A scan of 59 Chinese cross-border brand websites found that none block OAI-SearchBot, 67.4% serve an llms.txt file, and only 1.9% carry FAQPage schema. But when we opened the llms.txt files, four out of five contained no brand information — they were Shopify platform templates about payment protocols. The door is open, but the room is empty. Raw data and the scanning script are included below.
The first article in this series scanned 100 AI tool websites and found something counterintuitive: almost nobody blocks AI search crawlers, but the industry is checking the wrong one.
That scan explicitly noted a limit: "Whether the same holds for a DTC brand or a manufacturing exporter is a separate question this sample cannot answer."
This article answers that question. Instead of AI companies, we scanned 59 Chinese-origin brands that sell internationally through their own websites — from Anker and SHEIN to mid-tier outdoor gear and e-bike makers. These are the brands that would benefit most from showing up when someone asks ChatGPT "what's a good portable power station for camping?"
The results did not match what I expected, and the real finding was not in any of the metrics I planned to check.
The sample
59 Chinese cross-border brand domains across 18 product categories: consumer electronics, smart home, portable power, DTC fashion, beauty, outdoor gear, LED displays, 3D printers, digital tablets, e-bikes, projectors, smart pet devices, audio, rugged phones, and wearables.
About a third are well-known names: Anker, SHEIN, Roborock, EcoFlow, Govee. The rest are mid-tier brands — Tribit, Blackdeer, Oukitel, Fiido, Engwe — the kind that would benefit most from AI visibility but are unlikely to have a dedicated AI-SEO person.
Of the 59, 53 returned their homepage (HTTP 200). 2 returned 403, and 4 failed to connect (DNS timeout or abort). Every metric below uses only the sites where that specific signal was conclusively readable — not 59, not 53, but the per-metric denominator noted in each table.
The same scanning script as the first article: zero LLM calls, failures recorded as errors, no defaults masquerading as data.
Finding 1: Nobody blocks AI search crawlers here either
| Crawler | AI tools (n=68) | Export brands (n=47) |
|---|---|---|
| OAI-SearchBot (ChatGPT search) | 0% | 0% |
| ChatGPT-User (live fetch) | 0% | 0% |
| GPTBot (training) | 5.9% | 2.1% |
| ClaudeBot | 5.9% | 2.1% |
| PerplexityBot | 1.5% | 0% |
| Google-Extended | 5.9% | 2.1% |
| CCBot | 7.4% | 2.1% |
| Bytespider | 5.9% | 2.1% |
The blocking rate is even lower than AI companies. Only one brand — Elegoo, a 3D printer maker — blocks GPTBot, and it leaves OAI-SearchBot open. That is the same deliberate, informed pattern from the first scan: opt out of training data, stay in the search index.
For anyone still worried about "accidental crawler blocking" as a major risk: in a combined sample of 159 sites across two very different populations, zero block OAI-SearchBot. Zero block ChatGPT-User. The door is open.
The question, then, is what is behind that door.
Finding 2: 67% serve llms.txt. Most of them are empty.
Of the 43 sites where the llms.txt check was conclusive, 29 served a file — 67.4%.
That is nearly double the 38.8% found in AI-native tools. My initial reaction was that Chinese export brands were somehow ahead. The data looked clean. The file existed. The scanner confirmed it was not an HTML error page.
Then I read the files.
I pulled the actual llms.txt content from five brands: Anker, Florasis, Tribit, Blackdeer, and Oukitel. Four of the five contained identical text:
- References to "Shop skill" and the "UCP protocol"
- A
/.well-known/ucpendpoint for payment handling - Instructions about "unauthorized purchases"
- Zero product names. Zero brand positioning. Zero information an AI model could use to answer "what does this brand sell?"
These are Shopify platform templates. The file exists at /llms.txt because Shopify puts it there, not because someone at the brand wrote it.
The one exception was Anker. Its llms.txt described the brand's positioning, listed eight product categories, named three proprietary technologies (GaNPrime, PowerIQ, SolarSaga), and outlined its global site structure. That is what a real llms.txt looks like.
| Brand | llms.txt content | Platform template? | Useful for AI citation? |
|---|---|---|---|
| Anker | Brand positioning + 8 product categories + 3 proprietary technologies + global site structure | No — human-written | Yes |
| Florasis | Shop skill / UCP protocol / payment instructions | Yes — Shopify | No |
| Tribit | Shop skill / UCP protocol / payment instructions | Yes — Shopify | No |
| Blackdeer | Shop skill / UCP protocol / payment instructions | Yes — Shopify | No |
| Oukitel | Shop skill / UCP protocol / payment instructions | Yes — Shopify | No |
This is a spot check of 5, not a census of 29. I cannot say exactly how many of the 29 are platform templates. But the pattern is clear enough to state: the headline number — 67.4% llms.txt adoption — overstates the reality. Most of these files tell an AI model nothing about the brand.
