2026 E-Commerce Guide: Speed, AI and Analytics

In Brief

A 2026 e-commerce guide comes down to three measurable areas. Speed: stores that pull mobile LCP under 2.5 seconds get more orders from the same ad budget. AI: agents connected to live product and returns data close most customer questions without anyone waiting for a human. Data: teams that do not read cart abandonment, channel-level ROAS and customer acquisition cost every week keep pouring money into the wrong channel. Work through the three in that order and the conversion rate moves within a single quarter.

Last season a cosmetics brand in Istanbul spent 180,000 TL on ads and took 1,100 orders. That works out to 163 TL of ad spend per order. Three months later the same budget produced 1,380 orders. The campaigns had not changed. What changed was a 1.9 MB hero image converted to webp, three required fields removed from the checkout form, and the delivery promise moved from the cart to the product page. Technical debt hides inside the ad invoice, and nobody goes looking for it there.

In Turkey the squeeze is tighter still. Trendyol and Hepsiburada often show the same product cheaper than you do, next-day delivery has become the default expectation, and installment options still decide the moment of purchase. Your own store survives on something other than price: accurate stock, a returns policy readable on one page, a search box that works, and a checkout flow people recognise.

Order matters too. Raising ad spend before fixing speed is carrying water in a leaking bucket. Product data has to be cleaned before an agent goes live, because an agent spreads a wrong stock figure just as fast as a right one. Splitting budget before measurement is set up is guesswork with extra steps.

Putting a price tag on page speed

Core Web Vitals thresholds give everyone a shared vocabulary: LCP under 2.5 seconds, INP under 200 milliseconds, CLS under 0.1. Track those three with field data rather than lab tests. A large share of Turkish mobile traffic arrives from mid-range Android phones on connections that fluctuate through the day. A product page that opens in 1.4 seconds on office fibre can stretch to 5 seconds on a commuter train. Make decisions on the field number.

Start with these three files

• Hero images. Convert the first image on category and product pages to webp or avif and cap the width at 1200 pixels. On its own this step can win back close to a full second of LCP.
• Third-party scripts. A live chat widget, a heatmap tool, two separate ad pixels, and a testing script somebody forgot to remove when the campaign ended. Each one competes for the browser main thread. Delete what you no longer use and defer the rest until after the page has rendered.
• Fonts. One family with two weights covers most designs better than two families with six. Adding font-display: swap keeps text visible while the font downloads.

What speed is worth in orders

Picture a store with 100,000 monthly visitors and an average order value of 1,400 TL. Using conversion figures close to industry averages, the arithmetic looks like this.

State	Mobile LCP	Conversion rate	Monthly orders	Monthly revenue
Current theme, uncompressed images	4.1 s	1.4%	1,400	1,960,000 TL
After image and font cleanup	2.9 s	1.7%	1,700	2,380,000 TL
Deferred scripts and server caching	2.1 s	2.0%	2,000	2,800,000 TL

Not one extra lira of ad spend produces that 840,000 TL gap. Once the image and script work is done, attention moves to on-page decisions: the add-to-cart flow, where shipping information sits, how many fields the form asks for. For tested steps in that territory, seven ways to lift your conversion rate gives a longer checklist. Speed work is never finished in one pass either. Every new landing page, plugin or embedded video spends the budget again, so repeat the mobile measurement monthly.

Where AI agents actually earn their place

The word agent has taken a beating over the past year. In practice it means something simple: software wired into your store data that can take a step on its own. It reads an order number, pulls the status from the courier integration, opens a return request, checks stock. What separates it from a chatbot is not the conversation, it is the record it leaves behind.

Three concrete uses

• Pre-purchase questions. "Do you have this shoe in a 42, and will it reach Izmir tomorrow?" The agent checks the stock table and the courier cut-off rules and answers plainly. Because a wrong answer becomes a return, the stock the agent sees has to be live.
• Post-purchase tracking. A large slice of the support inbox is some version of where is my parcel. Handing that traffic to an agent frees your people for returns, damages and supplier chasing.
• Catalogue content. Writing descriptions for 3,000 products stopped being a human job. The agent drafts from product attributes and an editor corrects and publishes.

The order of setup

No agent goes live before the data is fixed. Wrong stock counts, missing variant names and a delivery cut-off written two different ways in two places all turn into faster-spreading mistakes once an agent repeats them. The sequence runs like this: clean the product attributes, consolidate returns and shipping policy onto one page, write the fifty most common questions and their correct answers into a table. Connect the agent after that.

Write the handover rules up front as well. Price negotiations, damaged goods, and returns past the window belong with a person. On Tecof AI agents that threshold is set from the dashboard and the conversation history arrives with the agent, so the customer never has to explain the problem twice. If you want the wider picture of where autonomous selling is heading, how agentic commerce works covers it in more depth. For a pilot, pick a single category, have a human review every reply for two weeks, and widen the scope once the error rate drops below five percent.

