Brand Agent Conversation Design: 30 Flows That Convert

"How many conversation flows does my Brand Agent need?" is the wrong first question, and it's the one almost every merchant asks us. The right first question is: what shapes of questions do my shoppers actually ask, and which of those shapes decide whether they buy? Get that taxonomy right and the flow count takes care of itself — usually landing between 20 and 50 for a focused category, more for a sprawling multi-category catalog.

What a conversation flow actually is

A conversation flow isn't a scripted Q&A pair. That's the old chatbot model — one input, one canned output. A flow is a pattern: a shape of question, the reasoning the agent needs to answer it well, and the data it needs to pull from. Write the flow for the pattern, not the exact wording, and the agent generalizes to the hundred phrasings you didn't anticipate.

Conversation flow

A named pattern of shopper intent — e.g. 'compare two products on a specific attribute' — paired with the reasoning logic and catalog data the Brand Agent needs to answer any phrasing of that intent well, not a single scripted exchange.

This distinction is why flow count is a weak proxy for quality. Ten well-designed patterns that cover the real shape of your category's questions beat two hundred narrow scripts that each answer one exact phrasing and fail on the next one. We cover the mechanics of what a Brand Agent needs to reason well in Shopify Brand Agents, explained — this article is about which patterns to build and in what order.

The five categories

Across every catalog we've enriched and every agent we've tuned, shopper questions cluster into five categories. They aren't equally weighted — some sit right next to the buy button, others show up after the order ships — but a Brand Agent that's missing any one of them has a visible hole in coverage.

1. Comparisons

"What's the difference between these two?" is the single most common high-intent question in ecommerce, and it's the one legacy chatbots handled worst — they either dumped both spec sheets or punted to a category page. A Brand Agent should reason across the specific attributes a shopper cares about and give a direct answer.

• "What's the difference between the Ridge Pack and the Summit Pack?"
• "Which of your two air fryers is better for a family of four?"
• "Is the Pro version worth the extra cost over the standard one?"
• "You have three shades of this color — how do they actually differ in person?"

The reasoning behind every one of these is the same shape: pull the relevant attributes for both SKUs, weight them against what the shopper said they need, and state a recommendation with the tradeoff named out loud. That reasoning only works if the catalog has the attributes structured — see catalog enrichment for AI if comparisons are coming back vague or hedgy.

2. Constraints

Constraint questions stack multiple conditions into one ask. They're the questions a good in-store salesperson answers instantly and a search bar answers terribly, because search bars can't do multi-attribute filtering in natural language.

• "I need something vegan, under $40, that ships by Friday."
• "Do you have this in a size 12 that's also machine washable?"
• "What's your quietest blender that still fits under a cabinet?"
• "Something gift-wrappable, under $60, that isn't the same thing I bought last year."

Constraint flows fail in one of two ways: the agent finds a product matching some of the conditions and presents it as a full match, or it declines outright because no SKU satisfies every condition and it doesn't know how to say so gracefully. Both are trainable. The agent should state which constraints are met, which aren't, and offer the closest alternative rather than silently dropping a condition.

3. Suitability

Suitability questions ask whether a product category-level, not SKU-level: is this kind of thing right for my situation. They're trust questions, and category-specific — a skincare suitability question looks nothing like a cookware one.

• "Is this moisturizer okay for sensitive skin?"
• "Would this cast-iron pan work on an induction stovetop?"
• "Is this backpack big enough to carry on for a week-long trip?"
• "Is this toy appropriate for a 4-year-old?"

Suitability flows lean hardest on structured compatibility and use-case data — exactly the "works with / suitable for" fields that most catalogs leave thinnest. If your agent is hedging on suitability questions with "you may want to check with a professional," that's usually a data gap, not a model limitation.

4. Objections

Objection flows handle the moment a shopper is qualified, interested, and hesitating — usually on price, but sometimes on trust, risk, or timing. This is where a scripted chatbot goes silent or deflects to a policy page, and where a well-trained Brand Agent earns the sale a human salesperson would have closed.

• "Why is this $30 more than the other brand's version?"
• "What happens if it doesn't fit — can I actually return it easily?"
• "Is this worth it, or should I just wait for a sale?"
• "How is this different from the cheaper one on the same page?"

