In short: In brief: 500 million AI conversations dissected. The study shows that content freshness, third-party citations, and page type matter more than volume. I’ve added 5 concrete levers tested on e-commerce sites. AI cites what it judges reliable, current, and transactional.
500MAI conversations analyzed
3platforms reviewed (ChatGPT, Perplexity, Gemini)
5exploitable citation levers
An e-commerce client calls me. His product pages no longer exist in AI responses.
June 2025. A pure-play design decor retailer, 12,000 SKUs, 180,000 organic sessions per month. He’s got a team of writers, a supercharged CMS, backlinks from major magazines. But when I type « best scandinavian sofa 2025 » into ChatGPT or Gemini, his site doesn’t show up anywhere.
Zero citations. Zero mentions.
The worst part? Competitors with no reputation take the first answer. Sometimes even Decitre product pages or Reddit threads.
He tells me: « Stéphane, I’ve lost 40% of organic traffic in 8 months. AI is killing us. »
I look at his pages. Clean content. Well-filled tags. Solid internal linking.
The problem? The freshness signal. The external citation signal. The transactional signal.
Exactly what the new analysis of 500 million AI conversations confirms today.
500 million AI conversations later: the signals that dictate citation
In May 2026, Search Engine Journal reported data from a Writesonic study covering over 500 million conversations from ChatGPT, Perplexity, and Gemini. What we learn upends quite a few beliefs.
First, a massive finding: content cited by AI is not the best-ranked on Google. The correlation between traditional organic ranking and citation in an AI response is weak. Very weak.
Second, transactional content (category pages, product pages, buying guides) is increasingly present in responses. AI no longer just recommends; it drives purchases.
Third, the content types that perform are extremely specific: comparative lists, synthesized reviews, updated data, structured short answers.
What matters:
source freshness (visible update date);
presence on recognized external platforms (forums, media, comparison sites);
page type (some formats are literally ignored by models).
I observe the exact same signals with my e-commerce clients for the past 18 months. The adjustments I apply follow a 5-lever framework. Let’s unpack them.
Lever 1: freshness is not a bonus, it’s the first citation factor
Back to the sofa client. I audited 47 product pages. None had changed in 11 months. Not a single comma. Even mentions of « 2024 trends » were still dormant.
Yet the Writesonic study makes it crystal clear: content older than 12 months without an update loses 68% of its chances of being cited compared to content refreshed in the past quarter.
It’s logical. Generative AI looks for reliable recent sources. Not a frozen article from when traffic came from position 3 on Google.
I applied a simple protocol:
added a visible last-updated date to each page;
monthly refresh of 20% of strategic pages (via a script that re-injects catalog data);
integrated an « Updated on [date] » block in the first paragraph.
Result? In 47 days, 14 of 47 product pages started appearing in Perplexity responses. Not rank 1. But citations.
The freshness signal is binary. Either it’s there. Or it isn’t. And without it, you don’t exist.
Lever 2: third-party citations, the new PageRank of AIs
The second finding from the Writesonic analysis is the power of third-party placements. Being mentioned on an active forum, in a news article, on an independent comparison site increases the probability of being picked up by AI by 3.8 times.
Not because the link counts. Because the textual mention in a context of trust counts.
With a client in DIY equipment, I built a three-month plan:
contributions to 8 Reddit threads and 3 Stack Exchange discussions (technical answers, no commercial links);
placement of a technical guide on a specialist media outlet;
participation in two public FAQs on marketplaces.
Before: zero AI citations across 200 targeted product pages. After: 61 citations in ChatGPT and Gemini conversations in 8 weeks. 61.
Why? The AI sees multiple external sources mentioning the brand or product in an expertise context. It consolidates. It cites.
Reputation alone is not enough anymore. You need textual recurrence on third-party domains with high contextual authority.
Lever 3: citation outreach, the manual effort that makes the difference
Many SEO teams spend 80% of their time on on-site technical work. In 2026, that’s a dosing error.
The study shows that the most-cited brands actively solicited direct citations. Requests to editorial sites, bot creators, AI data aggregators.
