AI in ecommerce: what we learned building a fashion commerce platform

On a fashion commerce platform we build and run, what mattered most for AI in ecommerce was testing features on real shopper inputs and handling the ways model calls go slow, fail or run up a bill. Four changes cut the AI stylist’s median response on live traffic from 30.1 s to 15.5 s. A fashion-tuned image embedding put the right product in the top five for 63.3% of 60 real test cases, where CLIP managed 38.3%.

A refused virtual try-on used to take about 110 to 115 s to fail, and now fails in about 20 s. A background job spent about $1,100 on 419,057 vision calls it didn’t need, and raised no errors. Underneath, the storefront serves cached HTML in about 70 ms.

This post pulls the five field notes into one account of AI ecommerce development, linking each deep dive. It also covers the commerce foundation every store has to settle early, what AI shopping agents ask of a store, and a kickoff checklist.

Five field notes, their numbers and how each was measured

All figures are from this platform.

The store foundation to settle before the AI features

Any ecommerce app development project has to settle five things early, whatever AI comes later.

Our standard build includes a global CDN with an enforced performance budget, technical SEO, WCAG AA accessibility and a consent banner.

Payments. Keep card numbers off your servers, so that PCI DSS scope stays small. SAQ A is the self-assessment questionnaire for merchants that have outsourced all account data functions to PCI DSS validated and compliant third parties.¹ That means the provider’s fields in an iframe, or a redirect to its payment page. The revised SAQ A, in force since 31 March 2025, dropped the payment page script requirements 6.4.3 and 11.6.1. It added a criterion that the merchant confirm its site is not susceptible to script attacks that could affect its e-commerce systems.¹ FAQ 1588 says it applies to pages that embed the provider’s form and not to redirects. A merchant can meet it with techniques like those in 6.4.3 and 11.6.1, or with the provider’s confirmation that its solution protects the payment page. Your acquirer or the card brands confirm which questionnaire applies.²

Speed. Agree a performance budget before the first feature ships, and fail any release that breaks it. At the 75th percentile of page loads, split by mobile and desktop, a good page has LCP within 2.5 s, INP of 200 ms or less and CLS of 0.1 or less.³ INP replaced First Input Delay as a Core Web Vital on 12 March 2024.⁴ In a study Google commissioned from Deloitte, a 0.1 s gain in mobile site speed went with 8.4% more retail conversions and 9.2% higher average order value.⁵

Product data. Google treats product pages where people can buy as merchant listings. Each needs a Product with a name and an image, and an Offer with a price above zero and a currency.⁶ Shopping agents read the same data, so generate the markup and the feeds from one product source.

Accessibility. The European Accessibility Act has applied to e-commerce services in the EU since 28 June 2025.⁷ The EU’s Your Europe portal lists online shops among the services covered, and exempts service businesses with fewer than 10 employees and under €2 million in turnover.⁸ Check the national law in each country you sell to.

Consent. In the EU, cookies that need consent can’t be set until the visitor gives it through a positive action. A vast majority of authorities in the EDPB’s cookie banner taskforce considered it an infringement when the layer with the accept button offers no way to reject. The report also says pre-ticked boxes don’t give valid consent, and that withdrawing consent must be as easy as giving it.⁹

Halving the AI stylist’s response time

The stylist streams its answers to shoppers and can act for them, for example by placing an order. Each reply runs a chain of model calls that classify the request, compose the answer, validate it and critique it, over hybrid keyword and vector retrieval. The calls run in sequence, so their times add up, and every generated token costs time.

We switched reasoning off for the compose step, and added hedged requests, which send a duplicate call when the first runs long. We moved the calls from Azure OpenAI to OpenAI’s API after seeing a slow tail on Azure. We also let retrieval return once it had enough results. Over four days of live traffic, the median fell from 30.1 s to 15.5 s, and the p90 from 40.3 s to 20.6 s. Alongside the latency work, the regression suite went from 18 of 47 passing to 47 of 47.

Hedging stopped paying once the tail shrank. In a later sample it fired 142 times, and only 4 of those duplicates came back first. A duplicate carries the same work as the original, so it only wins when the original was held up on the serving side, which the other changes had reduced. The algebra is in How we halved an AI stylist’s latency, and when hedging stops paying.

Visual search for ecommerce starts with the embedding

A shopper uploads an outfit photo, or a frame from a video, and gets back products that match. The pipeline finds the person, describes the garments in one Gemini call, boxes each garment with GroundingDINO through a hosted detection API and cuts it out with SAM 2. Each cutout is embedded and searched in a managed vector database, and a capped re-rank orders the candidates. A whole-frame vector would blend garments, person and background.

