From Answer Engines to Action Engines

Agentic optimization is the discipline of making a regulated product's legitimate decision context machine-legible enough for AI systems to classify, compare, constrain, and route it accurately. As the answer layer becomes an action layer, decision-frame fidelity must survive not only synthesis but inclusion, comparison, constraint, and routing — because a system can quote a product accurately and still leave it off the shortlist.

Part 2 of “Loss of Exclusivity Happens Twice Now.”

Executive premise

In Part 1, Signal Fidelity Group argued that loss of exclusivity now happens twice: first in law and commerce, then in the AI answer layer, where decades of molecule-level authority begin supporting every approved entrant.

The objective was decision-frame fidelity: ensuring the full, substantiated decision context survives AI synthesis, regardless of which legitimate sources are cited. That requirement now extends beyond the answer.

OpenAI introduced Instant Checkout and the Agentic Commerce Protocol in September 2025. In March 2026, it expanded ACP into product discovery, including richer product data, conversational refinement, and side-by-side comparison. OpenAI's commerce documentation describes ACP as the layer that lets ChatGPT ingest structured catalog data, understand merchant inventory, and surface relevant products in context.

The regulated-health implication is not that a general-purpose agent should prescribe a therapy, authorize a procedure, or purchase a prescription product. It should not. The implication is that software increasingly shapes the work between inquiry and decision:

retrieve → classify → compare → narrow → route → act

The human remains the principal. The decision pathway is becoming machine-mediated.

A system can quote a product accurately and still leave it off the shortlist.

The strategic choke point moved upstream

Checkout made agentic commerce visible. Product discovery revealed the deeper leverage. Before any transaction occurs, the system has already made consequential judgments:

• Which category does this product belong in?
• Which alternatives count as comparable?
• Which attributes and evidence matter to the user's question?
• Which constraints exclude an option?
• Which next step is appropriate and feasible?

Those judgments determine whether a product reaches the human decision-maker at all. For regulated products, the cost of a category error is higher than an irrelevant retail result. The system may conflate a drug and a device, collapse treatment and prevention, ignore an indication boundary, treat adjacent modalities as substitutes, or recommend a pathway that is unavailable in practice.

The transaction is downstream. The shortlist is the choke point.

A product feed is not a decision model

OpenAI's commerce documentation sets out feed and product specifications covering the fields a commerce system needs: stable identifiers, titles, descriptions, brands, categories, price, availability, seller context, variants, and eligibility flags. Those fields can make product information more current and usable. They cannot resolve an incoherent market position.

A regulated product must often be understood through relationships that do not fit cleanly into a retail catalog:

• regulatory class and approval status;
• specific population and approved use;
• clinical or operational job-to-be-done;
• evidence and label boundary;
• ordering, prescribing, administration, and interpretation requirements;
• coverage, payment, referral, capacity, and site constraints;
• substitute stack and non-product alternatives; and
• conditions under which the product is not appropriate.

If those relationships are absent or contradictory, the system will infer them from the wider public environment. It may inherit a competitor's frame, compare the asset to the wrong alternatives, or omit it.

The product feed is not the strategy. It is the serialization of a strategy.

Agentic optimization is therefore not a schema exercise. It is the discipline of making a product's legitimate decision context machine-legible enough for AI systems to classify, compare, constrain, and route it accurately.

Decision-frame fidelity at the action layer

Decision-frame fidelity was defined in Part 1 as the degree to which the complete, substantiated decision context survives AI synthesis. At the action layer, the test becomes stricter. The decision frame must survive four additional operations:

• Inclusion. Does the product enter the relevant option set?
• Comparison. Is it compared with the right alternatives on legitimate dimensions?
• Constraint. Are indications, evidence limits, access barriers, and workflow conditions preserved?
• Routing. Does the system suggest an appropriate, feasible next step?

An answer can be factually accurate at the sentence level and still fail at the decision level. That is why visibility, citation share, or schema completeness cannot be the only measures.

Six gates before the message

Before optimizing content, publishing structured product data, or evaluating an agent, a regulated organization should resolve six gates.