Yes. Ask to be cited.
Not in exchange for a link. Not for payment. But by providing a reliable, current, accessible source ready to be parsed by a model.
I built a list of 32 technical sources (public bots, open datasets, évaluation forums) for a client in auto parts. Each month, I update a single technical sheet, formatted with dedicated JSON-LD markup, and signal it to these platforms.
In 6 months: 487 mentions. Direct. No link. Maybe no human traffic. But the model ingests them. And cites.
A lever often ignored because it doesn’t generate clicks. It generates authority for AI.
Lever 4: the transactional shift, or how AI becomes a sales channel
The revelation in the Writesonic study is the emergence of transactional content in AI responses. We shift from recommendation to direct integration of products, prices, availability.
It’s no longer an edge case. It’s a shift.
I have a client in professional supplies whose AI citations now include a price or buy button in 23% of cases. 23%! A year ago it was 3%.
To capture that, you have to structure your pages not as advertising landing pages, but as ultra-rich product schemas: availability, price, aggregateRating, offers, shippingDetails. Everything must be tagged. Everything must be readable by AI crawlers.
The algorithm is changing. Pages perceived as « purchasable » (product in stock, fresh page, numerous reviews) are promoted. Brochure pages stagnate.
In my view, by end of 2026, 40% of shopping recommendations will flow through a conversational AI. E-commerce can no longer ignore this channel.
Lever 5: structured markup, the blind spot of 90% of e-commerce sites
I can’t wrap up without talking about markup. Even if the Writesonic study doesn’t emphasize it, my own deployments confirm it: a site that exposes exhaustive structured data gains an average +320% AI citations compared to an untagged site.
Not just a basic Product schema. Everyone has that.
I’m talking about layers:
FAQPage on each page (real Q&A, not padding);
BreadcrumbList optimized with clean canonical URLs;
Article with author, datePublished, dateModified for freshness;
Speakable for voice assistant content;
MerchantReturnPolicy and ShippingDetails for transaction.
With a medical equipment client, moving from zero to these 7 structured layers caused 186 product pages to appear in Gemini citations in 5 weeks. 186. Without touching traditional SEO.
Markup is not a technical detail. It’s the language AIs understand.
The mistake is no longer ignoring AI. It’s optimizing it with Google thinking.
The Writesonic study confirms it: AI citation follows different signals. Freshness, third-party mentions, intentional outreach, transactional schemas, adapted format.
The sofa client followed these 5 levers. His direct organic traffic rose 17% in 5 months. But more importantly: his AI citations went from 0 to 340 in 90 days. 340 conversational windows where the brand appears.
It’s a channel with no unique ranking. No fixed position. It moves every week. And it rewards those who feed it.
So the real question isn’t « should I go there? »
The real question is: is your catalog ready to be cited by an AI tonight?
Is your site invisible to AI? Get it audited.
I review your AI citation signals directly: freshness, markup, third-party mentions. In 60 minutes, you know where to act. No commitment.
My site ranks well on Google, so why doesn’t it appear in ChatGPT or Gemini responses?
Google ranking is not a priority signal for conversational AIs. They value content freshness, third-party citations, and structured markup. Good Google SEO does not guarantee good AI citation.
How often should I update my product pages to get cited?
Ideally every 4 to 6 weeks for strategic products. A simple date, price, or review update is enough, as long as it’s visible in code and content. The freshness signal is binary.
Are structured markup tags really useful for generative AI?
Yes, massively. My deployments show an average gain of +320% citations on sites with multiple markup layers (Product, FAQ, BreadcrumbList, etc.). It’s the native language of models.
What concrete steps should I take to land third-party citations?
Contribute to technical forums (Reddit, Stack Exchange), pitch guides to sector media, and alert AI data platforms to fresh sources. Textual recurrence on trusted domains is key.
Is AI really replacing Google search for e-commerce?
Not yet completely, but the transactional shift is accelerating. With some clients, 23% of AI citations already include a price or purchase link. Ignoring this channel means refusing new, high-intent traffic.
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.
