“AI cannot make a listing index. It can compile what humans already know about indexing into the back-end fields. That is exactly its job in the catalog ai workflow.”
The catalog ai workflow we run on every new listing launch, and on every catalog rebuild for a brand we onboard, is eight steps long. Three steps are pure human work. Two steps are rule-engine deterministic. Two steps are AI compile. One step is human verification after the listing is live. The order of the steps is the work. Reorder them and the listing does not index. That is the predictable failure mode of every ai-only Amazon agency model.
What follows is the catalog ai workflow step-by-step. Each step has its tool, its human role, and the role AI plays, or, more often, the role AI is not allowed to play.
Step 1 (human), Cerebro keyword pull on the actual category
The catalog strategist opens Helium 10’s Cerebro. They pull keyword data on the brand’s ASIN. They also pull on the top three to five competing ASINs in the same category root. Cerebro returns search-frequency rank, sponsored rank, organic rank, and historical depth for each keyword the ASIN ranks against.
The strategist does not feed the raw Cerebro export to AI. The raw export is roughly 800 to 4,000 rows. Most of those rows are noise. Branded competitor terms. Irrelevant cross-category drift. Misspellings. Accidentally indexed terms. The strategist filters Cerebro’s output down to the 80 to 200 rows that represent the actual target keyword set for the listing. This filtering is judgment work. It depends on the brand’s positioning, the category’s competitive shape, and which long-tail terms are worth chasing. AI cannot make this call without context the brand has not provided.
Step 2 (human), SmartScout organic-rank context
The strategist pulls SmartScout’s organic-rank report. They pull it for the brand’s existing listings where applicable. They also pull it for the top three competitors. SmartScout gives the temporal dimension Cerebro misses. Which keywords each ASIN has gained or lost rank on over the trailing 90 days.
This is how the strategist identifies the keywords the brand is currently bleeding rank on. That is the defense priority. And the keywords the brand has never targeted but where competitors are weak. That is the offense priority. This step is also human-only. The defense-versus-offense allocation is a strategy call. AI cannot make it without the brand’s positioning, margin tolerance, and inventory readiness.
Step 3 (rule engine), keyword tier assignment
The combined keyword set from Steps 1 and 2 hits the keyword_rules module in the ClearSight portal. The rule engine bins each keyword into three tiers based on deterministic criteria. The strategist team agreed on these thresholds in advance.
- Tier 1 (title-rank target), search-frequency-rank above category threshold, competitive density below threshold, direct customer-intent match. Usually 5 to 12 keywords per ASIN.
- Tier 2 (bullet-rank target), secondary intent, supporting attributes, qualifier terms. 15 to 40 keywords per ASIN.
- Tier 3 (Search Term Field + back-end attribute target), long-tail, low-volume, high-specificity terms. 60 to 150 keywords per ASIN.
This is deterministic. The same keyword with the same Cerebro stats gets the same tier assignment every time. The strategist can override manually. Sometimes does. But the rule-engine output is the baseline of the catalog ai workflow.
Step 4 (human), title and primary-bullet copywriting
The strategist writes the title and the first two bullets. They hit the Tier 1 keywords with natural-reading copy. This is the most expensive part of the catalog ai workflow on a per-hour basis. Also the most important. Title and primary-bullet copy converts the click into the purchase. It is the single largest predictor of behavioral signal alignment. That is the fourth indexing factor we wrote about in our AI listing-builder SaaS post.
The strategist does not generate this copy with AI. The strategist writes it. AI is allowed to review the draft for Tier 1 keyword coverage. It flags missing terms. But the writing itself is human work. Voice and conversion-rate optimization depend on judgment the LLM cannot replicate.
Step 5 (human), parent-child variation architecture
The strategist structures the parent-child variation tree. The goal is to allocate Tier 2 keywords across child ASINs. Not letting every child compete for the same Tier 1 terms as the parent. This is the architecture-design step.
It requires understanding which attribute the brand’s customers actually decide between. Size. Pack count. Flavor. Color. It also requires understanding how Amazon’s variation algorithm currently handles each family of attributes in the category. AI does not see the variation family. AI sees ASINs as isolated entities. This step is invisible to any ai-only Amazon agency model. That is why their listings consistently break parent-child structure or split indexing signal between siblings.
Step 6 (AI compile), Search Term Field and back-end attribute population
This is the first place AI does material work in the catalog ai workflow. The Tier 3 long-tail keyword set goes to Claude with three instructions. Compose the 250-byte Search Term Field to maximize unique keyword surface. No duplicate words. No stop words. No special characters. Map each Tier 3 keyword to the back-end attribute field where it is most likely to register. Sweetener. Age range. Pack count. Flavor. Flag any keyword that has no clean back-end home for human review.
