Schedule III Won't Save Your AI Visibility

The headline is tempting because it promises a clean before and after. Cannabis moves to Schedule III, the category becomes normal, and AI platforms finally start recommending dispensaries. That is not the state of the law or the state of search. The useful story is narrower and more difficult: federal movement does not remove the visibility systems that make cannabis hard to retrieve.

May 2024 was a proposed rule, not a completed rescheduling of marijuana generally. In April 2026, the Department of Justice issued a final rule placing FDA-approved products containing marijuana, and certain state-licensed medical products, in Schedule III. The DOJ rule is narrower than the shorthand most marketing summaries use. It does not turn every adult-use product, dispensary, or cannabis claim into an ordinary national category.

The broader proposal is a separate proceeding. The DEA hearing notice says the formal hearing on proposed marijuana rescheduling began June 29, 2026, after an earlier proceeding was terminated and a new notice was issued. The current DEA NPRM docket still lists hearing transcripts and post-hearing orders. This piece therefore treats whole-plant rescheduling as a live administrative question, not a settled marketing fact.

That correction improves the business question. A new federal status may change what investors, banks, regulators, and platforms are willing to do. It does not create a local knowledge graph, resolve state-by-state rules, or require an answer engine to recommend one dispensary over another. The legal announcement changes a boundary. The visibility work still has to be done.

The most useful evidence is not a dramatic claim that ChatGPT hates cannabis. It is the difference between surfaces. SOCi's 2026 Local Visibility Index, as covered by Cannabis Industry Journal, examined more than 350,000 business locations and 2,751 brands. Its finding was blunt: dispensaries were winning on Google while disappearing on ChatGPT.

That is not a contradiction. Traditional search can return a location because the user asks for a place, the business has a profile, and the query maps to a local result. An answer engine has a different burden. It needs enough trustworthy context to explain what the business sells, where it operates, whether the result is lawful for the user, and whether a recommendation would cross a policy or health-claim boundary.

The handoff cites the 2026 MG Magazine Cannabis Brand Visibility Index for a second signal: approximately 28 percent of cannabis prompts produced a refusal, hedge, or prominent disclaimer, the highest rate among the consumer categories it compared. The underlying coverage does not describe a normal ranking problem. It describes a category where the answer layer often narrows, qualifies, or declines before a brand can compete.

The marketing implication is uncomfortable. A dispensary can improve its Google listing and still remain absent from the conversation that starts with “what should I buy near me?” It can earn impressions on a conventional results page while the assistant gives a generic safety answer. Treating that as a traffic problem alone misses the category's actual failure: the system lacks enough usable, low-risk context to include the business.

Measure the gap by separating presence from usefulness. Record whether the business appears, whether the answer describes it accurately, whether the location is correct, whether the recommendation includes a safe next step, and whether the answer cites the business or only the category. A binary visibility score hides the difference between a correct local answer and a vague mention that cannot produce a visit.

Platform caution is not a single switch that a brand can negotiate away. Cannabis combines several sensitive variables: age restrictions, health and efficacy claims, local legality, product compliance, advertising rules, and the risk that a general answer sounds like medical advice. When those variables are uncertain, a model has a rational incentive to say less.

Policy inertia adds a second layer. A platform may update its legal assumptions slowly across countries, states, products, and model versions. The result is a category that is lawful in one context but still treated as a high-risk generic prompt. Rescheduling can move one federal input while the platform's safety and retrieval systems continue to use older or broader boundaries.

Low category priority makes the problem worse. A restaurant can expose a menu, price, address, and reservation path in predictable fields. A dispensary may need to explain inventory that changes daily, age or medical eligibility, delivery boundaries, product form, testing information, and claims it must not make. If those facts are scattered or stale, an answer system has fewer safe sentences available.

The answer is not to push harder on promotional language. In a constrained category, more persuasion can reduce the evidence quality. A claim such as “best for anxiety” creates a larger policy and substantiation problem than a precise description of product form, lab information, legal market, and store service. The visibility task is to make safe, useful facts easy to retrieve.

This is why the first useful optimization is usually a content boundary, not a keyword list. Write the answer a cautious system can repeat without changing its meaning. Say where the store operates, what kind of products it carries, what the customer must bring, and where current availability lives. Avoid asking a general-purpose assistant to infer legality, dosage, or medical suitability from a sales slogan. The safer sentence is often the more discoverable sentence.

Suppression does not create a blank answer. It creates a narrower answer, and the narrow answer tends to reuse the entities with the most consistent public evidence. That is why concentration matters more than an average visibility score. If the category is already cautious, the brands that are easiest to identify become the safest defaults.

