AI Search Works: How to Prove It With Real Tests
Summarize this article with AI
The AI Search black hole no longer exists
A client calls me on a Tuesday morning.
He invested $12,000 in content designed for AI Search. Result? Nothing. Not one extra citation. Not a single attributable click. Just a budget gone.
I hear this doubt often. « Does it really work? »
The answer is yes. But not with content thrown into the void. Not without a measurement method.
I proposed a protocol to him. No grand speeches. A test.
We isolated 40 pages. We changed their structure. We waited.
14 days later, the verdict came in: +47 citations in ChatGPT. +22% organic clicks from AI Search.
Numbers. Not an opinion.
AI Search is an experimentation field, not a black hole. And you can prove it.
Pour sortir du bruit et prouver ce qui fonctionne, voici le processus que nous avons suivi avec ce client.
Le protocole de test en 5 étapes
Comment isoler l’impact de vos modifications pour l’IA Search
A test protocol, not an opinion
Split testing is the method. Like CRO, but for LLMs.
ChatGPT, Claude, Perplexity, or Gemini won’t let you isolate a control group the way Google Ads does. You can’t say « show this page to 50% of users. »
We work around it.
My method has three pillars. A golden set of prompts. About thirty real questions your customers ask—neither too broad nor too specific. We freeze them. We run them again every 48 hours.
We build a structural control group. Of the 40 test pages, 20 keep their current form. 20 get a single modification: a new FAQ section, a table of contents, finer entity markup. One variable only.
We measure page by page. No global average. We compare each URL’s citation rate.
« Adding an FAQ section shoots citations up. Removing it sends them down. That reversibility is what turns correlation into causation. » — Mark Traphagen, seoClarity, cited by Search Engine Journal
This protocol takes 45 minutes to set up. And it delivers clear verdicts in two weeks.
Google shows its cards: Search Console arrives
On June 3, Google launched a dedicated report in Search Console. AI Overviews and AI Mode, page by page.
No more sampling. No more guesswork.
Suraj Lalchandani, project lead at seoClarity, sums it well:
« That was the hardest metric to get. Everyone was sampling. Everyone was inferring. Now Google gives it to you directly. »
Concretely, you see for each URL how many times it appeared in Google’s AI responses. Impressions, clicks, position. All first-party data.
I plug this report directly into my tests. The protocol no longer relies solely on a sample of prompts. It rests on actual traffic reality.
If Search Console shows a 22% jump in AI Overview clicks on modified pages, and control pages stay flat, you have your causal link.
Clean. Irrefutable.
Les résultats parlent d’eux-mêmes : les pages modifiées ont bondi de 3 à 65 citations. Voici comment se décompose cette progression.
Du gouffre au sommet : la décomposition des gains
47 citations ChatGPT, 7 dans Perplexity, 11 dans AI Overviews
The results: +47 citations, +22% clicks
Back to the Tuesday client.
His 40 pages split as follows: 20 control pages, 20 pages with a new FAQ section structured as question/answer plus FAQPage Schema markup.
After 14 days, I reviewed the results.
- Control pages: 3 ChatGPT citations. None in Perplexity.
- Modified pages: 47 ChatGPT citations. 7 in Perplexity. 11 additional citations in Google AI Overviews.
Overall organic traffic stayed flat. But traffic from AI Search features jumped 22%.
What’s interesting? Reversibility.
We then removed the FAQs from pages that had them. In 5 days, citations dropped 41%.
Proof.
And the cost? $0 in extra ads. Just structure.
Your architecture, this invisible advantage
I see it across 1,300+ deployed semantic clusters.
Sites that win the most AI citations are those with clear semantic architecture.
Semantic architecture rests on clear entities, complete pillar pages, and precise interlinking.
Why? Because an LLM doesn’t just read visible content. It models your domain’s semantic space. The more your site follows hub-and-spoke structure, the easier it is to cite.
In my client’s test, modified pages didn’t just have an FAQ. They fit into a cluster. A pillar page « electric bike buying guide » and its 9 satellite pages. The structure created meaning. The LLM rewarded it.
A site with 800 scattered pages and no logical internal links will struggle much more to break through. AI citation values clarity.
The fatal error: comparing averages
I see many e-commerce merchants look at overall traffic. « My organic traffic dropped 5%, AI Search is useless. »
Wrong start.
Overall traffic mixes pages performing well in AI with pages regressing. The average kills the signal.
What matters is performance per page on tested segments.
Another common error: looking at impressions without clicks. A citation isn’t a visit. You can appear in 200 ChatGPT responses without generating a click if your brand doesn’t inspire trust or the snippet is unclear.
So I set up specific UTM tracking. Every link from an AI result gets a parameter ?utm_source=chatgpt&utm_medium=ai. We track the conversion.
On my client’s case, of the 47 ChatGPT citations, 18 generated a click. 18 qualified visits, every 48 hours, from a free channel. Over a year, that adds up.
What you gain by testing now
Proving AI Search works doesn’t demand a $10,000 budget.
It demands a method.
Isolate 20 to 40 pages. Pick one modification. Launch tracking with fixed prompts and Search Console. Measure page by page.
In 14 days, you have an answer.
The question to ask: how many citations do you gain by structuring your pages as hubs instead of silos?
That’s what I do with my clients. We don’t debate. We test.
When do you launch your first 20 test pages?
What if we tested your first 20 pages together?
In 45 minutes, I’ll show you which pages win (or lose) citations, and how to build a test protocol that proves AI Search’s impact on your revenue.
Book a strategic call — 45 minFrequently Asked Questions
How do I know if my site already appears in AI Search results?
Since June, Google Search Console has an « AI Overviews & AI Mode » report. It lists each URL with impressions and clicks. For other LLMs, query 30 fixed prompts every 2 days and note the cited links.
How many pages do I need for a valid test?
Take at least 20 test pages and 20 control pages, ideally within the same semantic cluster. That way, thematic variations won’t skew your results.
How do I build a ‘golden set’ of prompts?
Take 30 real questions from your Search Console or onsite search. Mix informational and commercial queries. Rerun them every 48 hours to spot variations.
Is FAQ Schema enough to boost citations?
FAQ Schema improves indexing. But what triggers Google citation is a visible, well-structured FAQ section integrated into cluster architecture. Markup alone, without visible content, delivers a weaker result.
Can I measure AI Search ROI in revenue?
Yes. Add UTM parameters to useful links in AI citations. Track conversions from tagged sessions. It’s precise, but you must control which traffic sources you assign.
