SMG Swiss Marketplace Group: The Market Sold a Toll Bridge Because Software Got Cheaper

Larix Research · Market View · SMG is a CHF ~2.6bn company with syndicate research coverage — outside the undercovered universe Larix exists for. We publish views on covered names when we believe the prevailing framework is wrong, because the framework is what transfers to the names we do cover. No rating, no target. Disclosures at the end.

A chart and an income statement that disagree

SMG Swiss Marketplace Group — ImmoScout24, Homegate, AutoScout24, Ricardo, tutti.ch and the software and data assets around them — listed on SIX on 19 September 2025 at CHF 46 per share, the largest European IPO of the year to that date, and closed its first day at CHF 49. It now trades around CHF 26.40, a drawdown of roughly 46% from the first close.

Over the same period, the company reported: FY2025 revenue of CHF 332.0m, up 14.1% (12.4% like-for-like); adjusted EBITDA of CHF 180.2m, up 29.4%, with the margin expanding 6.4 points to 54.3%; cash conversion of 81.2%; net leverage down to 0.7x; a proposed maiden dividend of CHF 0.82 per share; and 2026 guidance of 10–12% revenue growth at a 56–58% margin, with mid-term targets of low-teens growth and low-to-mid-60s margins reaffirmed from the IPO prospectus.

At IPO, the CHF ~4.5bn valuation implied roughly 25x that year's EBITDA. Today's enterprise value of approximately CHF 2.7bn (market capitalisation ~CHF 2.6bn plus ~CHF 0.13bn net debt) implies about 15x 2025 EBITDA and roughly 13x the 2026 guide (~CHF 365m revenue at ~57% margin ≈ CHF 208m). EBITDA less capex on that guide approaches CHF 170m — a pre-tax cash yield above 6% on the enterprise, growing at ten-plus percent, with a 3.1% dividend on top.

The multiple halved. The business accelerated. Something in the market's classification system, not in the company, changed. Our contention: SMG has been filed under "software" during the great software repricing, and the filing is wrong.

The derating SMG was drafted into

The 2025–26 software sell-off has an internally coherent logic. Application software is priced on the durability of its revenue, and generative AI attacks that durability from three directions at once: it collapses the cost of writing replacement software, it lets customers build internally what they once bought, and — the newest vector — agentic systems threaten to disintermediate the interface itself. If an AI agent can execute the workflow, the argument runs, the seat-based tool renting you that workflow loses its claim on the budget. Whatever one thinks of the timeline, the direction is real, and the market has repriced the entire category accordingly.

SMG's revenue is subscription-heavy, digital, high-margin, and reported in the same spreadsheet column as SaaS. Its own communications lean into "AI-driven strategy." Its IPO was priced in the last warm month before the category froze. So when the category derated, SMG derated with it — a 25x asset marked to a 13x world.

The question that decides everything is whether a classifieds marketplace shares the vulnerability that justified the category's repricing. We think the honest answer is no, and the reasoning matters more than the conclusion.

What a marketplace actually sells

Application software sells capability: the ability to perform a task. Capability is exactly what AI deflates — when the marginal cost of producing working software collapses, so does the scarcity value of any given tool.

A classifieds marketplace sells liquidity: the probability that the counterparty you need is present. That probability is not produced by code. It is produced by two decades of accumulated two-sided participation, and it exhibits the property that makes these businesses the best in the listed universe when they win: the value of the network to each participant is set by the behaviour of all the other participants, which no competitor — human or machine — can replicate by shipping features. Switzerland's 4,008 paying real-estate agencies list on ImmoScout24 and Homegate because that is where Swiss property demand is; Swiss property demand goes there because that is where the agencies list. An AI lab can reproduce the website in a weekend. It cannot reproduce the reason anyone visits it.

