Builder project · Built at NYU SPS X Pulse Foundry
The Secret Sauce
Application that recommends changes for restaurants to help them rank higher in AEO searches

Tip: click @Pulse AI in the post to tag us.
From idea to reality
What we built
## The Secret Sauce **An AI growth agent that helps Café Athena get recommended when diners ask ChatGPT, Gemini or Claude where to eat.** Built for the owner, not an analyst: a guided six-step story you can understand at a glance, not a dashboard. ### The problem Diners now ask AI assistants "best happy hour in Pacific Beach" instead of scrolling Google. Café Athena has been on Garnet Ave since 1990 and has Zagat and San Diego Magazine awards, yet AI named it in only **35 of 99** diner questions we tested. It wins "Greek food" questions but disappears for parties (**0 of 15**), date nights (**1 of 9**) and happy hour (**3 of 13**), even though it hosts groups and runs happy hour every day. ### How it works 1. **Discover:** reads cafeathena.com live, the way an AI crawler does, and lists what's missing in plain language: *AI can't tell you're a restaurant* (no Restaurant schema), *no FAQ*, *mixed messages about happy hour* (homepage says "weekday", menu says "every day"). 2. **Results:** our 116-question bank across 11 categories, with where Café Athena ranks for each. The owner can ask any question live and watch Claude answer. 3. **Tailor:** the owner sets what matters (High / Medium / Low), keeps or deletes AI-suggested categories, and edits questions. A personalized score follows their priorities. 4. **Rivals:** head-to-head against Olympic Cafe and Spitz. Claude with web search answers 32 questions and counts who gets named in each category, then researches what each rival has online. 5. **Fixes:** 7 ready-to-publish fixes written only from facts on the site (Restaurant schema, FAQ page, events page, dietary guide, happy-hour copy, awards copy, press pitch), ranked by the owner's priorities. Each shows before/after, can be edited or rewritten by Claude, and is approved by the owner. Plus social posts matched to their priorities. 6. **Track:** projected lift, then a re-test. Claude answers the same questions against today's site and against the site with the fixes. It must quote the sentence that answers each question, and the server checks the quote exists, so the gain is grounded in evidence. **Under the hood:** a dependency-free Node server and vanilla JS front end. Every AI step runs through Claude Code in headless mode (`claude -p`), with web search where real-world answers matter. ### Proof it worked With four approved fixes, the re-test went from Café Athena named in **2 of 12** questions to **6 of 12**, and every new mention is backed by a quoted sentence from the new content. Date-night questions stayed unfixed, which honestly shows what's left to do. ### Why it matters - **Owner in control:** nothing is published without approval, every change can be undone, and nothing is emailed or posted automatically. - **Honest numbers:** recorded, live, projected and simulated results are always labeled. - **It keeps working:** a weekly re-test tracks visibility after the hackathon ends.