Use Case: Model Validation (Custom Column Only)

Filter false positives out of your search results

What This Does.

A single AI column that answers one question per listing — is this actually the model I’m searching for? — so keyword noise stops burying the real results. No detailed analysis, no elaborate display: one Yes/No column, built for speed.

Table of Contents #

1. The Problem
2. When to Use This
3. What You’ll Build
4. Setup Steps
5. The Prompt
6. Suggested Workflow
7. Adapting for Your Products
8. Troubleshooting

The Problem #

Search for a specific model number and eBay gives you every listing where that number appears — for any reason. Say you buy Intel i7-12700 processors:

• “16GB DDR4 RAM for 12700K systems” → it’s RAM, not a CPU
• “Motherboard bundle — supports i7-12700 (CPU not included)” → no CPU in the listing
• “Lot of 12 mixed processors” → 12 is a quantity
• “i5-12400 with 12-month warranty” → wrong chip, 12 is the warranty

Your keyword matched. The product didn’t. At high search volume, this noise costs real time.

When to Use This #

• You search for a specific model range with variants (i7-12700, 12700K, 12700KF).
• High search volume creates many false positives to wade through.
• You want fast filtering, not a full analysis readout.
• The determination is simple: is it or isn’t it your target model?

Same pattern works for GPUs (“RTX 4090” vs “lot of 4”), camera models, part numbers — any product where the model number also shows up as a quantity, spec, or accessory reference.

What You’ll Build #

In the grid the column header reads ai Model Match.

Keep the Display Template Minimal.

The value here is the column, not the panel. The profile still wants a display template — ask your assistant for a one-line template alongside the prompt and paste that in.

Setup Steps #

This assumes you have SKU Manager configured — if not, do the Quick Start Guide first.

1. Build the Profile #

1. Open Data → Ai → AI Settings and create or select a profile for this search.
2. AI Configuration: Config Mode Standard, Provider OpenAI (ChatGPT), Model gpt-6-luna, your API key, and the validation prompt (below) as the System Prompt.
3. Fields & Columns: tick the fields the prompt relies on — Title, Model, and any item specifics you use (Model Number, Storage, etc.). Under User-Defined Custom Columns (Results), tick Model Match — it’s auto-detected from the prompt. Ignore any junk keys the detector lists.
4. Display Template: paste the minimal template.
5. Click Finish Setup on a new profile, or Start Service [Save Configuration] on an existing one.

A new profile opens on Google (Gemini), so change Provider to OpenAI (ChatGPT) first. The app then selects gpt-6-luna, listed under Recommended.

2. Create the Filter #

1. Go to Home → Filters, then click the + button at the top of the Filters panel.
2. Condition: match your search — e.g. Alias contains your search’s alias, or a simple Title Is not blank catch-all.
3. Action: Apply AI Prompt → Profile: the one you just built.
4. Set a Filter alias, click Ok, then turn on the filter’s Enabled switch in the Filters list.

AI analysis runs only on listings a filter catches. The filter’s Action must be Apply AI Prompt, with your profile selected. A new filter starts switched off, so turn on its Enabled switch.

The Prompt #

A worked example for the i7-12700 family — adapt the model names and patterns to your product:

# Model Validation Prompt v1.0 — column-only output

## AI COLUMN CONFIGURATION
### "Model Match" — AI Column
Must return EXACTLY one of: Yes, No

| Value | When to use |
|-------|-------------|
| Yes | Listing is genuinely an i7-12700, i7-12700K, or i7-12700KF processor |
| No  | The search term appears for another reason (false positive) |

## ROLE
You validate whether an eBay listing is genuinely one of the target
models, or whether the model number appears for another reason
(quantity, RAM spec, motherboard compatibility, warranty months).

## VALID MODELS
- Intel Core i7-12700
- Intel Core i7-12700K
- Intel Core i7-12700KF

## DETECTION LOGIC
Priority order:
1. Model field (highest) — if populated, trust it first
2. Title — check for true model vs false-positive patterns
3. Description — cross-reference if unclear

False-positive patterns (→ No):
- RAM or motherboard listings that mention the CPU they support
- "CPU not included" anywhere in the listing
- Numbers that are quantities, warranty lengths, or speeds
- A different processor where the number appears incidentally

Model precision:
- 12700 ≠ 12700K ≠ 12700KF — check the suffix explicitly.

## OUTPUT FORMAT
```json
{ "Model Match": "Yes" }
```
Single field, single determination. Before responding, verify the
number refers to the processor model itself and the listing actually
contains the processor.

Suggested Workflow #

Days 1–2: Validation Mode — Hide Nothing Yet #

1. Run your searches normally and watch the Model Match column populate.
2. Review the No rows — are they really false positives?
3. Review the Yes rows — are they really your model?
4. On any wrong call, use the Feedback Trace link and take the trace to your AI project for a targeted fix — see Improving with Feedback.

When You Trust It: Work the Column #

Sort or filter the grid on Model Match so the Yes rows lead and the No rows drop out of your way. Audit the No pile occasionally — a systematic mistake there is a prompt fix, and you want to catch it early.

Adapting for Your Products #

1. Define your valid models — exactly which variants count as Yes.
2. List your false-positive patterns — how does your search term show up when it isn’t your product?
3. Swap them into the prompt — keep the structure, replace the specifics.
4. Test and refine with feedback traces.

Troubleshooting #

• Column not populating: is Model Match ticked under User-Defined Custom Columns? Is the filter enabled, with Apply AI Prompt and your profile? Is the AI server running? New listings fill the column.
• Column missing from the list entirely: Model Match must be a key in the prompt’s JSON example.
• Wrong determinations: capture them with the Feedback Trace link and bring the trace to your AI project. Usually it’s one missing false-positive pattern.

After you change the prompt or model, rows analyzed under the old settings read Outdated. To analyze one again, click the refresh icon in its Current AI Status cell, or click the row if it is not the selected one. This is paid, and the app does not ask first. To re-run a row that reads Analyzed, right-click it and choose Refresh AI analysis. The app asks first.

With Analyze new listings only selected, the first batch a search pulls when you press Start stays Not analyzed, except the row you have selected, which is analyzed. Click any other row in that batch to analyze it. Rows found later are analyzed as they arrive.

The app finds your columns by reading the JSON keys in your prompt’s example output. A name that is not a JSON key there is not listed. Open the Fields & Columns tab, or click the Auto Detect Custom Columns icon at the top of the columns panel, and the list reads the prompt again.

Result: clean search results, with keyword spam sorted out by a column instead of by your eyes.

See also: Setting Up AI Columns · Improving with Feedback · Quick Start Guide