Named
The agent mentioned you. It may never open your page.
Seed narrative · Agentic selection infrastructure
Agents don’t
rank you.
They pick you.
A flight recorder for buyer agents. Replay the task. See who survived, who got picked, and where the agent gave up. Selection share and completion are the scoreboard.
Synap1e is working on Picked. Pre-product; the system below is proposed.
01The problem
A name in an answer
is not a sale.
GEO measures the answer. A delegated buyer task keeps going: shortlist, verify, book, buy. A brand can clear the first three states and still lose at the fourth.
Cited
Your page supplied evidence. That evidence may favor someone else.
Selected
You fit the task. Now the agent has to finish it.
Abandoned
A gate, a missing fact, a broken handoff. The agent leaves or switches.
Three states of visibility. A fourth outcome that a mention rate misses.
Picked’s intended metric: selection share + task completion.
Even naming and citing are different events.
ChatGPT named tracked brands in 33.3% of a MarketScale B2B sample and cited them in 12.8%. Overall mention rate was 26.4%; citation rate, 17.1%. These are answer metrics from Jul 7–Oct 5, 2026, not purchase outcomes.[2]
Visibility can concentrate before the task starts.
In Broadcastwell’s 860 B2B answers, vendor-authored pages accounted for 77.4% of citations among the top 100 cited domains. The category leader appeared in 80% of its category’s answers; 35% of companies were named in zero. One engine, Jul 18–23, 2026; not a market-wide census.[3]
02The proposed product
Replay the buyer.
Record the decision.
One delegated task across ChatGPT agent mode, Claude computer use, Gemini agentic tasks, and Perplexity Comet. Log the observable retrieval set, applied criteria, selection, and abandonment. Access to hidden model reasoning is not assumed.
01Define the job
Task library
Canonical jobs with hard constraints and a definition of done. Reuse the job across brands; otherwise every replay becomes bespoke consulting.
“Payroll for a 40-person company, multi-state, demo this week.”
“Book a pediatric dentist in Round Rock this week who sees a 20-month-old.”
“Find a carry-on under $180 that fits a Tesla Model Y frunk.”
02Hold the task constant
Multi-agent replay
Same task, four agents. They do not behave the same. Record each system’s visible path instead of averaging away its failure.
- Claude checks more criteria.
- Gemini abandons faster on non-actionable pages.
- Perplexity fans out widest and transacts least.
Working replay hypotheses from the seed brief, not Picked benchmarks or guaranteed behavior.
03Find the deciding constraint
Selection graph
Who was retrieved? Who survived evaluation? Who was selected? Show which constraint flipped the decision, with the page or cited evidence attached.
Retrieved→Survived→Selected→Done / abandoned
Selection share: share of eligible canonical replays that choose the brand. Completion: share that reach the task’s definition of done. Neither is live buyer volume.
04Turn the trace into a patch
Abandon autopsy
Cookie wall. Login gate. Stock that disagrees with the page. Missing constraint. Contradicted claim. Missing from retrieval. Give the failure a reason code, then a specific patch.
Reason code→Feed / page / endpoint→Rerun the same task
A feed field, suitability block, booking endpoint, or fact correction. A patch is a testable recommendation, not a promised lift.
Canonical task / PAYROLL-040
“Payroll for a 40-person company, multi-state, demo this week.”
40 employeesMulti-state payrollDemo available this weekPublic evidence of coverage
Definition of doneA valid demo request submitted with a confirmation. No real submission has been made.
ChatGPT
Agent mode / illustrative trace
Selected you at #4- Retrieved 8 candidates.
Your brand entered at rank #4. - 3 survived the multi-state and demo constraints.
- Your coverage evidence matched the task.
Selected your brand. - Demo confirmation recorded.
Definition of done reached in this sketch.
Decision flipped on: suitable coverage.
Claude
Computer use / illustrative trace
Lost on a contradicted claim- Retrieved 7 candidates and checked coverage.
- Your page said all states.