The agent should answer objections with value and specifics, not defensiveness — name the material, the warranty, the durability difference, whatever earns the price gap. It should never invent a justification the catalog doesn't support. If there genuinely isn't a good answer to "why does this cost more," that's useful signal for merchandising, not a prompt to bluff.

5. Post-purchase

The category merchants build last and regret not building first. Post-purchase flows run after checkout — inside order-status conversations, follow-up emails, or a returning-customer's next session — and they're where retention and upsell actually happen.

• "How do I care for the leather so it doesn't crack?"
• "What accessories work with the thing I just bought?"
• "Can I exchange this for a different size before it even ships?"
• "I loved my last order — what's similar but not identical?"

These flows compound: a shopper who gets a good care-instructions answer six weeks in comes back for the accessory recommendation without a marketing email prompting them. We go deeper on this in post-purchase Brand Agents.

How to prioritize which flows to build first

Don't start from the taxonomy and try to write five categories' worth of flows from imagination. Start from what shoppers already ask you, today, in channels you already have.

1. 1

   Pull your existing question data

   Support ticket subject lines, on-site search queries, review Q&A sections, and live chat transcripts if you have them. This is real shopper intent, not a guess at it.

2. 2

   Tag each question by category

   Sort into comparison, constraint, suitability, objection, post-purchase. You'll usually find the distribution is uneven — some categories generate far more of one type than another.

3. 3

   Rank by frequency and purchase proximity

   A constraint question asked 200 times a month right before checkout outranks a suitability question asked 10 times a month by a browser three steps from buying anything.

4. 4

   Build the top 20-30 as real flows first

   Full reasoning logic, tested against edge phrasings, checked against your actual catalog data. Ship these before touching the long tail.

5. 5

   Backfill the long tail on a cadence

   Add flows monthly based on decline-rate data — every question the agent couldn't answer well is a candidate for the next batch.

How many flows is enough

For a single, focused product category, 20 to 50 well-built flows across the five categories typically covers the overwhelming majority of real shopper intent — not because that's a magic number, but because that's roughly where the marginal question stops mattering to conversion. A multi-category catalog needs the same taxonomy repeated per category, since a suitability question in skincare and a suitability question in cookware share a pattern but not a single word of content.

Resist the urge to chase a round number. A catalog with thin compatibility data will hit diminishing returns at 15 flows because the underlying facts aren't there to reason over, no matter how many patterns you script. A catalog with rich, well-structured data can support 60+ flows productively because there's genuinely more to ask about. Flow count should follow catalog depth, not a target set in a planning meeting.

Writing a flow so it generalizes

The difference between a flow that generalizes and one that only answers its exact example: the flow definition should specify the reasoning steps and the data fields the agent pulls from, not a fixed answer string. Write it like a brief for a new salesperson, not a script.

• Name the pattern, not the example — "multi-condition constraint match" not "vegan under $40 by Friday."
• List the data fields the agent needs — material, price, ship-by date, availability — so gaps show up before launch, not in a live conversation.
• Specify the failure behavior — what the agent says when no SKU fully matches, explicitly, so it never improvises a bad answer.
• Give 3-5 phrasing variants as test cases, not as the only inputs the flow should handle.

This is also why flow design and catalog work can't be separated. A perfectly written constraint flow still fails if "vegan" isn't a structured field anywhere in your product data — the agent has nothing to reason over. If flows keep coming back thin or hedgy in testing, check product attribute schema design before rewriting the flow itself.

Common conversation-design mistakes

• Writing scripts instead of patterns. A flow tied to one exact phrasing breaks the moment a shopper asks it differently, which is most of the time.
• Skipping objections. Merchants build comparisons and constraints, then leave price and trust objections to a static FAQ page — the exact moment a human closer would step in.
• Ignoring post-purchase entirely. Post-purchase flows are retention infrastructure, not a nice-to-have; skipping them leaves lifecycle value on the table.
• Guessing at priority instead of measuring it. Building flows in the order they occur to you, rather than in the order your own shopper data says they matter.
• Treating flow count as the success metric. A hundred shallow flows covering a narrow slice of real intent loses to thirty deep ones that match how shoppers actually talk.

How this connects to launch

Conversation flows are one of four inputs a Brand Agent needs to work well, alongside a complete catalog, your brand voice, and your policies — see the full breakdown in Shopify Brand Agents, explained. None of the five flow categories in this article do much good without the catalog underneath them; run the Agentic Commerce Readiness Score first if you're not sure whether your data can support them yet.