The 5WPR Cannabis AI Visibility Index cited in the handoff found Curaleaf, Trulieve, and Green Thumb Industries together accounted for approximately 17.5 percent of all cannabis-category AI citations. The figure is not a claim that those operators own the category. It is a signal that a few entities can become the reference layer when many smaller operators publish inconsistent or thin information.

This is a discoverability moat, not necessarily a product moat. A multi-state operator has more locations, more structured pages, more press coverage, and more opportunities to be described by third parties. A local dispensary may offer a better experience and still lose the machine's shortlist because its most important facts live in an inventory widget, a social post, or a directory that the model cannot confidently reconcile.

The practical response is not to imitate a large operator's tone. It is to close the evidence gap at the local level. Make the entity, location, service area, product categories, age rules, hours, and update dates explicit. Then publish useful explanations that a customer can understand without a promotional claim. Accuracy and consistency are the scarce inputs.

A local operator can also publish the context that a national directory cannot: which neighborhoods it serves, how medical and adult-use menus differ where applicable, how curbside or delivery boundaries work, and how the store handles products that are temporarily unavailable. These details are not filler. They give a retrieval system a reason to distinguish one legitimate business from the next generic category page.

Make the update cadence visible too. A page that says “open now” without a time or market context is weaker than a page that tells the reader when its store and menu information was last checked. Freshness does not guarantee inclusion, but it gives both the customer and the answer system a better way to judge whether the fact is still useful.

Begin with a machine-readable local record. Keep the business name, address, phone, hours, holiday changes, service area, license context, and store status consistent across the website and major local sources. If the answer engine cannot resolve which location is open and what it offers, no amount of brand language will rescue the recommendation.

Separate the product facts from the claims. A page can explain form, potency ranges, testing, ingredients, intended use as permitted, storage, and purchase constraints without promising a medical outcome. That separation gives the retrieval system useful nouns and qualifiers while reducing the chance that a generic prompt turns into an unsupported health answer.

Build a small question set and test it every month. Use the same prompts across Google, ChatGPT, and other answer surfaces. Record the exact wording, market, model or surface, cited sources, refusal language, and whether the store was named. The point is not to chase a single answer. It is to see whether the same business facts survive changes in prompt, location, and platform.

Treat compliance language as product infrastructure. If a product description requires a jurisdictional qualifier, put the qualifier next to the fact. If inventory changes daily, expose an update time. If a recommendation cannot be made safely, offer a neutral path to verified information. A disclaimer hidden in a footer is less useful than a clear boundary attached to the claim it qualifies.

Close the loop with a source record. Save the prompt, answer, cited page, timestamp, market, and outcome. If the answer changes after an inventory update or a policy change, the team should be able to tell whether the cause was missing data, stale data, a refusal boundary, or a competitor becoming the default source. That record turns AI visibility from a vague impression into a local operating signal.

Review those records with the same seriousness as local-search conversion data. A refusal is a product signal, not a reason to flood the web with more promotional pages. An incorrect citation is a source-quality problem, not proof that every platform should be abandoned. The discipline is to identify the exact boundary and improve the evidence that belongs on the safe side of it.

There is a natural connection to the site's broader visibility work. GEO is not SEO with a new name because the answer surface evaluates more than rank. And the regulated-market lesson from personalization liability applies here too: compliance can be a product feature when it makes the system safer to use and easier to trust.

A final whole-plant rescheduling decision, if it arrives, could change the category's economics and the confidence with which businesses, financial institutions, and platforms treat the market. It would not make state law uniform. It would not erase age or health-claim constraints. It would not automatically cause an answer engine to select a local store whose public evidence is thin.

That is why the right operating posture is two-track. Watch the legal process and date-stamp every claim about it. At the same time, improve the evidence a platform can use today: accurate local identity, safe product language, useful question-led pages, and a prompt record that shows what the category is actually being asked.

This two-track approach also keeps the business from making a dangerous timing bet. If whole-plant rescheduling is delayed, the visibility program still improves. If the rule changes tomorrow, the business already has the location data, product boundaries, and measurement record needed to understand what actually moved. Regulatory news should update the operating plan, not replace it.

The category does not need another promise that regulation will solve distribution. It needs an explanation of what changed, what did not, and what a business can measure while the boundary moves. Schedule III may open a door. Visibility still depends on whether the system can find a credible business on the other side.

The durable advantage will belong to operators that can answer both questions at once: “Are we allowed to do this here?” and “Can a customer or machine find the accurate answer?” Legal status is one input. Evidence quality is the operating system around it.