A third-party source said 2 states. CLAIM_CONFLICT
Your brand did not survive evaluation.- Selected the category leader.
Patch: reconcile the coverage fact.
Gemini
Agentic task / illustrative trace
Abandoned the demo- Retrieved 4 candidates.
Your brand fit the stated constraints. - Opened the demo path.
Scheduling required a login. LOGIN_GATE
No public booking endpoint.- Task abandoned.
No confirmation returned.
Patch: an actionable demo endpoint.
Perplexity
Comet / illustrative trace
Shortlisted; did not transact- Retrieved 12 candidates, the widest set in this sketch.
- 5 survived evaluation.
Your brand stayed on the shortlist. - Returned comparison evidence and demo links.
- No demo submitted.
Definition of done not reached.
Naming a shortlist did not complete the job.
Abandon / elimination reasons
CLAIM_CONFLICTCoverage contradicted → correct the fact on the page and in independent listings.LOGIN_GATEDemo scheduling blocked → expose a usable booking endpoint.No transactional attemptShortlist returned → completion stays incomplete; no success inferred.
Illustrative task-set selection share
22%vs61%
Your brandCategory leader
A separate, invented task-set summary, not calculated from the four displayed traces. Neither figure is a Picked customer result or an external benchmark.
Sketch only. No agent connection, live replay, or transaction runs on this page.Compare the same task set, environment, and agent version.
The proposed reason-code → patch taxonomy
COOKIE_WALLCritical facts blocked by consent UI → make the suitability block readable before nonessential cookie choices.
LOGIN_GATERequired action hidden behind an account → expose a supported booking endpoint and clear handoff.
STOCK_MISMATCHAvailability conflicts with the page → correct stock and availability in the feed field.
MISSING_CONSTRAINTFit cannot be established → publish the exact qualifying fact in a suitability block.
CLAIM_CONFLICTIndependent evidence disagrees → make a fact correction across the page, feed, and third-party record.
NOT_RETRIEVEDNo candidate-set entry → fix the relevant feed field or discoverable category suitability evidence, then rerun.
Fit can matter more than retrieval rank.
Suitability content: selection 11.4% without, 30.8% with, 2.7x at comparable rank. Correcting a wrong prior about the brand: 13.9% → 27.5%, about 2.0x. These are study observations, not promised Picked outcomes.[1]
The brand’s page is not the only witness.
Agents drew on prior brand beliefs in 81.6% of commands, checked 3.4 independent sources per claim, and eliminated candidates in 27.9% of conflicting-claim cases in the reported sample. The product must retain evidence, not infer an invisible thought process.[1]
The category boundary
Prompt panels
versus action traces.
Profound, Peec, Scrunch, and Ahrefs Brand Radar already sell the visibility answer: “were you named?” Picked’s proposed question is what happened after the buyer delegated a task.[9]
| Dimension | GEO visibility platforms | Picked / proposed |
|---|---|---|
| Unit of work | Prompt panels and generated answers | Canonical buyer task with constraints and a definition of done |
| Data | Mentions, citations, answer position, sentiment | Observable action traces, candidate survival, selection, abandoned steps |
| Scoreboard | Were you named, and how often? | Selection share and task completion on a controlled replay set |
| Recommended work | Visibility and content recommendations | A reason code tied to a feed field, suitability block, endpoint, or fact |
| Proof of value | Movement in the answer panel | The same task rerun after a patch, with a changed selection or completion outcome |
This is a funded category, not an empty market: Profound raised a $96M Series C at a $1B valuation in February 2026 and reported 700+ enterprise customers in that announcement. Its capital is a distribution risk for Picked, not evidence of demand for this exact product.[8]
03The wedge
Start with B2B software.
Earn the second vertical.
Payroll and HR first, until the trace is undeniable. Comparable constraints, visible claims, and a demo endpoint. Do not boil the ocean.
First / B2B software
Prove the patch loop.
