Welcome to the latest edition of Unpacking Commerce, our newsletter about emerging trends in retail, brands, and new commerce.
👉 TL;DR - four things in this edition:
GEO is SEO for AI answers: getting your pages quoted when ChatGPT or Google answers a question instead of showing links. Eighteen months in, two things are clear: being mentioned across the web counts more than backlinks, and for a shop the assistant reads the product feed before the page. What is still being measured is how much of it turns into sales.
Wero is a payment app owned by European banks. It moves money straight from the buyer’s account to the shop’s, with no Visa or Mastercard fee: 0.77% per payment in Germany, compared with about 1.8% for a card. It now replaces iDEAL, the method behind 72% of Dutch online purchases.
AI images in fashion are now on product pages, not just in tests: Zalando generates 90% of its content, Zara photographs a model once and dresses them digitally. At under €1 an image, the picture is no longer the bottleneck; the legal label and the models’ rights are.
💥EXIT💥Our very first investment closes a chapter this month: €40m in revenue, profitable, present in several countries, and a new shareholder on board. The story is in the newsroom below.
Where GEO stands after 18 months
1. What it is
🔍 GEO, generative engine optimisation, is the work of getting your page cited inside an AI answer in ChatGPT, Google’s AI Overviews, or Perplexity, rather than ranked in a list of links. The term comes from a 2023 paper that rewrote pages nine ways and measured what got cited more: adding sources, numbers, and direct quotes worked; keyword stuffing did not. The catch was visible in that first test: the metric is a share of the answer, so one page’s gain is another page’s loss. In practice, GEO means two things: a dashboard that asks the models thousands of questions and counts which brands come back, and the rewriting meant to move the count.
2. What has happened since
📉 The AI answer replaced the list of links as the default result on every major platform: Google’s AI Overviews in 2024, AI Mode worldwide in 2025 and past 1bn monthly users this summer, ChatGPT search since late 2024. The clicks went with the list: 68% of US Google searches ended without a click in early 2026. Shops see it in their numbers: across 2,500 German online stores, organic’s share of traffic fell from 42.5% to 33.3% in a year. What comes back from the assistants grows fast but remains a low single-digit share of most sites’ traffic. That gap, between the clicks lost and the clicks returned, is why a market exists.
3. Where the money went
🏦 Dozens of startups have entered the category in eighteen months, most of them selling the same dashboard. Profound raised $155m in three rounds, the last at a $1bn valuation; Peec in Berlin reached $10m of recurring revenue in 16 months and is, per Sifted, in talks at a $200m valuation. The barrier to entry is low: the product is prompts sent to public APIs, mentions counted, a dashboard on top. So the incumbents walked in. Semrush added it as a feature and took its AI products from $4m to $38m of recurring revenue in 2025; G2 lists 618 products in the category, up from seven in March 2025; Profound’s own entry plan is $99 a month. Kevin Indig’s verdict: “Monitoring tools have negative switching costs.” Hence the funded players’ pivot to agents that do the rewriting themselves. Whether that is a moat or next year’s checkbox is the open question.
4. The news this year
🗞️ Three events. Adobe paid $1.9bn for Semrush, an SEO company, and said why: “unlocking GEO for marketers as a new growth channel”. OpenAI dropped in-chat checkout in March and asked merchants for product feeds instead, then said on 31 August that ChatGPT Ads had passed $1bn of annualised revenue and opened self-service buying in Europe: a paid slot now sits under the same answer the tools measure. Cloudflare, which fronts more than 20% of web domains, blocks AI agents and training crawlers by default from 15 September on any new site’s page that carries an ad. Being read by a machine is now a setting the site has to open.
5. What the tests show
⚠️ Does it work? Four things are established. Being on Google's first page still helps, but less every quarter: 38% of the pages cited in AI Overviews ranked in Google's top 10 in March 2026, down from 76% eight months earlier, and only 12% of the pages ChatGPT cites rank there. Being talked about helps more than being linked to: across 75,000 brands, how often a brand is mentioned on the web predicts its presence in AI Overviews far better than its backlinks do (0.66 against 0.22). Rewriting your own pages does little: one benchmark found three successes in 54 tests. And the answers move all the time: ask the same question two days running and only a third to 40% of the cited sources are the same. What is not established is the money. The one study built to isolate cause and effect found more ChatGPT visits after optimisation, but not enough to rule out luck, and a review of 45 studies published in July concludes that "claims about GEO return on investment clearly outstrip the academic evidence". The one thing that works reliably is cheating. SafeGEO tested 22 ways of doctoring a product page so that a shopping agent picks it: leaving out the product's known flaws, adding false claims dressed in the language of experts and evidence, or hiding instructions in the page text that tell the model to recommend it. Flawed products made it into the agent's recommendations up to 83 points more often. Whether platforms keep allowing that is the question.
