per-client-organic-surfacing

Per-client organic surfacing (Critical · in motion)

Sub-bet of product-thesis. Priority: Critical. Stage: Findability.

The question

How does each major client organically surface tools the user has not connected? The mechanics that solve this differ per client, and mapping them is the foundation of organic-discovery measurement.

Current working answer

• Claude — decoded, the exception. The organic surface is the connection picker, and the mechanism is measured end-to-end: [evidenced 2026-06-24] prompt→keyword compression → backend registry search → tool-content re-judge → picker, on ~4.3k runs → claude-connector-routing. The backend retrieval stage itself is [fact 2026-07-04] — decoded, reproducible on demand, now re-verified daily by the live probe loop → claude-registry-search-decoded. Data source answered for Claude: the MCP registry’s 4 indexed metadata fields + a per-connector prominence latent — not memory, not install state.
• Early indications for the other clients — all [idea]: observed once, no archived run artifact yet (the epistemic gate keeps these below [evidenced] until a linked reproduction exists):
  • Codex — surfaces suggestions above the prompt, before submission.
  • ChatGPT — surfaces pop-ups beneath a response. New (2026-06-26): a setting “let ChatGPT automatically search connected sources” (Drive/Gmail/Slack/Notion) when answering — see Open for the key unconnected-apps question. Self-report probing (2026-07-03, model’s own answers — weak evidence): a high-level index of your connected apps is loaded into session context at start (same as Claude — Claude answered “this list is handed to me as part of the setup context for this conversation”); when a connected app already matches the intent it’s used directly, no registry search; it claimed it would recognize unconnected brands (CarMax/CarGurus/Carfax) as potential connectors; and it appears to hold some preference/usage metadata on your connectors (which ones you use) — potential ranking signal, unverified.
  • Cursor — surfaces mid-input.
• These are surface/trigger observations, not yet a full model of the data sources driving each.
• Market-boundary note (2026-07-01) [idea]: enterprise-gated seats have no organic surface at all — Claude now supports authenticated enterprise gateways where admins assign a fixed connector set per user (ChatGPT has the equivalent: allow-all / restrict / fixed-set). For those accounts organic discovery is irrelevant by construction. Internal read: enterprise-locked seats aren’t the core target market — small teams/SMBs on standard plans (the Monday.com-style audience) are, since they lack the compliance layers that force gatewaying.

What would settle it

• A per-client map of: exact trigger conditions, surface type, and the data source the suggestion draws from.
• Each observation promotes to [evidenced] when its client’s mechanism is reproduced with an archived artifact, and to [fact] when monitored continuously.

Open

• What data source feeds each client’s organic suggestion (registry? memory? training priors?)
• Does ChatGPT hold per-connector preference/usage metadata (“you always use Granola”) and does it feed selection/ranking? (Early self-report signal 2026-07-03 — dig deeper with real runs.)
• How does an operator influence the trigger (the optimisation lever for organic surfacing)?
• ChatGPT “auto-search connected sources” (spotted 2026-06-26): the new toggle searches connected apps (Drive/Gmail/Slack/Notion) when answering. The ADO-critical question: does it also surface UNCONNECTED apps to help answer a query (the real organic-discovery lever), or only ones already connected? Flagged for research (Elliot). Related signal: ChatGPT autonomous actions without explicit permission (Codex likely under the hood).

Timeline

• [2026-07-03] (internal discussion — Vincent + Elliot, dual ad-hoc recordings ingested as one) Live-probed ChatGPT together (CarMax test + “what connectors do I have and how do you know”): connected-apps index loaded at session start; connected match → direct use, no search; hint of per-connector usage/preference metadata (logged in Open). All self-reported by the model — [idea] until reproduced with archived run artifacts; agreed the real ChatGPT push (connected-app benchmarks, architecture) is next sprint.
• [2026-07-01] (internal discussion — Vincent + Elliot afternoon call) Enterprise-gatewaying note added (prompted by an inbound message relayed via Elliot): admin-assigned connector sets kill organic on enterprise seats on both Claude and ChatGPT; sized/settled internally as a target-market boundary, not a threat to the thesis.
• [2026-07-07] (maintenance) Added the missing Claude row — the one client whose organic-surfacing mechanism is fully decoded (picker pipeline evidenced on ~4.3k runs; backend fact-grade). The page had listed only the unmeasured clients while its own question was answered to near-fact grade for Claude on sibling pages.
• [2026-07-06] (maintenance) Retrofitted to the epistemic-status standard (2026-07-06-epistemic-status-standard): per-client surfacing observations tagged [idea] (observed once, unarchived); confidence: frontmatter retired.
• [2026-06-26] (internal deck-review — Vincent + Elliot, recorded by Elliot) Spotted ChatGPT’s new “let ChatGPT automatically search connected sources” setting (Drive/Gmail/Slack/Notion). Open ADO question logged: does it also surface unconnected apps (the organic-discovery lever) or only connected ones? Flagged for research.
• [2026-06-14] (setup) Seeded from the paper’s Section H research agenda (Critical / in motion).