A customer asks ChatGPT for a suitable product, hotel, restaurant or supplier. Your leadership team sees the conversation and asks a reasonable question: how do we get our business into that answer?
That question often produces an expensive category error. Being found in AI search, supplying a current product catalogue and giving ChatGPT access to a proprietary customer service are three different jobs. A website can solve the first. A product feed can improve the second. An app or plugin may be needed for the third.
The app is not the “advanced” answer by default. The right answer is the least complex route that can deliver the required customer outcome and produce evidence that it did.
Answer in 60 seconds
Use the public website route when the customer needs trustworthy public information and a useful destination. The operating problem is crawlability, retrievability, evidence and the quality of the landing journey.
Use a product feed when a merchant needs to supply structured, current catalogue data such as variants, price and availability. Feed acceptance can improve control over the product record; it does not guarantee that a product will be displayed or sold.
Consider a proprietary ChatGPT app when the customer needs a capability the public web or standard feed cannot provide: live or private data, account context, business-specific decision logic, a richer comparison interface or a controlled action.
Some businesses need more than one route. They should still be funded and measured separately. A crawler visit is not a citation; a citation is not a qualified visit; a feed acceptance is not a sale; an installed app is not proof of discovery or incremental demand.
In this article
- Start with the customer job
- The three routes are different contracts
- Route 1: improve public evidence and the destination
- Route 2: supply catalogue truth through a feed
- Route 3: build a proprietary customer capability
- The Route Decision Sheet
- Examples where the answer changes
1. Start with the customer job
“Presence in ChatGPT” is too vague to scope. Replace it with a customer job and a finish line.
Compare these four requests:
- “What does this company do, and is it suitable for a project in France?”
- “Show me a waterproof jacket in my size that can arrive by Friday.”
- “Which of my existing bookings can I move to next week?”
- “Prepare a compliant quotation using our account terms, but let me approve it before submission.”
The first request can often be served by well-structured public evidence and a good website destination. The second needs current product truth; a merchant feed may be the appropriate route. The third needs identity and private account data. The fourth adds business rules and a consequential action.
Technology should follow that difference. If the team starts by commissioning “a ChatGPT agent”, it can end up building authentication, interfaces and operational controls for a problem that a public page or feed already solves.
2. The three routes are different contracts
| Route | What the business supplies | What the customer can receive | Evidence that matters first |
|---|---|---|---|
| Public website and AI search | Crawlable public pages, coherent facts, evidence and useful destinations | Information, comparison context, citation and a link | Retrieval and citation across a controlled prompt set; qualified visits to the right page |
| Product feed | Structured catalogue records with stable identifiers, price, availability, variants and images | More current product discovery and a merchant-owned purchase handoff | Feed acceptance, eligible records and accuracy against the merchant system |
| Proprietary app or plugin | Controlled tools connected to live/private systems, optional UI, identity and action rules | Account-aware answers, structured decisions or real actions | Correct invocation, accurate state, completed valid actions and safe handling of exceptions |
These contracts can overlap. A retailer may maintain public buying guides, submit a product feed and add an app for compatibility checks using customer-owned equipment. A service company may need only public information and a strong enquiry path. A membership business may have little catalogue need but a clear account-servicing use case.
OpenAI currently separates these surfaces in its own documentation. It says any public site can appear in ChatGPT search, with OAI-SearchBot access helping content to be discovered and cited. Its merchant guidance says most merchants start with product feeds, while deeper apps are optional for businesses that need more control. Its developer guidance describes the additional value of an app as helping the model “know, do or show” something it could not deliver as well on its own. Publisher and developer FAQ, merchant guidance, app product guidance.
That documentation establishes available routes. It does not tell an individual business which investment will pay back.
3. Route 1: improve public evidence and the destination
Choose this route when the customer mainly needs public information: what the business offers, where it operates, which option fits, what policies apply and where to continue.
The work is broader than allowing a crawler. The answer has to be supported by pages that make the relevant facts clear, consistent and easy to retrieve. The cited destination then needs to help the visitor continue instead of repeating a generic brand message.
Use this proof ladder:
- Access: can the relevant crawler reach the canonical page and its important content?
- Retrieval: does a controlled test retrieve the intended page for the relevant question?
- Answer and citation: does the answer represent the fact accurately and cite the appropriate source?
- Qualified visit: does the visitor arrive on a page that matches the question?
- Business outcome: can the team connect that visit to a useful next action in its own analytics or CRM?
Each step is evidence for the next investigation, not proof of the step after it. Server logs can show that a crawler arrived; they cannot show that the business is cited. A citation can be verified in a test; it cannot prove representative customer exposure. An AI referral can be attributed; it cannot by itself establish that the sale was incremental.
A 2026 peer-reviewed study of 973 ecommerce sites found that ChatGPT referrals were measurable but still under 0.2% of sessions in its August 2024 to July 2025 sample. The same study found relatively stronger traffic shares and commercial outcomes in more complex product categories. It is descriptive, uses last-click attribution and covers an earlier period, so it supports monitoring and a complexity hypothesis rather than a current market-size forecast. Marketing Science study.
If the missing piece is the citation itself rather than the destination, our AI Visibility Audit measures where the business is retrieved and cited across a controlled prompt set before any content work starts.
4. Route 2: supply catalogue truth through a feed
Choose a feed when the decision depends on product records that change: price, availability, colour, size, seller, images, promotions or other catalogue attributes.