10.8%[1]
Q3 2026 category agentic share. Build one payroll/HR task library, sell repeated measurement, and require a before/after rerun.
Second option / Ecommerce catalogs
Make the feed testable.
12.3%[1]
Q3 2026 category agentic share. SKU constraints, stock disagreement, feed diffs, and competitor switches. Enter only after the first category works.
Second option / Bookable local
Measure the last step.
7.4%[1]
Q3 2026 consumer and local services share. Suitability and real appointment availability. An alternative second vertical, not a parallel launch.
The largest share is not automatically the best wedge. Travel booking was 14.6% in Q3 2026; the first market remains B2B software. These category shares are First Page Sage estimates, not Picked’s serviceable market.[1]
Timing signal, not TAM: under 5% of sites optimized for agent traffic, Digidop domain audit, May 2026, via Similarweb. After OpenAI’s Jul 10, 2026 model update, about two-thirds of ChatGPT Shopping recommendations came from merchant feeds rather than crawled pages, according to Profound’s analysis cited by Verlua in Sep 2026; these were monitored prompts, not purchases.[5][6]
Commerce context ≠ Picked TAM. McKinsey, Oct 17, 2025: U.S. B2C agentic commerce $900B–$1T by 2030, global goods $3–5T. Includes AI-shaped purchases, not only agent-completed checkout. Excludes services and B2B.[7]
04Intended model / not live pricing
Sell the rerun.
Reuse the task.
Charge for a repeatable selection diagnostic, not a dashboard full of names. These are proposed plans; no subscriptions, signed partners, or revenue are claimed.
Design partner
Discounted
Intended research arrangement
- Agreed data rights
- 50 tasks
- Weekly selection share
- Scope and discount to be negotiated
Team
$1,500
/ brand / month · intended
- 200 replays
- Four agents
- Abandon codes
- A specific patch list
Catalog
$4,000
/ month · intended
- SKU or location task sets
- Feed diff
- Competitor switch report
- Second-vertical offering
Enterprise
Custom
Intended private deployment scope
- Private tasks and SSO
- Raw observable traces
- Warehouse API
- Scope-based pricing
Gross margin only works if a category task library is reused across brands. Replay cost is not known yet. Measure runtime, retries, and human review before treating these prices as viable.
The intended moat
Schema and llms.txt are not the moat.
A First Page Sage study, Oct 2, 2026, covered 4,213 commercial prompts and 657 agentic tasks. Schema’s recommendation lift fell to 0.4 points after controlling for authority. Clarity of offering doubled recommendation rate, then leveled off. On agentic tasks, suitability and information consistency outweighed schema; this was a vendor-led study, not proof of Picked’s advantage.[4]
Trajectory data
Permissioned, versioned paths from retrieval to a completed or abandoned task. Evidence accumulates through repeated runs; none has been accumulated by Picked yet.
Category task graphs
Reusable jobs, constraint hierarchies, and definitions of done. A library that captures payroll buyers’ actual constraints is harder to copy than a markup checklist.
The patch loop
Trace → reason code → change → rerun. Keep the task and environment comparable, and record failures as well as wins. Causality still needs disciplined evaluation.
What we will not claim: no proprietary OpenAI or Google logs, no access to private chain of thought, and no platform partnership. Picked would record what its own authorized replay can observe.
05The investor question
Say the risks first.
The opportunity is a decision record that helps a brand change an outcome. The burden of proof is a successful patch and rerun, not the size of agentic commerce.
A platform can ship this.
First-party selection reports could compress the category. Picked would need independent cross-agent evidence and a useful patch loop; neither is proven.
The interfaces keep changing.
Agent UIs change monthly. Harness maintenance, task validity, and permissioned access are ongoing costs, not a one-time integration.
Replay can outrun the price.
If category tasks do not reuse, runtime and human review can consume the subscription. The first library must prove cost reuse as well as selection movement.
Suggested / not committed
$2.5–4M
seed
A proposed financing range. No round, investor commitment, customer traction, or signed design partner is claimed.