The evidence for this distinction is in the margin structure of the reference class, which we worked through in building our own screening rules. Auto Trader operates the UK car marketplace at a ~70% core operating margin, with management able to state that users spend six times more time on its platform than on all main competitors combined. CTS Eventim's ticketing network runs at ~46% EBITDA margins while its non-network live-entertainment business — same company, same management, no network — earns 6%. Rightmove and Scout24 tell the same story in property. Margins of this kind are not the reward for good software; software this simple would be competed to zero. They are monopoly rents on liquidity, and SMG's trajectory — 54.3% and guided toward the mid-60s — is the standard maturation curve of a classifieds franchise converging on its reference class, in a country small enough that it owns the #1 and #2 property portals and the #1 auto portal simultaneously.

The pricing evidence points the same way. Real-estate ARPA rose 14.7% in 2025 while the agency base grew to 4,008 — double-digit price extraction with net customer additions. Software companies raising price 15% in 2025 churned. SMG's customers paid, because the product is not the tooling; the product is access to the Swiss buyer, and there is one door.

The MCP bear case, steelmanned

Because the strongest version of the AI argument against marketplaces deserves a fair hearing, here it is, in the language of the systems being built right now.

An agentic AI with tool access does not browse. Give a capable model an MCP server — a standardised interface through which it can query structured inventory — and the entire demand-side experience of a classifieds portal (search, filters, alerts, comparison) is reproduced by the agent, better, inside a chat window. The user tells their assistant "find me a 4.5-room apartment in Zürich under CHF 3,200 and book viewings," and the assistant executes. In that world, the argument runs, the portal is demoted from destination to database: demand aggregates at the agent layer, the agent commoditises whatever it queries, and the marketplace suffers the fate every supplier suffers when someone else owns the customer relationship. This is aggregation theory with the aggregator upgraded from a search engine to an actor. It is not a stupid argument, and pieces of it are already operational: AI-generated answers are measurably reducing click-through to content sites, and classifieds portals receive a meaningful share of their top-of-funnel from exactly the search surface being cannibalised.

Three things are wrong with it as applied to SMG — and they are structural, not hopeful.

First: the agent has to query something, and the something is proprietary. The bear case quietly assumes the inventory is ambient — that listings exist somewhere neutral for an agent to aggregate. They do not. Swiss property listings exist because 4,000 agencies contractually pay SMG to host them; Swiss car inventory exists because dealers pay AutoScout24 to merchandise it. The listing is the marketplace's asset, contributed under commercial terms by suppliers who pay for the demand attached to it. An agent that wants Swiss property inventory must query SMG's systems, on SMG's terms, or query nothing. The correct analogy is not "AI replaces the software tool"; it is "a new client connects to the exchange." Exchanges do not fear new order-entry interfaces. They meter them. If agentic demand becomes real, machine-readable access to inventory is not SMG's disintermediation — it is SMG's next SKU, sold at a price that reflects exactly the liquidity monopoly described above. It is worth noting the company is already building in this direction: its stated investment in agentic capabilities targets automated workflows for the supply side — agents, dealers, sellers — which is the side that pays, and the side whose relationship an interface-layer AI cannot touch.

Second: the transaction has a body. Classifieds categories are heavy: apartments must be visited, cars inspected, a used sofa collected in Lausanne. The final commercial acts — the viewing, the escrow (SMG's MoneyGuard rails on Ricardo), the trust that the counterparty exists and the object matches the photos — live on infrastructure the marketplace operates in the physical jurisdiction. An interface layer can reroute discovery; it cannot absorb settlement. And in two-sided markets, whoever holds settlement and supply holds the pricing power, whatever happens to the front end. This, incidentally, is why the demand-side traffic erosion from AI search — the genuinely live version of the threat — hits content publishers catastrophically and portals only partially: a publisher's product is the pageview; a portal's product is the match, and matches are billed to the supply side, where SMG's revenue already sits.