This route gives the merchant more control over the structured record than relying on crawling alone. OpenAI’s current product feed specification requires a defined set of core fields and separates product discovery from checkout enablement. Its merchant page also says shopping is currently limited to the US and that direct applicants join a waitlist; eligibility must be checked for the target business and market. Product feed specification, merchant guidance.
Use a different proof ladder:
- Delivery: did the platform receive and process the file or API updates?
- Acceptance: were the records accepted without material validation errors?
- Eligibility: which products are eligible for the intended surface and market?
- Commercial accuracy: do title, variant, price, stock and destination match the merchant’s authoritative system?
- Display: can eligible products be observed for controlled relevant requests?
- Referral and purchase: did customers continue to the merchant and complete a valid order?
An accepted feed does not promise display. Display does not establish traffic at useful scale. An attributed order does not prove a new order. Keep those states separate in dashboards, supplier statements and board reporting.
Our catalogue-readiness test examines the detailed truth chain from source data to customer handoff. Run that work before buying a conversational layer that would only make unreliable catalogue data easier to ask about.
5. Route 3: build a proprietary customer capability
An app becomes relevant when the business can name something important that ChatGPT cannot reliably know, do or show from public pages and a standard feed.
Strong candidates include:
- live capacity across several locations;
- a customer’s own bookings, orders, entitlements or approved prices;
- compatibility logic based on a product already owned;
- a structured shortlist with business-specific constraints;
- an action such as rescheduling, reserving, requesting a quote or preparing an order;
- a confirmation interface where the consequences need to be explicit.
The app is a controlled integration with the systems that hold those facts and perform those actions. OpenAI’s current architecture supports remote MCP tools and optional interface components. Public distribution adds identity verification, production endpoints, policy and listing assets, test cases and a review process. Approval and publication do not promise prominent discovery. Plugin architecture, review requirements.
This route should pass three tests before budget is approved:
Capability test
Name the one customer job that improves because the business can supply live/private data, a controlled action or a better decision interface. “Brand presence” does not pass.
Operating test
Identify the authoritative system, access rules, confirmation point, failure path, support owner and audit record. A polished chat response cannot repair an ambiguous price or an unreliable booking API.
Distribution test
Explain how a relevant customer will encounter, connect or invoke the capability without assuming featured placement. A direct link, an existing customer journey or a named-brand use case can be tested. Total ChatGPT audience is not a defensible acquisition forecast for one app.
The next three parts of this guide take that route further: when the business case holds, how to compare quotes, and what a release has to prove before a customer action goes live.
6. The Route Decision Sheet
Use this sheet before requesting proposals. Complete one row for one customer journey; do not average several unrelated ideas into a high “readiness score”.
| Decision field | Your evidence | Route implication |
|---|---|---|
| Customer’s exact job | The request in the customer’s own language and the intended finish line | Public information favours the website; product comparison may need a feed; private state or action may need an app |
| Required data | Public or private; static or changing; source system and acceptable age | Public stable facts can be published; changing catalogue data favours a feed; private/live state needs controlled tools |
| Required action | Read, compare, draft, reserve, submit, pay, amend or cancel | Consequential writes increase app scope, confirmation and operating controls |
| Interface need | Plain answer, cards, comparison table, form or confirmation state | UI should be added only when it materially improves the decision or prevents error |
| Market and eligibility | Target country, account type, product/service category and current platform rules | A documented route in one market cannot be assumed elsewhere |
| Existing route | Website, marketplace, ecommerce platform, booking provider or owned assistant | Reuse it when it already delivers the job with acceptable data, control and cost |
| Minimum proof | Observable evidence from the appropriate ladder above | Set the pilot and reporting contract before build work starts |
| Non-proof | What the evidence cannot establish | Prevents acceptance, attribution or usage from being presented as incremental profit |
| Recheck date | Owner and date for volatile platform rules | Avoids designing a live service around stale documentation |
The output should be a route choice with conditions: improve the public path, prepare a feed, build a bounded app, combine specific layers or defer until the underlying data and customer journey are ready.
7. Examples where the answer changes
A multi-location operator may need public pages for discovery and a proprietary tool for live capacity across locations. A single-location business may receive more value from accurate public information and an existing booking provider than from its own app.
A retailer selling standard products may be well served by catalogue data and merchant checkout. A technical supplier with account-specific pricing and compatibility rules may justify authenticated tools, provided the final quote is validated by the commercial system.
A service marketplace may benefit from structured qualification before a quote, while a regulated adviser should keep consequential recommendations and commitments behind explicit controls and human review.
The sector label never decides the architecture. The customer job, data, consequence and proof do.
Buy the smallest sufficient route
AI chat search is becoming another practical place where customers may begin. A business should make itself useful there in proportion to the job customers need to complete. That can mean better public evidence, a cleaner catalogue feed, a proprietary app or a deliberate combination.
Do not fund the largest option because it sounds more strategic. Fund the smallest route that can produce the customer outcome, then require evidence appropriate to that route.
Bring one customer journey, its authoritative data source and the route you currently use. We can complete the Route Decision Sheet, identify the missing evidence and define whether the next investment is content, catalogue infrastructure, a bounded integration or no build yet. When the answer is a bounded integration, that is the scope of our AI Agents & LLM Products service.
Sources and scope
Platform pages were checked on 9 September 2026. OpenAI’s naming, market eligibility, shopping delivery and public review rules can change, so they must be rechecked before implementation or publication. The business examples are illustrative and do not describe an IZZY client. The route and proof ladders are IZZY’s decision framework; they are not a platform guarantee or a validated ROI model.