The question
- "Is my racket right for my style?"
- "Should I go faster or stay in control?"
- "Pre-made racket or custom blade and rubbers?"
PingPath case study
How I stopped an AI recommending products that don't exist
PingPath is my free quiz that recommends table tennis gear. The AI used to name items from memory. Now it can only pick what is in stock.
Why I built it
Players never know if their gear suits them
Table tennis players keep asking the same thing: is my equipment right for how I play? PingPath answers it.
The quiz asks
15 questions- skill level
- goal
- current setup
- strokes
- playing style
- table distance
- dominant wing
- handle
- grip
- pre-made or custom
- budget
- speed vs control
- indoor or outdoor
- how often
- accessories
What they get
Gear matched to how they play
- A blade, rubbers or pre-made racket picked for their style
- Only real, in-stock products, at real prices
- A link straight to the product page
01 / Problem
The AI recommended gear that wasn't there
A discontinued blade
The Butterfly Zhang Jike ALC. You can't buy it any more.
An out-of-stock rubber
Named with confidence. Not available to buy.
02 / Cost
Every miss was a lost sale
Wrong item namedfrom memory
Dead search linknothing to buy
Lost sale, confused customertrust takes the hit
03 / Fix
Give the AI a shortlist, not a free hand
The AI no longer answers from memory. It picks from a list of products that are in stock right now.
A: stock and price refresh every 6 hoursB: catalog snapshot rebuilds monthly
megaspin catalog apiclaudefastapivercelgithub actionsLinks resolve to real products
Buy links use real product IDs, resolved on the server. They are not searches. Amazon is the fallback.
Primary retailer: megaspin
I chose the retailer. The catalog, stock and price all come from it.
04 / Human role
The rules I gave the AI
I set the rules, chose the retailer and review the outputs. The AI works inside them.
Never invent a product
Pick only from the shortlist
Copy names exactly, quote the listed price
Say so when nothing fits
05 / Measure first
Few clicks? Check the links first
In August 2026 I saw lots of video views and few affiliate clicks. I audited the links before blaming anything.
Every link I checked carried the ID.
link audit, August 2026Every link I checked carried the tag.
link audit, August 2026The links were fine
So I added privacy-safe tracking to see where people drop off. No user IDs, cookies or IPs. Daily counts only.
tracking live from 2026-08-23Lesson: measure before you blame the tool
Without the tracking I would still be guessing.
06 / Results
What the funnel shows
130 completed quizzes, 40 affiliate clicks. These are event counts, not unique people.
Quiz to click
About 0.31 clicks per completed quiz. Events, not a per-person conversion rate.
source: app event log, 2026-08-23 to 2026-10-03
| Event | Count |
|---|---|
| Quiz completions | 130 |
| Affiliate clicks | 40 |
Clicks by retailer
source: app event log, same period
| Retailer | Clicks |
|---|---|
| megaspin | 29 |
| amazon | 11 |
Clicks by placement
source: app event log, same period
| Placement | Clicks |
|---|---|
| in the recommendation | 29 |
| cart button | 9 |
| add all to cart | 2 |
Daily activity, Aug 23 to Oct 3
Activity was strongest in the first two weeks, then thin. Days with no events show as zero.
source: app event log, daily counts
| Date | Quiz completions | Affiliate clicks |
|---|---|---|
| 2026-08-23 | 18 | 4 |
| 2026-08-24 | 26 | 4 |
| 2026-08-25 | 7 | 1 |
| 2026-08-26 | 6 | 7 |
| 2026-08-27 | 3 | 2 |
| 2026-08-28 | 7 | 5 |
| 2026-08-29 | 5 | 0 |
| 2026-08-30 | 0 | 0 |
| 2026-08-31 | 0 | 0 |
| 2026-09-01 | 1 | 0 |
| 2026-09-02 | 5 | 0 |
| 2026-09-03 | 0 | 0 |
| 2026-09-04 | 5 | 1 |
| 2026-09-05 | 9 | 5 |
| 2026-09-06 | 5 | 0 |
| 2026-09-07 | 7 | 5 |
| 2026-09-08 | 1 | 0 |
| 2026-09-09 | 2 | 0 |
| 2026-09-10 | 1 | 1 |
| 2026-09-11 | 0 | 0 |
| 2026-09-12 | 1 | 1 |
| 2026-09-13 | 0 | 0 |
| 2026-09-14 | 2 | 4 |
| 2026-09-15 | 4 | 0 |
| 2026-09-16 | 1 | 0 |
| 2026-09-17 | 1 | 0 |
| 2026-09-18 | 3 | 0 |
| 2026-09-19 | 0 | 0 |
| 2026-09-20 | 2 | 0 |
| 2026-09-21 | 0 | 0 |
| 2026-09-22 | 1 | 0 |
| 2026-09-23 | 1 | 0 |
| 2026-09-24 | 1 | 0 |
| 2026-09-25 | 0 | 0 |
| 2026-09-26 | 1 | 0 |
| 2026-09-27 | 1 | 0 |
| 2026-09-28 | 0 | 0 |
| 2026-09-29 | 0 | 0 |
| 2026-09-30 | 0 | 0 |
| 2026-10-01 | 1 | 0 |
| 2026-10-02 | 0 | 0 |
| 2026-10-03 | 2 | 0 |
quiz completionsaffiliate clicks
registered accounts.
from the earlier era when login was requiredsaved recommendations.
from the earlier era when login was required- Grounding fix ships
- Link audit: 4/4 and 4/4
- Funnel tracking goes live
- Data cut for this page
07 / What it means for your business
Same pattern, your customers
If an AI talks to your customers, fence it in with your own data.
Product finder
Recommends only what is in stock.
Quote bot
Quotes only from your price list.
Support assistant
Answers only from your own help docs.
Only sells what you actually have
How I measured this
Where each number comes from
- 130 completions, 40 clicks, 29/11, 29/9/2, daily chart: aggregate event counts from the app's event log, 2026-08-23 to 2026-10-03. No user IDs, cookies or IPs are stored, so these are events, not people.
- about 0.31 clicks per completion: 40 divided by 130. It is a ratio of events, not a conversion rate.
- about 2,870 products: the megaspin catalog size used by the fix.
- 4/4 and 4/4: my link audit in August 2026. Eight links in total.
- 45 accounts, 78 saved recommendations: from the earlier era when login was required.
- I measured no before-rate for bad recommendations. I make no sales, commission or revenue claims.
Your customers
Put AI in front of customers, safely
On a free AI Fit Call we look at where AI could help your customers, and what it should never be allowed to say.
Book a free AI Fit CallOne conversation. A clear view of where to start.