Third: the economics of the attacker don't clear. For an aggregator to demote SMG to a database, it must first assemble Swiss-scale liquidity outside SMG — convince the agencies and dealers to list somewhere with no buyers, in a market of nine million people, against an incumbent whose regulatory settlement with the Swiss Price Supervisor has just fixed the pricing framework for three years. Global AI platforms will not build Swiss classifieds liquidity for a market this size; the prize is too small and the supply too locked. Small-market density, usually SMG's growth constraint, is here its shield: Switzerland is not worth invading and impossible to invade cheaply — at once too small to attack and too profitable to leave.

What we would actually worry about

Independence means the bull essay gets audited too. Three risks are real and none of them is an MCP server.

Growth here is price, and price has a referee. Swiss classifieds are penetrated; the 14% revenue line is substantially ARPA extraction from a captive professional base — the same value-extension playbook that made Rightmove shareholders rich and Rightmove's customers organised. SMG has already been to see the Price Supervisor once; the settlement buys three years of certainty and simultaneously proves the ceiling exists. A model whose terminal assumption is decades of double-digit price increases on Swiss estate agents is assuming something Swiss institutions have demonstrated they will not tolerate indefinitely. The mid-60s margin target is credible; the duration of low-teens growth beyond it is the honest debate — and consensus, notably, models top-line growth well below the company's own guide.

The overhang is arithmetic, not narrative. TX Group, Ringier (~20%), Swisscom and General Atlantic still hold the large majority of the shares. Every lockup expiry and every strategic review at a parent is supply, in a stock whose free float is thin relative to its market capitalisation. Some portion of the 46% drawdown is not a verdict on the business at all; it is the mechanical pricing of anticipated paper.

The top of the funnel is rented from Google. The defensible parts of the franchise are supply and settlement; discovery is partially rented, and the landlord is redecorating. The metric that would falsify our comfort is direct/app traffic share versus SEO-referred traffic. If Swiss users' first query moves to an assistant and the assistant's operator ever does decide the market is worth owning — a Google, not a startup — the third argument above weakens. We assign this a low probability for a market of this size, not a zero.

The trade the market is offering

Strip it to the exchange being proposed. At ~13x forward EBITDA and a >6% growing pre-tax cash yield, the market prices SMG as software whose revenue durability is newly suspect. The audited record shows a two-sided liquidity monopoly compounding price at double digits with expanding margins, sub-1x leverage, a maiden dividend, and a three-year regulatory truce. The AI-disintermediation case, examined at the protocol level rather than the slogan level, terminates in a conclusion opposite to the one the market drew: agents make interfaces worthless and inventory more valuable, and SMG owns the inventory, the supplier contracts, and the settlement rails. The risks that remain — pricing-regulation ceilings, shareholder overhang, rented discovery — are real, specific, and none of them is the risk the derating was pricing.

We publish no rating on SMG and hold no position. We publish the framework, because eighteen months from now the same category error — liquidity businesses repriced as capability businesses — will be sitting in the uncovered small-cap universe this firm exists to examine, at sizes where nobody else will have done the work.

The unsolicited suggestion: point the agents inward

Everything above concerns what AI does to SMG's revenue. We think the larger effect over the next three years is on its cost line, and the company's public posture suggests it is underweighting the opportunity. SMG's communications — and its engineering organisation's own writing on how AI is reshaping product, design and research — describe AI as a feature layer: smarter search, fraud detection, decision-support for agents and dealers. All worthwhile. But the most economically consequential application of current AI for this specific company is none of those. It is agentic software engineering, pointed at SMG's own codebase.

Recall what SMG structurally is: a 2021 assembly of formerly competing platforms. It operates the #1 and #2 property portals as parallel brands, plus an automotive stack, a general-marketplace stack, and the software and data businesses bolted alongside. Assembled-by-merger platforms carry assembled-by-merger engineering: duplicated systems, parallel codebases doing the same job twice, harmonisation backlogs that outlive their business cases. And SMG's delivery model is verifiably distributed: the group operates a captive development centre in Ho Chi Minh City — SMG Vietnam, the former NVG Technology, occupying a full floor of Viettel Tower and describing itself as an R&D hub serving the group's fifteen brands, with agile teams split between Vietnam and Switzerland. That inventory of work has a specific technical profile: migration, deduplication, API harmonisation, test coverage, refactoring. Well-specified, verifiable, regression-testable. This is precisely the class of work where agentic coding systems are already strongest — not speculative frontier capability, but the demonstrated sweet spot.