The strategist reviews the AI’s output. The current edit rate at this step runs about 30 to 50%. Roughly one in three to one in two of the AI’s mappings get a manual override before submission. The most common edit is reassigning a keyword to a different back-end attribute the AI got wrong. Usually because the attribute field’s documentation has changed since the model’s training cutoff.
Step 7 (AI compile), first-draft A+ content modules
AI produces the first draft of the A+ content modules. Module text. Alt-text for module images. Comparison-chart copy. This output goes to the brand’s creative lead, not the brand directly. The creative lead redirects the draft against the brand voice, the visual hierarchy, and the design-system constraints. The redirected version goes to the designer for build. The designer’s output goes back through brand approval before the A+ content ships to Seller Central.
The AI compile here saves the creative lead the four to six hours it would otherwise take to produce a first draft. It is not replacing the creative lead. The branch where AI ships A+ content directly to Seller Central, common in ai-only Amazon agency models, is exactly the path that produces the off-voice, ungrouped, Cosmo-misaligned A+ content we keep auditing on inbound brands.
Step 8 (human verification, 14 days post-launch), Index Checker
Fourteen days after the listing is live, the strategist runs Helium 10’s Index Checker. They run it against the full keyword set from Steps 1 through 3. The output shows which keywords the listing is actually indexing for. Anything in Tier 1 that is not indexing gets diagnosed. Usually back-end attribute conflict. Or character-count truncation. Or a Cosmo-decomposition mismatch. Then re-shipped.
The current internal target is that 90% or more of Tier 1 keywords are indexing within 14 days of launch. Across the trailing 12 months of catalog launches on the book, the actual rate has been 87%. The gap (the 13% that need a second pass) is almost entirely back-end attribute conflicts. The model could not predict them at compile time. The strategist diagnoses them manually.
Where AI is used and where it is not in the catalog ai workflow
The catalog ai workflow uses AI in exactly two places out of eight steps. Step 6 (Search Term Field + back-end attribute compilation). And Step 7 (A+ content first draft). Both are downstream of the strategic decisions in Steps 1 through 5. Both are followed by human edit before the work ships.
AI saves the strategist roughly four to seven hours per launch on the work it does well. Compiling structured outputs from already-decided inputs. AI does zero hours of the keyword strategy. Zero of the parent-child architecture. Zero of the title copy. Zero of the indexing diagnosis.
This is the structural difference from the AI listing-builder SaaS model. The SaaS collapses all eight steps into a single LLM call. The catalog ai workflow runs the eight steps in order. The LLM enters at Step 6, not Step 1. The six-attribute completeness floor we hit before launching ads is the upstream version of this discipline.
Results across the catalog book
The numbers we track on the catalog work, trailing 12 months across the active client book:
- Tier 1 indexing rate at 14 days post-launch: 87% across all new ASIN launches
- Time to remediate failed indexing terms: median 6 days from diagnosis to re-shipped fix
- Net-new converting search terms surfaced per account per month: 22 to 60 depending on category and catalog size
- Strategist edit rate on AI’s Search Term Field + attribute compile output: ~38%
- Parent-child architecture errors at launch: trending near zero. The architecture is human-designed before AI sees the keyword set.
The numbers themselves are not what matters. What matters is the workflow shape that produces them. A sequence in which AI is downstream of human judgment. Not the other way around. The stack-reference post earlier in this series walks through the broader architecture. This post is the catalog ai workflow specific instance.
The next post in this series moves to the ads workflow. The same shape applies to bid logic, search-term harvesting, and campaign-structure work. With the same constraint that AI never owns the decision.
Reviewed by the SEO Team.
The Hybrid Stack, 10-post series. You are reading post 6 of 10.
Black-hat track, three archetypes of the ai-only Amazon agency:
- The three black-hat shapes of the 2026 Amazon agency
- Why offshore VA + ChatGPT shops are the most expensive cheap option
- Why AI listing-builder SaaS can’t get a 7-figure brand to actually index
- Why AI-first consultants who never log into Seller Central miss the work that moves money
White-hat track, the ClearSight hybrid stack:
- The ClearSight intelligence layer, stack reference
- Catalog AI/human workflow (you are here)
- Ads AI/human workflow
- Creative AI/human workflow
- Inventory AI/human workflow
- 12 months of the hybrid stack, results recap
← Previous: The ClearSight intelligence layer, stack reference | Next: Ads AI/human workflow →