity.ai. From Google AI Overviews.

Analyzing logs, I found that 47 product pages had been cited in at least one LLM-generated response in the past 30 days. 47. On highly intentional queries: « high-pressure hydraulic valve industrial 400 bars, » « gear pump vs vane pump comparison. » Pages we’d cleaned, structured, made fast.

Total organic traffic jumped from 2,300 to 10,100 monthly sessions. A +340% leap. Without a single line of new content. Without a single backlink. Just the technique.

And I see this across all my clients. Those with technical foundation scores above 85/100 on Lighthouse see on average 5.2x more LLM citations than those stagnating under 60. This isn’t fuzzy correlation. It’s a pattern.

The anti-recipe: what agencies forget to plug in

I still see too many e-commerce players bet everything on « AI-optimized » content. Writing prompts to please ChatGPT. Drafting longer text. Adding customer reviews by the hundreds. Wrong move.

AI doesn’t read like humans do. It scans structure, tags, authority signals, entity consistency. A 2,000-word article with poor architecture stays invisible.

« Optimizing for search engines remains relevant, and technical SEO is the foundation of AI Search. » – Dan Taylor, Search Engine Journal, April 2025.

Classic mistakes I fix every week:

• Neglecting schema.org. Without markup, your products are anonymous text. For LLMs, that’s a signal the page isn’t reliable.
• Leaving thousands of useless URLs indexed. These pages burn crawl budget. Worse, they dilute authority.
• Confusing speed with LCP. Fast hosting isn’t enough. You need to prioritize main content rendering.
• Ignoring structural internal linking. LLMs use connections between pages to establish themes and entities. If your linking structure is weak, silos collapse.

The good news? Fixing these takes 4 to 8 weeks. And the effect on AI citations can show up the next month.

Is your site ready to talk to AI?

You have an online store. Hundreds, maybe thousands of product pages. You’ve been doing SEO for years. Yet your pages aren’t breaking through in AI answers. The cause is often technical.

Here’s a checklist, no paid tools needed:

• Open Search Console. Look at the indexation report. Valid coverage rate above 90%? If not, your useful pages aren’t being seen.
• Run a Lighthouse audit on 3 product pages. Aim for a global score > 85 and LCP under 2.5 seconds.
• Check schema.org tags with Google’s test tool. A Product snippet should display.
• Count useless indexed URLs: parameters, duplicates, empty archives. Make them disappear.

Then measure. In 90 days, check your traffic sources. If a « perplexity.ai » line appears, you’re on the right track.

Technical SEO is the foundation, not a luxury. Without it, AI Search ignores you. With it, your pages become the source LLMs choose to cite.

So for AI, is your site a wasteland or an organized library?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn