How to build a site AI answer engines will cite — 371 articles across 7 languages

Saying "we do AI-answer-engine optimization" is easy and unevidenced. The only way to have proof is to build and operate a site whose explicit goal is being cited by AI answer engines, end to end.

I chose proxy and scraping-infrastructure comparison as the subject: prices and specs change often, structuring facts genuinely pays off, and there is demand across many languages.

Most of this work stops at "we did that" without ever being true in production.

You can add AI crawlers to robots.txt and have no guarantee the deployed file says so. You can emit hreflang and make things worse by declaring locales whose translations don't exist. You can ship an llms.txt that is quietly stale.

The other difficulty was that analytics on a young site return nothing. Using a fixed impression floor to find "almost ranking" queries returns zero rows for weeks.

Turn every "we did that" into a test.

The AI-crawler allow-list lives in its own pure data module, and robots.txt is generated from it. On top of that there is a test that fetches the deployed production robots.txt and checks it against that same list. It verifies the output, not the configuration.

hreflang is computed per article from the set of translations that actually exist. Locales that haven't been translated are never declared.

What I gave up was publishing speed. 53 articles × 7 languages means each additional language raises the marginal cost of every article. I chose to keep the pillar-and-cluster structure intact across all languages rather than mass-produce shallow posts.

Pricing data is not something to be scraped out of article prose. It is published as a CC BY 4.0 machine-readable dataset at its own URL, with Dataset structured data attached. The reuse terms travel with the data, and the disclaimer ships as a field of the dataset so that context survives even when the data is lifted out on its own.

Locale design puts English at the root (cleanly serving as x-default) with the other six prefixed. zh targets Traditional Chinese (zh-Hant / zh_TW) and pt targets Brazil (pt-BR), with the language tag, hreflang value and OG locale kept as separate fields — reuse one string for all three and it is guaranteed to be wrong somewhere.

IndexNow submits only to the engines that participate, with the fact that Google does not participate noted in the code — so nothing that doesn't work is written up as though it does.

Search Console queries are classified as answer-engine fan-out or human and fed into an editorial backlog. Because a fixed floor returns nothing on a young site, the impression floor is adaptive.

72 commits between 2026-07-17 and 07-27. 53 articles × 7 locales = 371 MDX files, with 3 pillars and 50 clusters per locale. 69 test files against an 80% coverage floor.

CI runs a quality gate (content build → formatting → ESLint at zero warnings → type check → content validation → coverage) alongside a dependency audit, Gitleaks over full history, and Semgrep SAST.

The AI-crawler allow-list is 10 agents: GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, ClaudeBot, Claude-User, Google-Extended, Applebot-Extended and CCBot. A deploy verification confirms those 10 are actually present in production's robots.txt.

The page you are reading was built the same way. Structured data, llms.txt, hreflang, and the constraint that no number appears unless it can be counted — everything tried on ProxyFacts is applied to this portfolio itself. Reading about answer-engine optimization on a site that does it is the easiest kind of evidence to check.

Building hreflang across 7 languages from only the translations that exist, making data citable through a machine-readable dataset, and proving a measure actually reached production via a test — all of it ports directly to a content platform you are about to build.

I also take on multilingual content-platform design as fractional CTO / technical advisory.