The Listing Auditor
A pipeline that audits a brand's live Amazon listing and writes the email about what it found.
- Timeline
- 5 months · 5 campaigns
- Role
- Designed and built the qualification and audit pipeline, with an Outlyst team
- Status
- Client engagement
- Primary stack
- Clay · OpenAI · Smartlead · Marketplace data
Headline metrics
- What it tackles
- Selling Amazon optimisation means competing with fifty agencies making the same claim, and the buyer cannot check any of them.
- What it delivers
- A pipeline that qualifies a brand through cost-ordered gates, audits its live Amazon listing across eleven attributes, and writes the first email from what it found, with a second branch for brands not on Amazon at all.
The problem in full
An Amazon performance agency could run marketplace accounts at a standard its own clients vouched for. Filling that capacity was the constraint. Prospect research was being done by hand, store by store, which caps a team at a handful of prospects a week no matter how many people are on it. The founder's brief was precise: reverse engineer what a full-time researcher would do with two hours to audit a single prospect's Amazon presence, and turn that into a system that runs at list scale.
The harder problem was the pitch itself. If you sell Amazon optimisation, your email competes with fifty other agencies saying they can improve someone's Amazon performance. The claim is unfalsifiable and free to make, so it is worth nothing. But an Amazon listing is public. Every deficiency is sitting on the product page where anyone can look at it. An email that names the actual problem can be verified by the recipient in ten seconds, in another tab. Nothing has to be believed.
Two facts about the market shaped the engineering. Nothing could be written about a brand until the system knew whether it was on Amazon at all, and a wrong answer in either direction ends the conversation in the first sentence. And roughly half of the qualified stores had no marketplace presence to audit, so a design that only knew how to write about a listing would have discarded half of everything the client had paid to find.
System design
I designed and built the pipeline as four workflows chained into one, in Clay, run with an Outlyst team. Qualification is ordered by cost. The cheap question runs first: is this an e-commerce store selling physical products, checked against the site's storefront platform, product catalogue and cart. A category exclusion layer then removes what the agency does not service, fresh grocery beyond pantry goods, cooked food brands, guitars, and items bulky enough that the merchant has to fulfil them itself, before any email exists to send. Only records that survive every gate reach the paid step, the live marketplace scrape.
The Amazon question is answered twice, by two independent methods, across brand name similarity, store status and product matching. Only when both checks come back negative is a brand classified as not on Amazon. Each record carries its own evidence, so every routing decision can be inspected after the fact.
For a brand already on Amazon, the pipeline scrapes its live product detail page at the moment of sending and reads it against eleven attributes: brand registry status, title character length, A+ content, bullet formatting, promotional video, photo count, brand story, review count and rating, storefront, price and Prime eligibility. Each is scored against the standard the agency applies to its own client work, and title limits vary by category, so the check varies with them. The email is then written from two or three most important selected findings, not all eleven. The model is handed the scored record and asked only to phrase it, so everything it can say is something the scrape observed.
Brands not on Amazon get the second branch instead of the bin. The pipeline searches the marketplace inside the prospect's own product category, pulls the top-selling competitors with their names and their rankings, and sends that across as a scoreboard: here is who is dominating your category, and you are not on the field. Both branches end in a sendable email, which is what made roughly half the addressable market usable.
The calls that made it work
Order the pipeline by cost
Scraping a live product detail page and putting a model over it costs real money per record and is worthless on a record that was never going to qualify. Everything cheaper runs first, so the expensive step only ever sees brands worth writing to. That ordering is the shape of the pipeline, not an optimisation added later.
Answer the Amazon question twice
Telling a brand about a listing it does not have, or pitching an Amazon launch to somebody already selling there, ends the conversation in the first sentence. Two independent checks, and only a double negative classifies a brand as not on Amazon. Unglamorous, and it is the difference between a smart email and an embarrassing one.
Cap the email at two or three findings
The audit produces eleven data points. A list of eleven faults reads as an attack from a stranger. Two or three reads as somebody who actually looked. The system deliberately withholds most of what it knows, because the email's job is to prove attention, not to deliver the full report.
Route the no, do not delete it
A qualification step that only deletes records makes the list smaller and the client poorer. Routing brands without a listing to a competitor scoreboard kept every qualified record in play and opened a second revenue route the client had not been working. Building it cost one extra prompt and one extra search.
The results in full
Main Campaign 3, the largest of the five, put 8,243 emails in front of 4,490 prospects and returned 259 replies, 44 of them positive: a 5.77% reply rate, and 44 replies from people who wanted the audit continued. That share is the return on an email that names a verifiable fact about the recipient's own listing. Across all five campaigns the pipeline reached 15,148 prospects with 28,905 emails.
The client kept more than a campaign result. A listing audit that a researcher would spend two hours on now runs at the scale of a list, across thousands of brands. The second branch made roughly half the addressable market usable and gave the agency a route to brands not yet on Amazon. And the exclusion layer did its work invisibly, keeping unsuitable brands off the client's calendar entirely.
Continue reading
The Model Router
Two trained classifiers behind an LLM agent that decides which one to call.
Compliance, Sold Cold
23 campaigns selling anti-money-laundering software to the compliance officers who buy it.
Outbound in Norwegian
A 25.26% reply rate across 18,062 Norwegian prospects, in a language I do not speak.