FashionSigLIP led, and the platform uses it. With 60 cases, each hit rate’s 95% interval spans up to about 12 points either side. At that precision the test separates FashionSigLIP from CLIP, and can’t separate CLIP from DINOv2. DINOv2 keeps one job, spotting exact duplicates of catalogue images. The method is in Comparing FashionSigLIP, CLIP and DINOv2 for visual product search.

AI product images that fail quietly

Try-on renders a garment onto a photo of a person with Gemini image generation. In the first version, the parser dropped the reason for a safety refusal, so the retry loop treated it like a timeout and spent five retries. The shopper saw an error after about 110 to 115 s, and the failure was cached with no expiry. A refusal is now final on the first attempt, so a refused try-on fails in about 20 s, and a cached failure lapses after 30 minutes.

Background-removal cutouts also lost their alpha channel when saved, and one image came out 92.7% black. Compositing onto a chosen background before any step can drop alpha prevents it. For second-hand items photographed by their sellers, the platform regenerates a studio-style shot with Gemini. Early outputs sometimes rotated the item or recoloured its metal hardware, and lowering the temperature to 0.2 fixed both.

Retry rules and alpha traps are in How AI image pipelines fail quietly, and how to catch it. Generated images shown to shoppers can also carry marking duties, covered in our post on AI content marking rules.

Spend guardrails for LLM batch jobs

A background job sent items that already had a classification back to a vision model. Its lookback window came from a variable named for two days that held 730. The job made 419,057 calls it didn’t need, for about $1,100, roughly $0.0026 each. No call failed, so the only signal was the bill.

Rate limits cap throughput, and the field note shows budget alerts lagging behind spending, so provider controls catch a job like this late. The guardrails it recommends act sooner.

• Select only items that the current model and prompt version haven’t processed yet.
• Estimate a run’s cost before the first paid call, and stop above a threshold unless a person approves.
• Put the unit in every duration’s name, and give the lookback an upper bound.
• Write a ledger row for every paid call, and enforce a daily budget per job in code.

The code is in How a backfill bug ran up a $1,100 LLM vision bill.

Ecommerce site speed from caching HTML by device

The storefront is a Next.js app that renders each page on the server, with different HTML for phones and desktops. Before the change, every request waited for a full render. We now cache that HTML at Cloudflare’s edge, with the device type in the cache key, so phones and desktops get separate copies. Cache hits return the HTML in about 70 ms. On mobile, the PageSpeed score went from 65 to 87, and LCP in the lab came down from 14.1 s to 3.8 s, still above the 2.5 s that counts as good.

Anything that differs per visitor, such as account, cart and checkout pages, has to bypass the cache, through rules ordered after the caching rule. Each deploy needs a purge. The rules, purging per device type and a stale-while-revalidate trap are in Caching server-rendered Next.js HTML at the edge, by device type.

What agentic commerce asks of a store

AI shopping agents search, compare and buy for a person, reaching a store through protocols that connect them to its catalogue, checkout and payment provider. Three are worth tracking.

Agents work from product data, so feeds need complete prices, availability and images. Google’s UCP integration requires an active Merchant Center account with eligible products.¹³ Checkout should be callable as an API, with your systems pricing each order and accepting or declining it. Check which protocols your payment provider supports. Under Google’s checkout the retailer stays the seller of record.¹⁴ Our comparison of MCP, A2A and function calling covers the agent-side layers.

A kickoff checklist for AI in ecommerce

9io has kicked off products in healthcare, ecommerce, edtech and fintech. For an AI commerce product, we’d settle these first.

1. Confirm your PCI DSS questionnaire with your acquirer, and keep card entry in the provider’s hosted fields or page.
2. Set a performance budget against Core Web Vitals at p75, and keep personal pages out of any HTML cache.
3. Generate product markup and feeds from one product source.
4. Build to WCAG AA from the first component, and ship a consent banner with reject beside accept.
5. Time each step of a chained AI feature, log its tokens, and tune reasoning effort per step against a regression suite.
6. Build a test set from real shopper inputs before choosing a model, and report intervals next to hit rates.
7. Sort model responses into success, refusal, empty or transient before any retry, and expire cached failures.
8. Give every job that calls a paid model a cost estimate, a ledger and a daily budget.
9. Keep product data, checkout and payments ready for agent protocols, and recheck their status each quarter.

Frequently asked questions