Catalog Visibility Agent Workflow

Catalog Visibility Agent Workflow

A buyer question can contain far more information than a short product title can hold. Each Reach Dog selling guide supplies six curated questions across discovery, evaluation, and purchase friction. The catalog visibility agent workflow checks all six against one verified product record and creates four useful drafts for every supported connection without stretching the product facts.

This is currently a do-it-yourself workflow for an LLM. It is not yet a one-click Reach Dog catalog operation, and Reach Dog does not publish or schedule the resulting content. Each selling guide provides free category-specific instructions that a merchant can copy into an LLM and run with their own catalog facts. The guide and agent blueprint are free; model usage depends on the LLM provider the merchant chooses.

Most merchants can stop there. Technical teams that want to connect the workflow to a product feed can use the Python-friendly build and deployment reference. It explains the input contract, tools, state, tests, review queue, and human approval gate without presenting autonomous publishing as a current Reach Dog feature.

What the workflow does

The workflow takes six conversational buyer questions through five stages:

1. Keep the questions separate. Preserve all six buyer needs, phases, required facts, and claim boundaries.
2. Check product relevance. Compare each question with verified catalog facts before writing anything.
3. Choose one supported angle per question. Do not merge separate buyer decisions into one generic theme.
4. Draft four assets per supported question. Create a product-page or FAQ answer, Facebook or Instagram post, short-video script, and X post.
5. Review and hand back. Mark every draft Publish, Verify, Revise, or Reject and return final approval to the merchant.

The maximum is 24 reviewable drafts when all six questions are supported. The purpose is not to force every product into every question. A good result may support two questions, hold three for verification, and reject one. Partially supported and rejected questions produce no marketing drafts.

Start with evidence, not an empty prompt

Give the workflow a small approved evidence packet:

• the six buyer questions supplied by the category guide;
• product IDs and current titles;
• structured product attributes and specifications;
• approved product descriptions and claim language;
• testing, certification, compatibility, care, shipping, and policy details that apply to the category;
• the dated Reach Dog guide and its buyer-language evidence;
• brand voice and channel limits.

Do not ask the LLM to fill missing facts from general knowledge. If a product fact is missing, the workflow should label it MERCHANT VERIFICATION NEEDED and avoid using it as a claim.

One question, five dimensions

Use the five dimensions as a reading framework, not as five mandatory marketing claims.

Some questions will reveal two dimensions. Others may reveal all five. The workflow should state what is explicit, what is reasonably implied, and what remains unknown.

Check relevance before creating content

For each candidate product, require one of three decisions:

• Supported: verified product facts answer the question directly.
• Partially supported: some facts fit, but a specific fact must be verified before publication.
• Reject: the product facts do not support the connection.

This step keeps the workflow commercially useful. It prevents the LLM from creating attractive content for a product that cannot deliver what the buyer asked for.

Create one angle per supported question

The six-question set contains exactly two questions from each buyer phase:

• Discovery / context: situations, occasions, environments, and early questions.
• Evaluation / specification: materials, fit, dimensions, compatibility, construction, and comparisons supported by facts.
• Friction / objection: setup, care, delivery, returns, uncertainty, or another barrier to purchase.

An angle is not distinct merely because the opening sentence changed. Each supported question should answer a different buyer decision and use the product facts relevant to that decision.

After completing the six-question pass, a merchant may select the strongest supported question and ask for 5–10 additional angles or a ten-post campaign. That is an expansion pass, not a substitute for checking all six questions first.

Match the content to the channel

Every supported question produces four different assets:

• Product page or FAQ: a direct answer followed by the exact supporting specification.
• Facebook or Instagram: a relatable situation, one useful answer, and a clear next step.
• Short video: a 30-to-45-second spoken script with a question-led opening, concise explanation, suggested shots, and factual close.
• X: a compact observation or question with one supported takeaway.

The workflow should not copy one paragraph into four formats. It should preserve the evidence while changing the structure, pacing, and level of detail for the channel. Every asset must carry a separate facts-used trace.

Required output

Ask for a structured result containing:

1. all six buyer questions and their buyer phases;
2. the five-dimension breakdown for each question;
3. assumptions and unknowns;
4. the product reviewed;
5. Supported, Partially supported, or Reject for each question;
6. supporting catalog facts for each decision;
7. missing evidence;
8. one distinct angle for every Supported question;
9. four channel drafts for every Supported question;
10. the exact product facts used by each draft;
11. claim review;
12. Publish, Verify, Revise, or Reject status;
13. merchant approval notes.

Stop conditions

The workflow must stop instead of drafting when:

• no product is supported by the supplied facts;
• the buyer question depends on a missing specification, test, certification, compatibility statement, or policy;
• the requested wording would create a health, safety, performance, environmental, or regulatory claim the evidence does not support;
• the product category, buyer question, or catalog record is too ambiguous to identify the correct item.

Merchant approval remains the last step

An LLM can organize evidence and accelerate drafting. It cannot know whether an unpublished specification is correct, whether a claim has been approved, or whether a draft matches a merchant’s current inventory and policies.

Before publishing, confirm the product match, supporting fact, claim wording, offer, price, availability, and channel requirements. Then approve, revise, or reject each asset.

For a shorter three-stage prompt workflow, see Using Selling Guides With AI. For how Reach Dog currently analyzes evidence, see Using the Reach Agent.