6. What remains open
❓ Five questions remain. Does it pay? No vendor has published a test with a control group, and Adobe's finding that AI-referred shoppers convert 60% better says nothing about optimisation: those shoppers arrive with their choice already made. What is left once the hype passes? Tracking is turning into a feature of tools companies already own, so the category could end up as a tickbox in an SEO suite, a line in the PR budget, since mentions are what count, or a column in the product feed, and those three futures are worth very different amounts. Who is in charge? The levers are spread across SEO, PR, the team that manages the product catalogue and, since 15 September, IT for the crawler settings: four departments that have never shared a budget. Will today's work still count next year? ChatGPT cited a source in 1.6% of answers a year ago and 6.8% now, and no study yet measures what happens once ads sit under the answer. And is the text even the problem? What an assistant needs from a shop is price, stock, availability, and specifications. Adobe found product pages 66% machine-readable, the worst score of any page type, and OpenAI and Google now ask retailers for a product feed rather than a better page.
Wero: banks take back the checkout
1. What it is
💳 Wero is a payment app owned by European banks. When you pay with it, the money goes straight from your bank account to the shop’s, with no Visa or Mastercard in between. It is run by EPI, the European Payments Initiative, a Brussels company the big eurozone banks set up in 2020 to own a payment system of their own: 18 shareholders, among them BNP Paribas, Crédit Agricole, Deutsche Bank, the Sparkassen, ING, KBC and Rabobank, plus the processors Worldline and Nexi. Wero began in 2024 as a way to pay friends. Online checkout went live in Germany in November 2025, with Lidl, Decathlon and Rossmann among the first shops, in Belgium in March 2026 and in France this month. EPI counts 58m registered users, many of whom never chose it: their bank moved them over from the national wallet Wero replaced.
2. What is at stake
Three things. Who takes the fee: every card payment in Europe pays Visa or Mastercard, which handle two-thirds of euro-area card transactions and whose scheme fees rose 34% between 2018 and 2022 while interchange stays capped by law at 0.2% and 0.3%. A bank-owned wallet keeps that margin, and the data, with the banks. Whether Europe can build a network at all: the banks’ first attempt, a European card scheme, collapsed in 2022 when 20 of the 31 founding banks walked out. Wero is the second attempt, by a cheaper route: no new network, but a merger of the bank-transfer wallets each country already had. Giropay closed at the end of 2024, Paylib in early 2025, Payconiq on 4 December 2025, and iDEAL is next. And the cost of failure: four working national methods dismantled for a brand shops do not yet trust. Ten months after the German launch, none of the country’s ten largest online retailers had integrated it, and three quarters of retailers surveyed by EHI were waiting to see.
3. The iDEAL test
iDEAL is what makes the bet credible. It is the Dutch bank-transfer button behind more than 1.5bn payments a year, about 72% of everything bought online in the Netherlands, in 350,000 shops, the only one of the merged methods that already beats cards at checkout; Giropay, Paylib and Payconiq were mostly used to pay friends. EPI bought it in 2023 and is folding it into Wero in stages: a combined iDEAL | Wero logo since 29 January 2026, all Dutch banks connected to Wero in October 2026, payments moved onto Wero’s rails after that, the iDEAL brand gone by the end of 2027. If Dutch shoppers keep paying the same way under a new name, Wero inherits Europe’s one proof that a bank transfer can be the default online, and a template for the countries next door. If they do not, it loses its only working e-commerce base. Two concessions already show where it hurts: EPI has pushed buyer protection to the last step of the migration, around mid-2027, and has had to promise Dutch shops that prices stay at iDEAL’s level until the end of 2028, because iDEAL costs them about 20 to 35 cents a payment and Wero is priced as a percentage.
4. What changes for a shop
Three things. The fee: in Germany PAYONE charges 0.77% per Wero payment, against about 1.6% for PayPal; Mollie lists Wero at a flat €0.32 in the Netherlands, €0.39 in Belgium and 0.90% + €0.25 elsewhere, against 1.80% + €0.25 for European consumer cards. On a €500 sofa, that is €9.25 by card, €4.75 by Wero in Germany or France, 32 cents in the Netherlands. The cash: payment is taken instantly and there is no chargeback, so an issuer cannot pull the money back, which ends card fraud losses and, until purchase protection ships, the buyer’s safety net with them. The features: no recurring payments and no one-click, the shopper scans a QR code or confirms in the bank app; subscriptions are in pilot with Orange’s Sosh, instalments in preparation with no date, and iDEAL’s own in3 instalment product is being wound down because Wero cannot carry it. What has not changed is the shopper’s side of the button: EPI’s own merchant lead counted about 300 live German shops in March, and EPI has still published no tariff, only “a small percentage fee with built-in caps”. Does the fee gap move a shop before its customers know the name?
AI fashion imagery is ready
1. Where the technology stands
🖼️ Generated fashion imagery has passed the only test that counts: the big retailers put it on product pages. Zalando told Reuters in May 2025 that 70% of its editorial campaign images were AI-generated; by Shoptalk Europe this year, the figure for machine-generated site content was 90%. H&M has run digital twins of 30 real models since July 2025, each image labelled and watermarked. Zara photographs a model once and lets AI dress them in the rest of the range. Zalando’s own line is that the output is no better than a photographer’s: “it is really about how new, how relevant it is”.
