Cerca has a functioning place-memory and social product plus a substantial Discover MVP. The live audience is still a small cohort: existing behaviors show useful activation and valuable taste data, but there is not yet clear evidence of a durable place-saving habit.
The unanswered question is not whether Cerca can build an LLM recommender. It is whether Cerca's first-party taste context produces recommendations compelling enough that people return for another real decision.
Do not pivot away from saves, ratings, and friends. Recast them as Cerca's compounding intelligence layer, with Discover as a possible new front door.
A standalone "LLM for places" would compete directly with products such as Ask Maps, which already combines conversational recommendations, saved and search history, extensive place data, and action completion, and Yelp Assistant.
Cerca's differentiated hypothesis is personal and trusted-network taste—not the chat or interview interface itself.
- Keep the existing six-answer interview for the first pilot, but do not treat profile completion as the product's success metric.
- Preserve the existing place, rating, guide, and social functionality. Do not deprecate or reposition it until Discover demonstrates repeat use.
- Continue using personal ratings, interests, and place anchors in recommendations.
- Defer friend-derived prompt context until the core loop works; a sparse friend graph would make the first test harder to interpret.
- Add minimal recommendation feedback: Good fit or Not for me.
- Track outbound place opens.
- Do not build reservations, full save-resolution, or additional profile machinery for this test.
Track:
- discover_interview_started
- discover_interview_answered
- discover_profile_ready
- discover_run_created
- discover_run_completed
- discover_run_failed
- discover_recommendation_opened
- discover_recommendation_feedback
- discover_second_session_started
Properties may include question number, profile version, category type, location mode, rank, recommendation count, latency, and signal counts.
Never send answer text, profile prose, exact location, friend identities, or private notes.
Invite 12–15 existing non-test users with at least five meaningful ratings or saves. Require at least eight to complete the interview and one real recommendation run.
- Owner: Arvin
- Decision date: Day 29 after the first cohort invitation
Before access, ask each participant about their last real place decision:
- What triggered it?
- What tools or people did they use?
- How long did it take?
- What was unsatisfying?
- What did they ultimately choose?
Then let them use Discover for a current decision. Do not ask whether they "like the idea." Observe whether they inspect a recommendation, give fit feedback, and return without prompting for another decision.
At least 5 of 8 activated participants create a second Discover run on a different day within 14 days.
- At least 60% of completed runs produce a Good fit response or outbound place open.
- Users can explain why the results were more personally relevant than their normal workaround.
- Fewer than 10% of runs fail technically or return no usable recommendation.
- Negative feedback is attributable to fixable ranking or context errors rather than "Google Maps or ChatGPT already does this just as well."
- Continue: At least 5 of 8 repeat, with acceptable recommendation quality. Next, test bounded friend-derived evidence and integrate recommendation saving into Cerca.
- Change: Three or four repeat, or users value results but interview friction suppresses activation. Shorten or prefill the profile flow before retesting.
- Stop or reposition: Two or fewer repeat despite eight activations and technically sound results, or most participants prefer their existing workaround.
These are directional founder thresholds for a small cohort, not statistically significant experiment claims.
$math-validate-startup
It should consume:
- the current Discover product and technical design;
- the recent behavioral baseline;
- interview notes centered on the last real decision;
- the analytics contract and thresholds above.
Keep provisional:
- that an explicit taste interview is necessary;
- that social signals improve results;
- that repeat use will drive free growth;
- that the broader place-discovery market is reachable.
- $math-diagnose-metrics after the four-week cohort to evaluate activation, recommendation action, and repeat-decision behavior.
- $math-design-growth-experiments only after the repeat-use threshold is met, to test onboarding, distribution, and external positioning.
Fundraising, pricing, and market-size work remain out of scope until repeated decision use is demonstrated.