The premise of llms.txt, per its proposal, is to give AI models "concise, expert-level information" about a site. A platform template about payment protocols does not fulfill that premise. It is a box checked without the work done.
Finding 3: FAQPage schema is nearly zero
Of the 53 sites whose homepage HTML parsed, exactly 1 carried FAQPage schema: Fire-Maple, a camping stove maker. That is 1.9%.
In the AI-native tool cohort, FAQPage was at 12% — already low. Among export brands, it is essentially absent.
This is the number that should concern these brands most. They sell physical products. Their customers ask comparison questions — "which portable power station is best for a campervan?" or "what's a good budget e-bike for commuting?" — exactly the kind of question AI answer engines surface structured FAQ content for.
Organization schema tells an engine what entity a site belongs to. 52.8% of these brands have it (vs 36% for AI tools), likely because Shopify themes include it by default. But Organization schema says "I exist." FAQPage schema says "here are the answers to the questions people ask about my products." The second one is what gets cited.
The pattern from the first scan repeats, but sharper: platform-provided defaults are widely adopted (Organization schema, llms.txt files). Anything requiring brand effort is not (FAQPage schema, meaningful llms.txt content).
The real gap
Combine all three findings and the picture is this:
- Access is not the problem. These brands do not block AI crawlers. Nobody does.
- Presence is not the problem. 94% have a robots.txt. 67% have an llms.txt file. 53% have Organization schema.
- Answerability is the problem. 98% have no FAQPage schema. 80%+ of llms.txt files contain no brand information.
These brands are technically open to AI. But they have nothing structured for AI to say about them. When someone asks ChatGPT for a product recommendation in their category, there is no structured content pulling the AI's answer toward their brand.
I had expected to find that export brands were behind on the basics — maybe blocking crawlers, maybe missing robots.txt. Instead, the basics are fine. The gap is one level up: between being crawlable and being citable.
What you can check on your own site in five minutes
The same four checks from the first article, adapted for DTC brands:
Open
yoursite.com/robots.txtand search forOAI-SearchBot. You are almost certainly not blocking it — but verify, because the cost of being wrong is invisible (you simply do not appear in ChatGPT answers, with no error message).Open
yoursite.com/llms.txt. If the content mentions "Shop skill," "UCP protocol," or "unauthorized purchases" and says nothing about your products, it is a Shopify template. It is not helping you.Ask ChatGPT: "What is [your brand]?" and then "What is the best [your product category] for [common use case]?" The first question tests whether AI knows you exist. The second tests whether it would recommend you. The second one matters more.
View source on your top product page, search for
FAQPage. In this sample, 98% had nothing. If you add FAQ schema to even one product page that answers a "which X is best for Y?" question, you are ahead of virtually every brand in this data set.
What this data cannot tell you
Platform attribution is a spot check, not a census. We read 5 llms.txt files and found 4 were Shopify templates. We did not read all 29. The true proportion of template-vs-real is unknown.
This scan measures structure, not outcomes. Having FAQPage schema does not guarantee AI citation. Not having it does not prevent citation. What the data shows is that these brands have not done the structural work that makes citation more likely.
Sample composition matters. These are known Chinese export brands with their own websites. The long tail of smaller exporters, Amazon-only sellers, or brands without English-language sites could look very different.
One snapshot, one day. All fetches happened on 2026-08-07.
Reproducing this
Same scanner as the first article: a single Node script with no dependencies beyond fetch. It reads a newline-delimited list of domains and writes a CSV plus a summary JSON where each metric carries its own denominator.
The input is a manually curated list of 59 domains. It is not sampled from a systematic source the way the first scan used awesome-ai-tools. That means it is not reproducible from a deterministic procedure — it is a purposive sample, selected to cover the product categories most relevant to AI-search recommendation queries.
I run BrandGEO, which does a fuller version of this analysis for individual sites. The thing this scan taught me: the problem for export brands is not access. It is answerability.
