Open ChatGPT. Type the question a customer would type before buying what you sell. Not your brand name. The problem your product solves, in the words a stranger would use.
Press enter.
Two or three products come back. Maybe four.
Was one of them yours?
If not, understand what just happened. That shopper did not scroll past you. They did not see you on page two. You were not part of the conversation at all, and there is no page two in a conversation.
The Number That Should Worry Every E-commerce Team
Adobe Digital Insights tracked a 4,700 percent year-on-year increase in AI-referred traffic to retail sites. That is not a typo, and it is not the interesting part.
The interesting part is what happened to that traffic when it arrived. Adobe Analytics found AI-referred shoppers converting 42 percent better than traditional organic visitors in the first quarter of 2026, up from roughly 31 percent in studies the year before.
Sit with that inversion for a moment. AI referrals used to convert badly, because early AI answers sent curious browsers rather than buyers. Now they convert better than any other organic channel, because the AI has already done the comparison shopping. The person arriving on your product page has read the reviews, weighed the alternatives, and decided.
They just decided somewhere you could not see, in a conversation you were not part of.
AI referral conversion, against traditional organic
Adobe Analytics. The shopper arrives having already chosen. Being in the answer is the entire game.
Why Your Product Page Cannot Save You
The instinctive response is to improve the product page. Better copy. More detail. Richer descriptions of quality and craftsmanship.
It will not work, and the reason is structural rather than editorial.
AI models are built to discount self-asserted claims. Your page says your moisturizer is excellent for oily skin. Of course it does. Every moisturizer page says something similar, and the model knows that the incentive behind that sentence is identical across every brand in the category.
So it looks elsewhere for evidence.
A September 2025 study from the University of Toronto found generative engines display a systematic bias toward earned media, third-party editorial, independent publications, over brand-owned content. Muck Rack's analysis put the figure higher still: roughly 82 percent of links cited by AI engines originate outside the brand's own website.
Read that again if you have been investing in content marketing.
Eighty-two percent of the citations that decide whether you appear in an AI recommendation come from places you do not control.
How AI weighs two kinds of evidence
Your product page
One source. One obvious incentive. However well written, the model reads it as a claim from a party with a reason to make it.
A forum thread about your product
Six independent users. Specific experiences. No incentive to lie. The model reads it as evidence rather than assertion.
Same product, entirely different trust weight. Authenticity signals, not polish, drive citations.
The Reddit Story Everyone Gets Half Right
Here is where most GEO advice becomes actively dangerous, and I want to be precise about it.
You will read that Reddit presence multiplies your AI citations. You will also read that Reddit collapsed and no longer matters. Both claims circulate confidently. Both are half true, and the missing half is what protects you.
In September 2025, OpenAI quietly adjusted its retrieval weights. Semrush tracked Reddit's citation share inside ChatGPT falling from roughly 60 percent to 10 percent over about six weeks. No announcement. No warning. Brands that had built their entire AI visibility strategy on Reddit watched it evaporate.
The following month, Reddit sued Perplexity in federal court in Manhattan over unauthorised scraping.
Then it partly recovered. Tinuiti's first-quarter 2026 data found Reddit citation share growing at least 73 percent between October 2025 and January 2026.
So which is it?
Neither, and the answer depends entirely on which engine is answering.
Reddit citation share, by engine
Tinuiti Q1 2026. The same content strategy produces radically different results depending on which engine your customer opened.
Perplexity draws roughly a quarter of its citations from Reddit. Gemini cites it in one response in a thousand. ChatGPT's weighting has swung by fifty percentage points in six months.
The lesson is not "invest in Reddit" or "abandon Reddit." It is that any strategy anchored to a single platform carries concentration risk that can reprice overnight, without notice, because a model provider adjusted a weight in a configuration file.
That happened. It will happen again.
What Structured Data Actually Does Now
Schema markup used to be a rich-snippet tactic. You added Product schema, you got a star rating in Google results, click-through improved by perhaps thirty percent.
That is no longer what it is for.
Structured data is now the dataset AI engines read to decide what to recommend. Research indicates 65 percent of pages cited by Google AI Mode and 71 percent of pages cited by ChatGPT carry advanced structured data. It is not a formatting preference. It is the machine-readable substrate of your entire catalog.
Which makes the most common implementation error genuinely expensive.
The JavaScript trap
Most modern e-commerce platforms inject schema after page load, through JavaScript. It renders correctly in a browser. It validates in testing tools. Everything looks right.
AI crawlers read the server-side HTML head before JavaScript executes.
Identical schema, one of them invisible
window.addEventListener('load', function() {
var s = document.createElement('script');
s.type = 'application/ld+json';
s.text = JSON.stringify(productData);
document.head.appendChild(s);
});
Result: the crawler reads the head, finds nothing, and moves on. Your price and stock data may as well not exist.
{ "@context": "https://schema.org",
"@graph": [
{ "@type": "Product", "@id": "#prod" },
{ "@type": "Offer", "availability": "InStock" },
{ "@type": "AggregateRating" }
] }
Same information. Same schema types. One reaches the AI intact, the other never existed.
