I audited eight stablecoin and RWA platforms for AEO. The pattern surprised me.

A couple of days ago I read HIFI’s comparison page on best stablecoin infrastructure. It names four providers: HIFI, Zerohash, Stripe (via Bridge), and Mastercard (via BVNK). It does not name OpenTrade. It does not name Ondo. It does not name Plume.

That kind of comparison page is now training data for the answer engines — ChatGPT, Perplexity, Claude, Google AI Overviews. An institutional buyer asking “who runs the rails for stablecoin yield products” in a chat box is getting an answer shaped, in part, by who shows up in pages like that.

So I ran a small audit. Eight digital-asset platforms — two of HIFI’s four (HIFI itself and Zerohash; Stripe Bridge and BVNK skipped because their AEO surface sits behind larger parent-company sites), plus six excluded from the comparison entirely (Centrifuge, Maple, Plume, Ondo, OpenTrade, Backed). I checked the surfaces that AI crawlers and agents look at: llms.txt, .well-known/agent-card.json, JSON-LD schema, and AI-bot rules in robots.txt. Root domain and the common subdomains.

The data is below. It went in a direction I did not predict.

The audit, 7 June 2026

Surfaces covered: count of reachable subdomains with at least one of llms.txt, agent-card, JSON-LD, or AI-specific robots rules. Stripe Bridge and BVNK skipped — both sit behind larger parent-company sites where the AEO surface isn’t separable. Audit re-runnable via check-agent-surfaces.ts in our open tooling.

What surprised me

I expected the AEO posture to be uncorrelated with which platforms get cited. Crypto sites have historically optimised for crawlable token data — CoinGecko, DefiLlama, block explorers — not for AI answer surfaces. So I expected to find a fairly random scatter of llms.txt presence across the eight.

Instead, the AEO posture spans the full range across both groups, and being named in HIFI’s comparison turns out not to track surface quality. HIFI itself and Zerohash — the two from HIFI’s list I could audit at company level — both sit at 2/6, the same level as Maple and Plume. Ondo at 3/6 has a stronger posture than either of HIFI’s named peers. Centrifuge runs the most mature surface posture in the table at 5/6, despite not being in HIFI’s list. OpenTrade at 1/6 is genuinely the bottom. (I excluded Stripe Bridge and BVNK from the audit because their AEO surfaces inherit from the parent companies, which is a different measurement question.)

So the real signal isn’t that HIFI’s omissions all have weak AEO. It’s that AEO posture and being in a competitor’s comparison are two independent vectors. Ondo has a polished agent-readable surface and is still not in HIFI’s post. The answer engine learns the category shape partly from each platform’s own surfaces and partly from who other platforms mention by name. Strong on one and weak on the other still leaves you out of the composed shortlist.

The cleanest contrast in the table is OpenTrade vs HIFI. Both run developer documentation with llms.txt. Only HIFI also publishes Organization, FinancialService, ContactPoint, and PostalAddress schema on the marketing site, plus an agent card on docs. OpenTrade’s marketing site has none of that. The HIFI comparison post then writes HIFI’s positioning explicitly into the training data. OpenTrade is, structurally, not in the conversation.

Why digital-asset companies are unusually exposed

Three reasons this matters more for digital-asset infrastructure than for, say, a Nordic dental clinic.

1. The buyer prompt is long and structured. An institutional buyer doing platform diligence is not asking the chat box “best stablecoin platform.” They are asking something like “Who provides regulated stablecoin issuance with EMI licensing in the EU, supports off-ramping to local banks in 10+ jurisdictions, and has SOC 2 Type II?” The answer engine has to compose a shortlist from structured signals across the candidate set. Companies that publish FinancialService schema with explicit jurisdiction and license fields show up in those composed lists. Companies that publish marketing prose do not.

2. The category is YMYL. “Your Money or Your Life” pages — finance, regulated services, anything where bad advice causes harm — have always been held to higher standards by Google. The same instinct now runs through the LLM-grounding stacks: models lean harder on Wikipedia-style structured signals (Organization data, ContactPoint, founded-on date, regulatory registrations) before they’ll quote a financial-services site in an answer. The bar for “trust signal” is higher in your category than in most.

3. Your competitors are publishing the comparison pages. The HIFI page is a good example, but it is not unusual. Stablecoin platforms, RWA platforms, and DeFi protocols write each other’s comparison pages every week, mostly inside their own blog or docs. Each of those posts is a small bundle of training data that defines who “exists” in the category for an AI model. If you are not in those pages, the answer engine learns the shape of the category without you in it. Counter-attack is to be in the next round of those pages, on your own terms, with a structured comparison post of your own — and to make your marketing site easy to ingest when an AI crawler arrives.

What to actually do

The fix surface for a digital-asset company is small and concrete. In rough priority order:

1. Move llms.txt to the root domain. Three of the seven platforms with docs-level llms.txt have nothing at the marketing root. docs.opentrade.io/llms.txt is great for developer agents; it does not help when a journalist or institutional buyer pastes the homepage URL into ChatGPT. Both surfaces matter, for different audiences.
2. Publish proper Organization / FinancialService JSON-LD on the homepage. Name, legal entity, founding date, parent organisation, the regulatory licenses you hold (with jurisdiction), the products you offer, ContactPoint, PostalAddress. This is the data the models reach for when they need a trustworthy structured statement of what your company is. HIFI’s root JSON-LD is the cleanest example in the eight: five schema types, all loadable in one parse.
3. Allow the AI bots in robots.txt, explicitly. Zerohash names GPTBot, ClaudeBot, PerplexityBot, and ChatGPT-User. Centrifuge names GPTBot, CCBot, and anthropic-ai. Defaults vary by hosting provider; an explicit allow removes ambiguity. If you also want to deny training but allow real-time grounding crawlers, name them separately.
4. Publish your own structured comparison post. If your competitors are writing the category-defining pages, write one yourself. Name the platforms you compete with directly, including yourself, with structured criteria. This is the single highest-return piece of AEO content a digital-asset company can ship in 2026, because it gets ingested by the answer engines as a category artifact — not as marketing.
5. Add an agent card on docs. Lowest-effort, lowest-impact today, but trending toward higher impact as A2A and agent-mesh patterns mature. HIFI and Backed have already done this. Centrifuge has gone further and put one on the app itself.

The whole list is three to ten days of work at a typical platform engineering team. It is not a content strategy. It is a configuration change with copy attached. The compounding return is “your company shows up in the next round of training data the answer engines ingest about your category.”

One caveat

The audit table is a snapshot from 7 June 2026 of agent-readable surfaces, not a verdict on which platform is best. Centrifuge has the most mature surface posture of the eight; that does not make it the best RWA platform. HIFI’s comparison page is unsurprisingly favourable to HIFI; that does not make their evaluation wrong, only interested. Read the table for what it is: a structural audit of how each platform is talking to the AI infrastructure layer, not to humans.

The thing I will keep watching is whether the bottom-of-table platforms show up in the next round of category comparisons that get written this summer. If they don’t, that is the answer engines doing their job correctly with the inputs they were given.