Getting found in AI search

The answer to a typed question is now frequently a written paragraph rather than a list of blue links. Ask for a moisturiser suited to someone over forty and the model names three or four brands and cites its sources. Landing in that list calls for different preparation than classic SEO, because the model is not reading your page so much as parsing it.

What a model can read

• Complete Product schema on every product page: price, currency, availability, brand, and GTIN where you have one. A missing field is enough to drop the product out of comparison answers.
• A question and answer block near the foot of the page. Eight to ten questions serve both the shopper and the model, and each answer should stay within three or four sentences.
• A clean specification table: capacity, material, warranty period, washing instructions. Models quote from tables far more readily than from long prose.
• Accessible HTML. A page that buries its facts inside an image reads as an empty page on the model side.

How to measure it

Requests from AI crawlers show up separately in server logs. In your analytics, collect sessions whose referrer carries a model name into their own segment. Volume looks modest at first, yet the add-to-cart rate for those sessions comes in above the organic average in most stores, because the visitor arrives at the decision stage rather than the research stage. For the mechanics and for which content tends to get cited, the work on AEO and GEO is a solid starting point. Once measurement is running, find your three most-cited pages and copy that structure across the other categories. Getting the template right once beats writing bespoke content for every product.

Tying ad budget to a measurement

Two formulas carry most of the weight on the advertising side. ROAS is ad revenue divided by ad spend. CAC is ad spend divided by the number of new customers won. Teams that read only one of them pile budget into whatever channel looks most efficient and quietly starve the top of the funnel.

How a 108,000 TL monthly budget splits

Channel	Spend	Revenue	ROAS	New customers	CAC
Google Search, non-brand	45,000 TL	202,500 TL	4.5	120	375 TL
Meta retargeting	25,000 TL	137,500 TL	5.5	40	625 TL
Meta cold audience	30,000 TL	75,000 TL	2.5	95	316 TL
Email and SMS	8,000 TL	96,000 TL	12.0	12	667 TL

A manager reading only the ROAS column shuts down the cold audience campaign. Yet a third of all new customers come from there, and it acquires them more cheaply than anything else. Turn it off and the retargeting pool drains within two months, the email list stops growing, and total revenue falls in month three. Read the channels as a chain that feeds itself, not as four independent scoreboards.

When the measurement itself is broken

Numbers only mean something if the tracking is sound. In a store without server-side event forwarding, a good share of iOS traffic disappears and Meta under-reports its own contribution. Even blunt instruments like a coupon code or a one-question post-purchase survey give you a more honest picture than the platform report alone. Set aside one day a month to reconcile dashboard revenue against the accounting figure. If the gap runs past ten percent, hold off on budget decisions and repair the tracking first.

Balancing stock, shipping and marketplaces

When your own store and the marketplaces draw on one warehouse, the real risk is not price, it is stock sync. If the last unit sells on Trendyol and stays purchasable on your site for another fifteen minutes, you have bought yourself a cancelled order and a dented reputation. Shorten the sync interval to minutes and hold a buffer on your fastest-moving lines.

Real ways to differentiate

• Product laddering. List your three bestsellers on the marketplace and sell the bundle only on your own site. That creates separation without cutting price.
• Easy returns. Doorstep collection, a return code in one click, and a refund within two days. This is where you can beat marketplace standards outright.
• A delivery promise you keep. State on the product page that orders placed before 15:00 ship the same day, then hold to it. A promise you never made costs less than one you miss.
• Content for members. How-to videos, care guides, spare part lists. Marketplaces do not carry any of it.

Order in the warehouse

No automation works on a catalogue without SKU discipline. When colour and size variants are named by different rules across the catalogue, the integration misses matches and the reports come out wrong. For picking routes, shelf labelling and barcode scanner layout, warehouse management and logistics flow lays out a workable framework. Do not tie yourself to a single courier either. Regional performance between Yurtici, Aras and MNG varies widely, and reporting delivery times province by province, then switching carriers where service lags, brings the return rate down visibly.

Payments, consent and compliance

Checkout is the most expensive point in all the traffic you paid for. Every additional field there is written straight into the abandonment rate. On the card screen, ask for the cardholder name, card number, expiry and CVV, and nothing else. When the billing address matches the delivery address, a single checkbox should be enough.

Expectations specific to Turkey

• Installments. In high-basket categories, show the installment table on the product page so it is not a surprise at checkout.
• Bank transfer discount. Passing part of the saved card commission to the customer lowers your payment cost and speeds up cash flow.
• 3D Secure. Making it mandatory cuts fraud, though the bank confirmation screen can be slow depending on the issuer. Report failed 3D attempts separately.
• Cash on delivery. It carries return risk, but for a first-time buyer it is a door into trust. Keeping it open below a set order value is a sensible middle ground.