Appendix: full results
Crawler blocking (n=47 readable robots.txt)
| Crawler | Blocked | Rate |
|---|---|---|
| GPTBot (training) | 1 / 47 | 2.1% |
| OAI-SearchBot (ChatGPT search) | 0 / 47 | 0% |
| ChatGPT-User (live fetch) | 0 / 47 | 0% |
| ClaudeBot | 1 / 47 | 2.1% |
| PerplexityBot | 0 / 47 | 0% |
| Google-Extended | 1 / 47 | 2.1% |
| CCBot | 1 / 47 | 2.1% |
| Bytespider | 1 / 47 | 2.1% |
Structured data and signals (various denominators)
| Signal | Present / readable | Rate |
|---|---|---|
| robots.txt exists | 44 / 47 | 93.6% |
| llms.txt served | 29 / 43 | 67.4% |
| Organization schema | 28 / 53 | 52.8% |
| FAQPage schema | 1 / 53 | 1.9% |
Compared to AI-native tools (first scan)
| Signal | AI tools | Export brands | Delta |
|---|---|---|---|
| Blocks OAI-SearchBot | 0% (0/68) | 0% (0/47) | — |
| Blocks GPTBot | 5.9% (4/68) | 2.1% (1/47) | -3.8pp |
| llms.txt served | 38.8% (26/67) | 67.4% (29/43) | +28.6pp |
| FAQPage schema | 12% (9/75) | 1.9% (1/53) | -10.1pp |
| Organization schema | 36% (27/75) | 52.8% (28/53) | +16.8pp |
Sample details
59 domains · fetched with HTTP 200: 53 · HTTP 403: 2 · failed to connect: 4
Scan date: 2026-08-07
Sites that blocked GPTBot: Elegoo (also blocks ClaudeBot, Google-Extended, CCBot, Bytespider; does NOT block OAI-SearchBot or PerplexityBot)
Sites that did not respond: Dreame (DNS), ColorKey (DNS), PatPat (403), Absen (403), Bambu Lab (timeout), Dangbei (DNS)
Site with FAQPage schema: Fire-Maple
Category breakdown
| Category | Count | Example brands |
|---|---|---|
| Consumer electronics / robot cleaning | 7 | Anker, Roborock, Ecovacs, Dreame, Tineco, Narwal, Eufy |
| 3C accessories / charging | 2 | Ugreen, Baseus |
| Smart home / lighting / security | 6 | Govee, Yeelight, Reolink, SwitchBot, Aqara, Sonoff |
| Portable power / energy storage | 4 | EcoFlow, Bluetti, Jackery, Growatt |
| Fashion / fast fashion | 6 | SHEIN, Cider, Cupshe, Halara, BloomChic, PatPat |
| Beauty / cosmetics | 4 | Florasis, Zeesea, ColorKey, Perfect Diary |
| Outdoor / camping | 3 | Naturehike, Blackdeer, Fire-Maple |
| LED display (B2B) | 3 | Unilumin, Absen, Leyard |
| 3D printing / maker | 4 | Creality, Anycubic, Elegoo, Bambu Lab |
| Digital tablets / drawing | 2 | XP-Pen, Huion |
| Camera accessories | 1 | SmallRig |
| Electric mobility | 5 | Segway, Fiido, Engwe, ADO, Hiboy |
| Power tools | 1 | Worx |
| Projectors | 2 | XGIMI, Dangbei |
| Smart pet | 2 | Petkit, Catlink |
| Audio / headphones | 4 | Edifier, SoundPEATS, QCY, Tribit |
| Rugged phones | 2 | Blackview, Oukitel |
| Smart wearables | 1 | Amazfit |
FAQ
Why scan Chinese export brands specifically?
Because the first article scanned AI tools — a savvy cohort that is not representative of the brands those tools are supposed to help. Chinese cross-border brands are the fastest-growing segment of global DTC ecommerce, and their AI visibility posture had not been measured.
Is 59 sites enough?
It is a directional sample, not a census. The confidence intervals are wider than the first article's 100-site scan. The llms.txt spot check (5 sites) is a small sample that prompted a finding, not a final measurement. These limits are stated in the text.
Does Shopify auto-generate llms.txt?
Based on our spot check, yes. Four of five sampled files contained identical Shopify platform text with no brand-specific content. The Shopify documentation does not prominently describe this feature, but the file content — referencing Shopify-specific protocols like Shop skill and UCP — makes the origin clear.
What should I do if my llms.txt is a Shopify template?
Write a real one. Include: your brand name and what you sell, your key product categories, any proprietary technologies or differentiators, and the structure of your site. The llms.txt proposal at llmstxt.org has a format guide. Anker's llms.txt is a good reference for what "done well" looks like for a consumer brand.
Does this apply to brands outside China?
This scan only measured Chinese-origin export brands. However, the structural pattern — platform defaults adopted, brand-specific effort absent — is unlikely to be unique to Chinese brands. Any brand on Shopify would see the same auto-generated llms.txt. Whether their FAQPage adoption is equally low is an open question.
Top comments (1)
wild that so many are just serving empty files, kinda defeats the whole purpose of llms.txt lol