The problem
Generative engine optimisation is a market full of claims and short on evidence. Agencies say they get businesses recommended by ChatGPT; almost none publish a prompt set, a method, or a measured result. If Dubai Foundry was going to sell GEO, the consultancy itself had to be the proof: a site built exactly the way we tell clients to build, measured the way we tell clients to measure, with the numbers recorded whether they flatter us or not.
What we built
A conversation-first consultancy site. BoostSearch sells no packages. Every page routes to one action, a conversation, and the site says plainly what we do not do: no 90-page automated reports, no promises to control what AI says, no prices without understanding the business. The design is its own system, ink on paper with a highlighter motif, built on our design tokens.
The GEO stack we recommend, live. Every page is complete server-rendered HTML that reads without JavaScript. A JSON-LD entity graph ties the organisation, its Dubai service area and its parent studio together. robots.txt explicitly welcomes every AI crawler, llms.txt hands assistants a curated map, IndexNow pings Bing on every deploy, and a visible canary token in the footer proves when an index has re-read the site.
A free AI readiness checker. boostsearch.ai/check fetches any site the way an AI crawler would and reports eight mechanical checks: crawler access, llms.txt, structured data, sitemap, content without JavaScript, indexability. No email gate. The tool is the demonstration; the conversation is the conversion.
A measurement pipeline. Every morning a cloud routine expands our seed prompts into the search queries AI assistants run, searches them for real, and records whether boostsearch.ai appears and which competitors do. Monthly, a probe asks the real models the real prompts through the Claude and OpenAI APIs with native web search, and records the fan-out queries they actually issue and the businesses their answers actually name. Crawler hits per bot family are aggregated from server logs. The daily report lands in our inbox with the honest number in the subject line.
Why these choices
Measurement before optimisation. The site launched, was baselined at zero, and is now held deliberately unchanged while the measurement runs. Optimising against a signal you have not received is guessing. When the numbers move we will know what moved them.
Real observation over simulation. The answer probes read the fan-out queries from the models themselves rather than guessing what they might search. Those observed phrasings calibrate the daily watch.
The same one-box discipline as Wipe Board. A Next.js standalone build on a small Hetzner server behind Caddy, deployed by a GitHub Action on every push. No moving parts that wake anyone up at night.
What happened
The honest version: it is early. The site went from empty folder to production, with TLS, analytics, Search Console and Bing verified and sitemaps submitted, inside a day. Google, Bing and OpenAI's crawlers visited within hours. Day-one answer probes confirmed what we expected: nobody, including us, has yet claimed the AI answers for generative engine optimisation in Dubai. That baseline, and the pipeline that watches it move, is the product. We will publish the climb as it happens, including the flat parts.