This single misconfiguration silently disqualifies thousands of Indonesian product pages from AI recommendation, and nothing in Search Console will ever tell you.
Nesting matters more than presence
Isolated schema blocks are weaker than a unified graph. Product, Offer, Review and AggregateRating nested inside a single @graph configuration lets the engine understand that these facts describe one entity rather than four unrelated fragments.
Adding FAQPage schema on top of that correlates with a 44 percent increase in AI search citations, and FAQ blocks on buying guides have shown a 2.7 times citation lift.
Where in the Page the Citation Comes From
SparkToro's 2026 analysis found 44.2 percent of all LLM citations extracted from the first 30 percent of a page. The final 30 percent contributes only 24.7 percent.
Models behave like impatient editors. They read the top, take what answers the question, and stop.
So the answer-first structure is not a stylistic choice. A concise, declarative summary in the first forty to sixty words, directly answering the query, without hedging, without a paragraph of brand narrative preceding it.
Comparison tables extract at an 82 percent frequency, translating to roughly a 2.34 times impact on citation rates. Pages maintaining 120 to 180 words between headings receive 70 percent more citations than pages with sections under 50 words, because that density gives the model enough context to extract a coherent thought.
None of this makes the page worse for humans. It mostly makes it better.
Indonesia Breaks the Western Playbook in Three Places
Global GEO advice assumes a market that does not exist here.
Perplexity punches seven times above its global weight
Perplexity holds roughly 11 to 21 percent of Indonesia's AI chatbot share, against a global average nearer 2 to 3 percent.
The reason is specific and knowable. In May 2025, Telkomsel bundled Perplexity Pro subscriptions into its data plans, putting the tool in millions of hands overnight.
A GEO strategy copied from a Western agency would underweight Perplexity badly, and Perplexity is precisely the engine that draws a quarter of its citations from Reddit. The strategic implications compound.
Machine translation is penalised
Generative models detect unnatural phrasing and down-rank it during citation selection. Running your English product page through a translation layer produces content that reads as machine-generated, because it is.
Natively written Bahasa Indonesia, paired with English through hreflang, is the requirement. This also positions you for Sahabat-AI, the 70-billion parameter model developed by Indosat Ooredoo Hutchison and GoTo, trained explicitly on Bahasa Indonesia, Javanese and Sundanese.
Marketplaces have their own AI layer
Lazada runs AI Lazzie, an OpenAI-powered shopping assistant launched in late 2024. Shopee runs Choki. Both read product listings and review text to generate recommendations inside the apps where Indonesians already shop.
Optimizing only for ChatGPT misses the assistant your customer meets while browsing.
It No Longer Stops at a Recommendation
This is the part most brands have not internalised, and it arrives faster than the last shift did.
The Agentic Commerce Protocol, governed by OpenAI and Stripe, lets AI agents browse products, select variants, verify pricing, and complete checkout inside the chat interface, on the shopper's behalf. Google and Microsoft developed the Universal Commerce Protocol toward the same end.
The purchase completes without the shopper visiting your website. Without seeing your design. Without reading your carefully written brand story.
The agent reads machine-readable attributes: price, availability, shipping terms, return policy. It ignores everything a human would notice.
If your catalog feed is not connected, the transaction happens with a competitor whose feed is. Not because your product is worse. Because the agent could not read yours.
What the agent needs, and what it discards
Server-side schema
Product, Offer, AggregateRating in the HTML head. Not injected. Not deferred.
Live pricing and stock
Real-time accuracy. Nightly batch updates make your data look stale, and stale data gets deprioritised.
Crawler access
GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot permitted in robots.txt. Blocking them is invisible and fatal.
Specific reviews
"Lasts twelve hours on oily skin" is quotable. "Great product, fast shipping" gives the model nothing.
Website design carries zero weight. The agent cannot see it.
Where to Actually Begin
Not with content. Not with Reddit.
Check robots.txt first. If GPTBot is disallowed, nothing else you do matters, and this is the single most common reason a product is invisible to ChatGPT.
Then check whether your schema renders server-side. View source, not inspect element. If the JSON-LD only appears after JavaScript runs, you have found the problem.
Then look at your reviews. Not the average rating, the language. Do they describe who the product suits and in what circumstances, or do they say "recommended, five stars"? The first is an asset for AI citation. The second is noise.
Only after those three does off-site work make sense. And when it does, it should span earned media, video, review aggregators and communities rather than concentrating anywhere, because September 2025 demonstrated exactly what concentration costs.
All of this depends on traditional ecommerce SEO underneath it. AI engines rely on search crawlers to reach your catalog in the first place. A store Google cannot index properly will not be cited well by ChatGPT either, however immaculate the structured data.
The full discipline sits inside Generative Engine Optimization for e-commerce, and the broader framework across every AI surface inside our GEO practice.
