Traditional competitor matching for repricers works by keyword search: extract keywords from your listing title, query eBay's Browse API, take the top N results, treat them as competitors. It's fast, cheap, and includes 5-15% wrong products because keyword matching doesn't understand product identity. A "Toyota Camry oil filter" search returns Camry oil filters + Corolla oil filters + Camry air filters + generic filters that just mention Camry in the description.
Wrong-catalog matches cause the two worst repricer failure modes: catastrophic underpricing (matching against cheaper wrong products) and phantom overpricing (matching against pricier wrong products with different function). Vision verification is what stops both, by settling product identity before a price is ever compared.
How keyword matching fails
Example: your listing is a 2013 Toyota Camry XLE 2.5L headlight assembly, driver side. Keyword extraction produces: "2013 Toyota Camry headlight driver."
eBay Browse API returns 20 candidate listings. Analysis:
- 14 are actually 2013 Camry driver headlights ✓
- 3 are 2013 Camry passenger headlights (wrong side)
- 2 are 2013 Camry fog lights (wrong part type but title mentions "headlight")
- 1 is a 2010 Camry driver headlight (wrong year)
If your repricer treats all 20 as competitors and computes median: median is pulled by the 6 wrong matches. Passenger headlights typically cost more (different assembly complexity). Median inflates. Your pricing drifts upward chasing a phantom.
The vision verification pipeline
For each keyword-matched candidate, run a secondary check:
Step 1: Fetch candidate's primary photo.
Step 2: Compare against your listing's photos using a vision similarity model.
Step 3: Cross-check item specifics: does candidate's MPN, brand, fitment year match yours?
Step 4: Assign confidence score (0-1).
Step 5: Only include candidates above 0.7 confidence in your competitor pool.
Result: 20 keyword matches → 14 verified matches. Wrong products excluded. Median accurate.
Fail-closed policy
When the verifier is uncertain (confidence between 0.5-0.7), the safe default is to SKIP repricing that SKU rather than include ambiguous matches. Failing closed prevents catastrophic errors at the cost of occasional missed repricing opportunities.
Older repricers fail-open — include everything and hope for the best. Modern verifier-backed repricers fail-closed by design.
Where vision beats humans
Human eyes miss subtle product differences that vision AI catches:
- Same brand, slightly different sub-model (Camry LE vs XLE trim variants)
- Same product family, different generation (2013 vs 2014 body style change)
- OEM vs aftermarket variants (visually identical to untrained eye)
- Left vs right side items (many categories are asymmetric)
We measured this rather than estimating it. On 380 (our listing, competitor) pairs labelled by hand in September 2026, Profitio's matcher holds 96.6% precision — and 92.9% on the pairs that were held out of rule-tuning entirely, which is the number that actually predicts how it behaves on your catalogue. Five false matches in 331 judged pairs. Precision is the figure we tune for, because a wrong match feeds a different part's price straight into your repricer, while a missed one only costs you a competitor you never needed.
What still requires humans
Vision AI can't verify claims like "genuine leather" vs "faux leather" from photos alone. Can't distinguish 24k gold plating from 14k. Can't tell if an electronics unit is tested-working vs untested. Anything requiring physical handling or destructive testing is beyond vision AI.
For those categories, keyword matching + human spot-check remains the workflow. Fashion authenticity, high-value jewelry, medical/industrial equipment.
Cost economics
Vision API calls cost ~$0.001-0.003 per verification. For a repricer running on 1,000 SKUs with 15 candidates each, that's $15-45/day in verification API calls. Real cost across a month: ~$500-1,500 depending on volume.
Compare to the alternative: without verification, a wrong-catalog mispricing incident on 200 SKUs (typical) leaks $2,000-8,000 in margin over a week before detected. One prevented incident pays for months of verification API costs.
The margin protection math
The cost of a wrong match is not spread thinly — it is concentrated and it compounds. One mismatched competitor pulls a SKU down to a price that was never for your product, the repricer treats that price as market truth on every later run, and the SKU stays there until somebody notices by hand. That is why verification belongs before the comparison rather than in a report afterwards: by the time a wrong price shows up in your margin report, it has already been live for days.
Profitio's repricer includes AI competitor verification by default with fail-closed policy. 14-day free trial.