1. Classify

What exactly is the asset in regulatory, clinical, and commercial terms — and what is it not? The required output is a category-correct identity statement and a set of explicit exclusions. A vague category can create a precise error.

2. Define the decision

Which precise clinical, operational, or economic decision does it change? The required output is a bounded job-to-be-done with its audience and context. A product can be scientifically important and commercially peripheral if it does not change a decision someone can and will make.

3. Bound the evidence

What is approved, independently supported, company-asserted, inferred, unresolved, or unsupported? The required output is a claims wall and an evidence-state ledger. The boundary is not where strategy goes to die; it is what lets an organization remain ambitious without presenting tomorrow's possibility as today's product truth.

4. Follow the decision system

Who orders, pays, earns, uses, interprets, absorbs cost, or can veto adoption? The required output is a decision-owner, incentive, and access map. A representation that ignores coverage, capacity, referral patterns, administration burden, or site availability can be technically accurate and practically false.

5. Map substitutes and friction

What will users do instead, and where can workflow defeat adoption? The required output is a substitute stack alongside a capacity, referral, training, logistics, and access map. The named competitor is rarely the whole competitive set.

6. Position, encode, and test

What can the asset credibly own now, what can it earn next, and what would disprove the recommendation? The required output is a position, a falsifier, a semantic spine, and a measurement plan.

These gates are not a longer content brief. They are the minimum reasoning required before a product is made actionable.

The new keywords are relationships

Traditional search optimization trained organizations to manage nouns and phrases: brand, molecule, disease, category, indication. Agentic systems need verbs and relationships:

• is
• is indicated for
• helps decide
• requires
• is supported by
• differs from
• is available through
• should not be used for

Once the six gates are resolved, Communications can encode three stable semantic relationships.

Identity

[Asset] → is → [category-correct product truth]. Identity should preserve regulatory class, status, intended use, and meaningful exclusions.

Utility

[Asset] → helps [audience] decide → [specific choice in a bounded context]. Utility identifies the decision the product changes — not a generalized benefit.

Credibility

[Asset's role] → is supported by → [verifiable proof within its evidence boundary]. Credibility names the source and its limit rather than substituting reputation for evidence.

These relationships can become a common spine across product pages, clinical explainers, media materials, professional education, structured metadata, product feeds, expert commentary, and other trusted sources.

Worked example: why “Alzheimer's test” is not a useful category

FDA approved TAUKLARIFY (florquinitau F 18 injection) in August 2026 as a PET radiodiagnostic drug for brain imaging in adults with cognitive impairment being evaluated for Alzheimer's disease, to identify patients with tau neurofibrillary-tangle pathology. Safety and effectiveness have not been established for evaluating non-Alzheimer's tauopathies. (FDA; Lantheus)

FDA had previously cleared a blood-based in vitro diagnostic that aids in identifying amyloid pathology. It follows a device pathway, uses a blood draw and laboratory workflow, and is not intended for screening or stand-alone diagnosis. (FDA)

Both products relate to Alzheimer's disease. They are not interchangeable representations of an “Alzheimer's test”:

• Regulatory form. An FDA-approved PET radiodiagnostic drug, versus an FDA-cleared in vitro diagnostic device.
• Biological target. Tau NFT pathology, versus amyloid pathology.
• Workflow. Radiopharmacy, PET imaging, scanner time, and trained interpretation, versus a blood draw and laboratory analysis.
• Approved job. Identify tau NFT pathology in the labeled population, versus aid in identifying amyloid pathology in the specified population.
• Key boundary. Not established for non-Alzheimer's tauopathies, versus not screening or stand-alone diagnosis.

The 2025 updated appropriate-use criteria reinforce the importance of context. Among 17 scenarios assessed for tau PET, five were rated appropriate, six uncertain, and six rarely appropriate. (Rabinovici et al.) A system that sees only the noun “test” may compare the products on the wrong terms.

Context is part of the product truth.