I wrote Cited or Silent specifically about this problem, because the gap between brands that understood the citation mechanics and brands that were still writing longer product descriptions was widening every quarter, quietly, in a channel nobody was measuring. The companion volume, Found Before They Search, covers the search foundations everything else rests on. Both are free as gated editions, and both are published in paperback and hardcover and listed on Google Play Books and Apple Books.
What Silence Costs
Nothing measurable, at first.
Your Google rankings hold. Your marketplace listings perform. The quarterly report looks fine.
Meanwhile a shopper somewhere asks an AI what to buy, receives two product names, and buys one of them. That transaction never appears in your analytics as a loss, because it never touched your site. There is no impression to compare against, no click-through rate to decline, no red flag anywhere.
The channel converting 42 percent better than your best organic traffic is running, right now, and you are either in the answer or you are not.
There is no page two in a conversation.
Frequently Asked Questions
What is the difference between SEO and GEO for an online store?
Traditional SEO ranks your product and category pages in a results list so shoppers can click through and compare. GEO structures your schema, content and off-site presence so a conversational AI selects, quotes and recommends your product directly, often without any click. SEO is a ranking game across many positions. GEO is closer to binary: you are named, or you are absent.
Does GEO replace our existing SEO work?
No, they compound. AI engines rely on search crawlers to access and index your catalog before anything else can happen. A store Google cannot index properly will not get cited well by ChatGPT either, however good the structured data. Technical SEO is the foundation GEO sits on.
Why does AI cite Reddit and forums instead of our product page?
Because generative models are built to discount self-asserted brand claims. A University of Toronto study found systematic bias toward earned media over brand-owned content, and Muck Rack's analysis put roughly 82 percent of AI-cited links outside the brand's own website. Six independent forum users describing specific experiences read as evidence. Your product page reads as a claim from a party with an incentive.
Should we build a Reddit presence to improve AI visibility?
Carefully, and never exclusively. In September 2025 OpenAI adjusted retrieval weights and Reddit's ChatGPT citation share fell from roughly 60 percent to 10 percent within six weeks, without announcement. It later partly recovered, growing at least 73 percent between October 2025 and January 2026 according to Tinuiti. Perplexity still draws around 24 percent of its citations from Reddit while Gemini cites it in 0.1 percent of responses. Any strategy anchored to one platform can reprice overnight.
Does JavaScript-injected schema work for AI crawlers?
No. AI crawlers read the server-side HTML head before JavaScript executes. Schema injected client-side after page load renders correctly in a browser and validates in testing tools, but the crawler finds nothing. This single misconfiguration silently disqualifies thousands of product pages, and nothing in Search Console reports it.
How long until GEO work produces results?
Technical changes such as server-side schema deployment and robots.txt corrections can shift citations within four to eight weeks, since AI crawlers re-index considerably faster than traditional search. Building genuine third-party consensus takes longer, typically three to six months of consistent presence before it meaningfully affects citation frequency.
What is the Agentic Commerce Protocol?
An open-source standard governed by OpenAI and Stripe that lets AI agents browse products, select variants, verify pricing and complete checkout inside a chat conversation on a shopper's behalf. Google and Microsoft developed the Universal Commerce Protocol toward the same purpose. The transaction completes without the shopper visiting your website, which means the agent reads machine-readable attributes and ignores design entirely.
Do Indonesian shoppers actually use AI for product research?
Indonesia-specific polling on pre-purchase AI use remains thin, and global benchmarks serve as directional proxies rather than verified local figures. What is measurable: ChatGPT dominates Indonesian AI chatbot traffic, and Perplexity holds roughly 11 to 21 percent locally against a global average nearer 2 to 3 percent, largely because Telkomsel bundled Perplexity Pro into data plans in May 2025.
Does machine-translating our content into Bahasa Indonesia work?
No, and it can hurt. Generative models detect unnatural phrasing and down-rank machine-translated content during citation selection. Natively written Bahasa Indonesia paired with English through hreflang is required, which also positions the brand for Sahabat-AI, the 70-billion parameter model from Indosat and GoTo trained on Bahasa Indonesia, Javanese and Sundanese.
Sources & References:
- Adobe Digital Insights - AI-referred traffic growth to retail sites, year on year.
- Adobe Analytics - AI referral conversion premium, Q1 2026, against 2025 baseline studies.
- University of Toronto - study on generative engine bias toward earned media over brand-owned content, September 2025.
- Muck Rack - Generative Pulse report, proportion of AI-cited links originating outside brand websites.
- Semrush - citation tracking of Reddit's share within ChatGPT following the September 2025 retrieval weight adjustment.
- Tinuiti - Q1 2026 AI Citations Trends Report, Reddit recovery and per-engine citation share.
- SparkToro - 2026 analysis of citation extraction by page position.
- Arfadia Digital Indonesia - AI Citation Rate Report 2026. arfadia.com/resources
- Arfadia Digital Indonesia - Digital Marketing Benchmark Indonesia 2026. arfadia.com/resources