KVKK and IYS

Two separate obligations run alongside each other on the data side. The privacy notice informs, while explicit consent has to sit in its own unticked box. One checkbox tucked under the order form that bundles marketing permission with contract acceptance is treated as invalid. If you send commercial electronic messages, those permissions also have to be registered with IYS, and a single bulk SMS to an unregistered list can turn into an administrative fine larger than the year's ad budget. Set the cookie banner to opt-in by default as well. Building this properly takes a day, while cleaning it up afterwards takes months.

A 90-day implementation plan

Nobody tackles all of the above in one week. With a small team, a realistic calendar looks like this.

Week	Work	Owner	Success measure
1-2	Mobile speed audit, image conversion, removal of unused scripts	Developer	Mobile LCP under 2.5 s
3-4	SKU and variant naming rules, shorter stock sync interval	Operations	Cancellations from stock errors near zero
5-6	GA4 ecommerce events, server-side forwarding, coupon-based source tracking	Marketing	Dashboard and accounting revenue within ten percent
7-8	Product schema, question blocks, standardised specification tables	Content	Valid schema across every category page
9-10	AI agent pilot in one category, written handover rules	Support	Half of routine messages closed automatically
11-13	Budget redistributed on channel and acquisition cost data	Marketing	New customer count rises while CAC holds

A plan that stays in a table does nothing. Give every row an owner and a date, and let the weekly meeting discuss only the success measure column. If you would rather run most of these steps from one place, the Tecof e-commerce platform keeps speed, stock and agents in the same dashboard instead of scattering the team across three separate tools.

Be honest about resourcing. This calendar does not need a full-time squad: one developer day, two operations days and one marketing day per week covers it for most stores. If part of the work goes to an agency, write the success measure column yourself and ask which number comes from which source before you sign off on the reporting format.

Tomorrow morning the first job is clear. Open analytics, list the five product pages with the most traffic over the past 28 days, download their hero images, convert anything above 300 KB to webp and upload it again. You will be done before lunch and see the effect in field data within two days.

Frequently Asked Questions

How long does it take to launch an online store in 2026?

With a hosted platform and a template, the technical setup takes about a week. The real time goes into product data: photography, descriptions, variants, shipping rules and returns copy. Three weeks is realistic for a fifty-product catalogue. If the catalogue is messy, double it.

Which tool should I use to measure page speed?

Do not rely on one. PageSpeed Insights shows both a lab score and field data, and the field data is what you should act on. The Core Web Vitals report inside Search Console groups pages together, which suits weekly tracking better. Run your checks at the same hour and on the same device class each time.

Will an AI agent replace my support team?

It does not replace them, it shortens the queue. Once repetitive shipping and stock questions are handled, your people can spend time on returns, damages and wholesale enquiries that need judgement. Weak handover rules make the agent push on the wrong topic and cost you the customer, which is why every reply needs human review for the first fortnight.

Why invest in my own site when I already sell on marketplaces?

Marketplaces keep the customer data and can change commission rates without asking you. On your own store the email address, order history and consent records stay with you, which makes the second sale far cheaper than the first. A healthy balance treats the marketplace as a discovery channel and pulls repeat purchases back to your site.

What ROAS should I aim for?

There is no single correct figure. At a 40 percent gross margin your break-even ROAS lands near 2.5; at 60 percent it drops to about 1.7. Calculate your own break-even point first and set the target above it. In categories with strong repeat purchase, a low ROAS on the first order can be perfectly acceptable.

What is the minimum I need for KVKK and IYS compliance?

Publish the privacy notice, keep explicit consent in its own unticked box, and register marketing permissions with IYS. Put retention periods into a written policy and cover data sharing with partners in your contracts. For anything more detailed, work with a lawyer who handles Turkish data protection regularly.

Where should I start on a small budget?

Speed. Compressing images and clearing out scripts costs time rather than money, and the benefit spreads across every channel. Measurement setup comes second, email and SMS flows third. Increasing ad spend before those three are done is the most expensive move available to you.

Do orders really come from AI search?

Volume is still small next to organic search. The intent behind those sessions is sharp, though, so the conversion rate can run high. Product schema and question blocks are work you would do for SEO anyway, so the extra cost is low and the position is worth holding.

Will switching themes fix my speed problem?

Sometimes, though more often it postpones the problem. The slowness usually traces back to uncompressed images and accumulated third-party scripts rather than the theme itself. Measure first: is the server response slow, are the images heavy, is a script blocking rendering? A theme change made without that answer reproduces the same problem in a new design.