The loss-of-exclusivity application

At loss of exclusivity, the answer and action layers can compress a choice faster than conventional communications systems can respond. When an originator and a generic share a molecule and an accepted equivalence conclusion, the default chain is easy to construct:

same molecule → accepted equivalence → available → lower price, where true → choose

That chain should not be defeated by disputing settled evidence or inventing unsupported superiority. The defensible opportunity is to make the larger, substantiated decision legible where it genuinely matters:

• diagnosis and subtype;
• approved use and patient fit;
• treatment versus prevention;
• formulation, delivery, administration, or service differences;
• initiation, recurrence, monitoring, and continuity;
• access, support, and supply;
• appropriate alternatives; and
• when the molecule is not the answer.

The originator may possess more evidence and infrastructure. But unassembled proof loses to complete decision logic. The objective is not more brand-only copy. It is a neutral, sourceable decision framework that lets both people and machines understand what equivalence answers — and what it does not.

A cross-functional operating model

Agent-ready communications cannot be owned by one function in isolation.

• Medical and Scientific. Evidence, clinical context, limitations, and unresolved questions.
• Regulatory and Legal. Approved claims, disclosures, and boundaries.
• Market Access. Coverage, payment, documentation, and economic constraints.
• Commercial. Business objective, priority segment, and desired behavior.
• Digital, Data, and IT. Feeds, schemas, APIs, identity, and technical implementation.
• Communications. Cross-source coherence, trusted-intermediary strategy, and externally verifiable meaning.
• Analytics. Observation of selection, claims, sources, variation, and action routing.

Communications does not own the underlying truth of every input. It owns whether those truths arrive as a coherent decision.

At the answer layer, ambiguity changes meaning. At the action layer, ambiguity changes behavior.

What to measure

A defensible measurement frame should capture more than brand presence.

• Question coverage. Which consequential questions and decision stages are represented?
• Selection. Does the product enter the relevant shortlist, and beside which alternatives?
• Classification. Is the asset placed in the correct category and decision layer?
• Claim fidelity. Which approved or evidence-supported statements survive synthesis?
• Source influence. Which domains and documents shape the answer?
• Boundary fidelity. Are indications, limitations, uncertainty, and conditions preserved?
• Action routing. What does the system suggest the person do next?
• Variation. How do results change by platform, context, date, and repeated run?

The objective is not to claim control over a model. It is to observe the decision environment, strengthen legitimate source inputs, and measure whether decision-frame fidelity improves.

Three actions for Communications leaders

Add actions to the fifteen questions

For every consequential question, document the downstream decision: choose, switch, delay, escalate, refer, monitor, or decline. You cannot evaluate an agentic answer without knowing what action sits downstream.

Find the verbs and attach proof

Write the Identity, Utility, and Credibility relationships for each priority product. Attach the source and the evidence boundary to every relationship. If different functions cannot agree on those relationships, a product feed will not solve the problem.

Audit the shortlist

Run the questions across multiple AI surfaces, contexts, and repeated observations. Measure not only whether the brand appears, but how it is classified, which alternatives are present, which claims and sources survive, and what next step is recommended.

Conclusion

Part 1 asked whether a brand's evidence survives into the answer. Part 2 asks whether the decision that evidence supports survives into the action.

A product feed can transmit fields. It cannot repair a category error, define a credible differentiation, or establish the evidence boundary. That work has to happen before the feed — and before the message.

Decision-frame fidelity has to survive all the way into the action.

Sources

• Signal Fidelity Group, “Loss of Exclusivity Happens Twice Now”, August 25, 2026.
• OpenAI, “Buy It in ChatGPT: Instant Checkout and the Agentic Commerce Protocol”, September 29, 2025.
• OpenAI, “Powering Product Discovery in ChatGPT”, March 24, 2026.
• OpenAI Developers, Agentic Commerce Protocol, accessed September 1, 2026.
• FDA, Novel Drug Approvals for 2026, TAUKLARIFY entry dated August 13, 2026.
• Lantheus, FDA Approval of TAUKLARIFY, August 14, 2026.
• FDA, First Blood Test Used in Diagnosing Alzheimer's Disease, May 16, 2025.
• Rabinovici et al., Updated Appropriate Use Criteria for Amyloid and Tau PET, 2025.