Why this could be fundable
A new decision surface.
A narrow, falsifiable wedge.
Build a reliable harness for the same buyer job across agents. Start with payroll and HR. Show that a recommended change moves selection share on a controlled rerun. Then learn whether buyers will pay to repeat the measurement.
Proposed use of funds
Agent browser runtime · Eval harness · First B2B task library · Recruitment of 15 design partners as a target, not existing relationships.
Proof to raise against: selection share moved on a rerun after a recommended patch.
The intended 18-month plan.
Milestones, not completed work or commitments.| Window | Build | Evidence to earn |
|---|---|---|
| 0–4 months | Harness for two agents and 50 payroll/HR tasks. | Stable, reviewable traces for the same task; measure replay cost. |
| 4–9 months | Four agents, abandon taxonomy, ten design partners, first before/after. | Partners are a recruitment target. Show selection share movement after a patch. |
| 9–14 months | Second vertical and self-serve Team plan. | Choose ecommerce catalogs or bookable local; prove task reuse before broadening. |
| 14–18 months | API, private task sets, renewal on selection share. | Customers renew for decision evidence, not additional mention charts. |
Source notes.
Research windows matter. These are dated industry studies and projections, not Picked traction. The console is an illustrative product sketch. Agent-specific replay behavior is a hypothesis to test. The pricing, financing, and roadmap are proposed.
- [1]First Page Sage · Oct 6, 2026. Agentic SEO, Explained. U.S. search-share estimates, category shares, and the command study (n=2,417; Mar 4–Jun 10, 2026). Rank, suitability, prior-belief, verification, completion, and switch figures are reported observations in that vendor-led sample; they are not universal agent behavior or Picked’s results.
- [2]MarketScale · Jul 7–Oct 5, 2026. The State of GEO: AI Visibility for B2B Companies. ChatGPT: 33.3% named, 12.8% cited. Overall: 26.4% mention rate, 17.1% citation rate. A tracked B2B sample; neither metric measures selection or completed transactions.
- [3]Broadcastwell · Jul 18–23, 2026. The 2026 State of Generative Engine Optimization. 860 B2B answers; one engine held constant. The 77.4% vendor-authored citation share covers the top 100 cited domains. The 80% figure is a median category-leader rate; 35% of companies were named in zero. Not peer reviewed.
- [4]First Page Sage · Oct 2, 2026. Impact of Structured Data on AI Rankings. 4,213 commercial prompts, 657 agentic tasks. Schema: 0.4-point recommendation lift after authority controls. Clarity doubled recommendations, then plateaued; suitability and consistency mattered more on agentic tasks. A study of observed associations, not a guaranteed intervention effect.
- [5]Digidop domain audit, May 2026 · via Similarweb. Similarweb’s agentic search guide, Jun 1, 2026, attributes the under-5% figure to that audit. The separately linked Digidop article is dated Apr 26, 2026. This page uses Similarweb’s May-audit attribution; the public article is not the underlying audit dataset.
- [6]Profound analysis · cited by Verlua, Sep 2026. Agentic Commerce Optimization and Profound’s Shopping analysis. About two-thirds of ChatGPT Shopping recommendations came from merchant feeds after the Jul 10, 2026 update. Monitoring prompts, not actual shopper purchase logs.
- [7]McKinsey · Oct 17, 2025. The agentic commerce opportunity. U.S. B2C $900B–$1T by 2030; global goods $3–5T. Includes AI-shaped purchases, not only agent-completed checkout. Excludes services and B2B. Not TAM for Picked.
- [8]Profound · Feb 24, 2026. Series C announcement. $96M Series C, $1B valuation, 700+ enterprise customers reported at that time. Historical competitor context, not a claim about Picked or an updated customer count.
- [9]Visibility incumbents. Profound, Peec, Scrunch, and Ahrefs Brand Radar. The comparison describes the visibility baseline. It does not claim these companies cannot build agent replay or already lack every proposed Picked capability.