The Vietnam centre is the tell, and it cuts in a direction management should sit with. It means SMG has already harvested the conventional axis of engineering efficiency — the wage-arbitrage axis — as far as it goes. The honest corollary is that agentic coding saves less per seat when the seat is priced in Ho Chi Minh City rather than Zürich. But two components of the opportunity survive that discount entirely. The first is the coordination tax: a build organisation split across eight time zones pays for its cost advantage in hand-off latency, specification overhead, and the serialization of work around two continents' working hours — and this tax, unlike wages, is precisely what agents collapse, because an agent implementing against a well-specified backlog does not wait for anyone's morning standup. The second is velocity itself: SMG spends roughly CHF 150m a year below the revenue line, engineering is the largest controllable component, and the guided march to low-to-mid-60s margins assumes conventional operating leverage. An organisation that has already optimised the price of an engineering hour has exactly one lever left — the number of hours a unit of software requires — and that is the lever agentic development pulls.

The strategic version of the argument matters even more than the margin version. For a company whose history is acquisition and whose geography caps organic volume growth, integration cost is the tax levied on every future deal. Cut the cost of absorbing a codebase by an order of magnitude and the M&A arithmetic changes — targets previously too small to integrate profitably become accretive, and the consolidation runway in adjacent niches extends accordingly.

The industry is already producing examples of what this transition looks like when it arrives late and under duress. In June, US proptech Opendoor wound down its entire India operation of roughly 250 people, its CEO explaining that manual workflows once handled offshore had been unified and re-staffed with small AI-native teams closer to the customer. We cite it as a warning about sequencing, not as a template: Opendoor is restructuring from weakness — quarterly revenue down 37.5%, losses doubling — and executing the crude version of the shift because distress removed its choice about pace. Nothing in our argument asks SMG to dismantle Vietnam. The opposite: the centre's accumulated knowledge of fifteen codebases is precisely the asset an agent-first organisation needs most, redeployed from writing code toward specifying, reviewing and orchestrating it. The concrete recommendation is narrower and more demanding than any headcount decision: standardise the development process itself on AI agents — make agentic pipelines the group-wide default path for implementation work, with engineers as specifiers and reviewers, rather than leaving adoption to individual teams' enthusiasm. Organisations that standardise early, by choice, keep their people and compound the productivity. Organisations that wait tend to make the transition the way Opendoor just did.

There is a pleasing symmetry available to management here, and it doubles as the compressed version of this entire note. Generative AI deflates the price of software capability. That is a catastrophe if capability is what you sell — which is why the SaaS complex derated. SMG does not sell capability; it sells liquidity, and it buys capability, in quantity, every year. A liquidity monopoly whose input costs are deflating is on the winning side of the very trade the market punished it for. The company should act like it: fewer AI press-release features, more AI in the software factory. The stock market filed SMG with the sellers of software. Operationally, it has the chance to become one of Europe's most aggressive buyers of what AI actually cheapens.

Sources and method

SMG is covered by sell-side research; this note is a market commentary, not an initiation of coverage, and contains no recommendation, rating, or price target. Figures are drawn from company reporting (FY2025 results, 18 March 2026; H1 2025 results; SIX IPO records), the company's own public materials including its Vietnam development-centre site, press reporting on third parties cited in the text (Opendoor, Inc42, 11 June 2026), and market data as of 3 July 2026. We received no compensation from SMG Swiss Marketplace Group Holding AG or any related party, and the analyst holds no position in SMG securities. Our research process and conflicts policy are described in our methodology and disclosures pages.