2. What the price change does
Zalando puts production at three to four days instead of six to eight weeks and costs down 90%. In the companies we look at, a generated on-model image now costs less than €1, often a few tens of cents. What stays in the bill: Zara pays models the standard shoot fee even when they never come back to the studio, and H&M asks for consent on each use. The saving is on the studio, the crew, the logistics and the calendar, not on the talent. The consequence is bigger than the saving. At the old price, one image per item, shot once, amortised over the season; at well under a euro, the image becomes a variable that can be reset like a price: a version per market, per weather, per audience, per week, and tested against the others. In a back-planned calendar, the weeks saved do not lengthen the season; they move the imagery deadline closer to launch, so late additions, reorders and reactive drops get on-model pictures in time instead of a flat lay or nothing. That makes imagery a merchandising tool before it is a marketing one, and the question for a retailer is no longer whether to generate, but what to hold constant.
3. Where the constraint moved
To the label. Since 2 August 2026 in the EU, a generated image must carry a machine-readable mark and, for a realistic likeness, a visible one; in New York, since 19 June 2025, a model’s digital replica needs separate written consent covering purpose, pay and duration. The label is where the money leaks: in the Nuremberg Institute for Market Decisions’ experiments on consumers in the US, UK and Germany, the same ad marked as AI-made was judged less natural and less useful, and willingness to buy fell; in a survey of 411 shoppers, 71% could not tell generated from shot, 59% wanted a label anyway, and 31% grew sceptical once told. This is why H&M and Zara pay real models for pictures they never posed for: a twin of a named person survives the label, a synthetic one does not. Guess found out in July 2025, when an AI model in Vogue with a footnote credit drew a wave of reader anger. The talent fee has become the price of the disclosure, and the asset library, built to store files, now has to store rights: who, for what, until when.
4. Who makes the last picture
Increasingly not the retailer. Google’s try-on renders a retailer’s listing on a body built from the shopper’s own selfie, in the US since December and in Australia, Canada and Japan since October; Meta wants ads generated end to end by the end of 2026. In both cases, the platform generates the final image from the retailer’s product data and the platform’s picture of the customer. The retailer’s job shrinks to the input: a clean garment shot, consistent across the range, with the attributes right, which is the same discipline the product feed already demands. Whether a fashion retailer still needs an image pipeline in three years, or only a product-data pipeline and a rights ledger, is the question nobody in the sector has answered.
💥 EXIT 💥
Some exits mean more than others.
As we were laying the very first bricks of SPRING’s investment strategy, we made our first-ever investment in Hardloop, then a young online retailer of outdoor sports gear with the ambition of becoming Europe’s Backcountry.
What won us over at the time?
🔹 A hybrid social-media-and-commerce model that lowers customer acquisition and retention costs while diversifying revenue
🔹 Curious, attentive, pragmatic and highly agile founders
🔹 An exceptional capacity for work and execution
🔹 A rare rigour, essential in e-commerce, where margin is won at every level
A few years on, the results speak for themselves:
✅ Over €40m in revenue (from €270k when we invested)
✅ A profitable company, with EBITDA growing too
✅ A remarkably efficient e-commerce model, built with little capital
✅ An international footprint, rare for a young B2C company
Congratulations to Julien, Guillaume and the whole Hardloop team for how far they have come.
A new chapter opens today with a new shareholder coming on board to take the company to its next stage of growth. We have no doubt you will keep writing this story and achieving great things, with the team and the backing you now have. 🙌
🔗 NIQ: retail’s share of EU consumer spending falls for the fourth year running - Retail took 31.9% of household spending in 2025, the rest drifting to travel, leisure and hospitality; turnover grew 2.1% against 2.5% inflation, so the sector shrank in volume.
Inditex: first-half sales up 7.6% to €19.8bn, gross margin 58.7% - Net income €2.98bn, up from €2.79bn, with growth in stores and online. The mid-market is collapsing around it; Zara’s owner keeps taking share and margin at once.
ECDB: Bol widens its lead over Amazon in the Netherlands - Bol sold €5.8bn in 2025, up 12%, against €3.3bn for Amazon.nl, up 2.7%; the gap is now 75%. AliExpress took third place, and Amazon’s average commission runs at 19.7% against Bol’s 14.2%.
WARC: retail media passes $200bn as growth slows - $200.4bn this year, growth outside Amazon down to 9.8%, the lowest WARC has recorded, and more than 20 ads per page at the largest retailers: a channel that converts existing demand without creating any.
👍 If you like Unpacking Commerce and want to help it grow, please share this newsletter with your colleagues, followers, and friends. Have a great day!
About us
Spring Invest is a European investment fund dedicated to companies that are shaping the future of commerce. We champion doers who build innovative companies making commerce better, from enabling technologies to new commerce models and everything in between. More info about our investment thesis 👉 here.